fix(annotations): coverage floors for the Tag dialog round

Floors for the new images-favorites.mjs and images-tagging.mjs, and the
measured raises of image-tagging-state.mjs and image-tagging.mjs.
activity-scan.mjs keeps main's floors.

.github/workflows/viewerjs-check.yml

@@ -16,9 +16,19 @@ concurrency:
   group: ${{ github.workflow }}-${{ github.event_name == 'pull_request' && github.ref || github.run_id }}
   cancel-in-progress: ${{ github.event_name == 'pull_request' }}
 
+# Timeouts are the measured job time plus at least 50%, and never under 20 minutes for a browser job,
+# because runner speed varies. Before sharding (main 2cecfcf, one worker): frontend coverage 32m10s,
+# page workflows firefox 18m59s, webkit 19m26s. First sharded run (PR #274 run 39712909047, 3 workers,
+# no npm cache, where npm ci stalled for 5m in six of eight jobs): coverage shards 11m19s/10m47s/10m49s
+# (tests 5m20s/4m42s/4m46s), merge 5m52s (report and check 4s), firefox shards 7m53s/9m27s (tests
+# 4m41s/3m29s), webkit shards 12m25s+ (cancelled at the old 12-minute limit after its 267 tests passed in
+# 6m29s) and 4m53s (tests 3m08s). With the npm cache hit (run 39725176320 at cbc8a61, npm ci 0-2s):
+# coverage shards 7m34s/6m04s/5m22s (tests 6m38s/4m53s/4m04s), merge 54s, firefox 7m57s/5m21s (tests
+# 7m08s/4m34s), webkit 8m16s/5m03s (tests 7m26s/3m58s); a cache miss adds up to 5m to npm ci, which
+# the 20-minute limits absorb.
 jobs:
   frontend-coverage:
-    name: frontend coverage (chromium and touch)
+    name: frontend coverage (shard ${{ matrix.shard }}/3)
     runs-on: ubuntu-latest
     # Layout comparisons depend on the fonts each browser falls back to, so browser jobs run in the
     # same pinned Playwright image as local runs (tag = @playwright/test in tools/vendor/package-lock.json;
@@ -26,12 +36,15 @@ jobs:
     container:
       image: mcr.microsoft.com/playwright:v1.64.1-noble
       options: --user 1001
-    # This job runs every page workflow twice (chromium, then touch) on one worker with V8 coverage on, so it
-    # takes about twice a single-browser job. Measured 2026-10-20: whole job 18m26s-24m57s (main 5ded3cf
-    # 22m15s; PR head 6359999 was cancelled at the old 25-minute limit while still passing), coverage step
-    # alone 20m55s-22m17s+, against ~12-15 minutes for the firefox/webkit jobs. 40 minutes leaves headroom
-    # for suite growth while still stopping a hung run; re-measure if a run passes 30 minutes.
-    timeout-minutes: 40
+    timeout-minutes: 20
+    # Public-repo ubuntu-latest runners have 4 vCPUs: three Playwright workers, each with its own test
+    # server, leave one CPU for the servers and the runner.
+    env:
+      PM_TEST_WORKERS: 3
+    strategy:
+      fail-fast: false
+      matrix:
+        shard: [1, 2, 3]
     steps:
       - uses: actions/checkout@v7
       # The image's own Python 3.12 and Node 24: setup-python/setup-node install the runner's toolcache
@@ -50,24 +63,107 @@ jobs:
           "$venv/bin/python" -m pip install --no-cache-dir -r requirements-dev.txt
           echo "$venv/bin" >> "$GITHUB_PATH"
       - run: python --version && node --version && node --version | grep -q '^v24\.'
+      # npm's download cache, keyed on the lockfile; inside the container ~ is the job's HOME (/github/home),
+      # which npm also uses. node_modules itself is still built fresh by npm ci.
+      - name: Cache npm downloads
+        uses: actions/cache@v6
+        with:
+          path: ~/.npm
+          key: npm-${{ runner.os }}-playwright-v1.64.1-${{ hashFiles('tools/vendor/package-lock.json') }}
       - run: npm ci --prefix tools/vendor --ignore-scripts
-      - run: npm run --prefix tools/vendor check
-      - run: npm run --prefix tools/vendor test
-      - name: Frontend unit, viewer, workflow, and coverage checks
-        run: npm run --prefix tools/vendor test:coverage
-      - name: Upload frontend coverage and Chromium traces
+      # Shard 1 also runs the node units and the viewer suite, so the vendor pins and the tooling tests
+      # run on the lightest shard instead.
+      - if: matrix.shard == 3
+        run: npm run --prefix tools/vendor check
+      - if: matrix.shard == 3
+        run: npm run --prefix tools/vendor test
+      - name: Page workflows with coverage, shard ${{ matrix.shard }}/3
+        run: npm run --prefix tools/vendor test:coverage:shard -- ${{ matrix.shard }}/3
+      - name: Upload raw coverage
+        uses: actions/upload-artifact@v7
+        with:
+          name: coverage-raw-${{ matrix.shard }}
+          path: coverage/js/raw/
+          if-no-files-found: error
+          retention-days: 3
+      - name: Upload Chromium and touch traces
+        if: always()
+        uses: actions/upload-artifact@v7
+        with:
+          name: frontend-traces-shard-${{ matrix.shard }}
+          path: tools/vendor/test-results/
+          retention-days: 14
+
+  frontend-coverage-merge:
+    name: frontend coverage (merge)
+    needs: frontend-coverage
+    runs-on: ubuntu-latest
+    # Same image as the shards: the raw files name sources by the container's workspace path.
+    container:
+      image: mcr.microsoft.com/playwright:v1.64.1-noble
+      options: --user 1001
+    timeout-minutes: 10
+    steps:
+      - uses: actions/checkout@v7
+      - run: node --version && node --version | grep -q '^v24\.'
+      # npm's download cache, keyed on the lockfile; inside the container ~ is the job's HOME (/github/home),
+      # which npm also uses. node_modules itself is still built fresh by npm ci.
+      - name: Cache npm downloads
+        uses: actions/cache@v6
+        with:
+          path: ~/.npm
+          key: npm-${{ runner.os }}-playwright-v1.64.1-${{ hashFiles('tools/vendor/package-lock.json') }}
+      - run: npm ci --prefix tools/vendor --ignore-scripts
+      # Each artifact lands in its own coverage-raw-N directory, so the next step can count them.
+      - uses: actions/download-artifact@v8
+        with:
+          pattern: coverage-raw-*
+          path: ${{ runner.temp }}/coverage-raw
+      # coverage:merge only checks that raw files exist; merging two of three shards would report
+      # partial coverage, so a missing or empty shard artifact fails here.
+      - name: Verify every shard's raw coverage arrived
+        shell: bash
+        env:
+          STAGED_RAW: ${{ runner.temp }}/coverage-raw
+          EXPECTED_SHARDS: 3
+        run: |
+          found=0
+          for shard in $(seq 1 "$EXPECTED_SHARDS"); do
+            directory="$STAGED_RAW/coverage-raw-$shard"
+            count=0
+            if [ -d "$directory" ]; then count=$(find "$directory" -maxdepth 1 -type f -name '*.json' | wc -l); fi
+            echo "coverage-raw-$shard: $count raw files"
+            if [ "$count" -gt 0 ]; then found=$((found + 1)); fi
+          done
+          echo "shard artifacts with raw coverage: $found of $EXPECTED_SHARDS"
+          if [ "$found" -ne "$EXPECTED_SHARDS" ]; then
+            echo "::error::Expected raw coverage from $EXPECTED_SHARDS shards, got $found of $EXPECTED_SHARDS"
+            exit 1
+          fi
+          staged=$(find "$STAGED_RAW" -mindepth 2 -maxdepth 2 -type f -name '*.json' | wc -l)
+          mkdir -p coverage/js/raw
+          # cp -n never overwrites; a clash shows up as fewer merged files than staged ones.
+          find "$STAGED_RAW" -mindepth 2 -maxdepth 2 -type f -name '*.json' -exec cp -n -t coverage/js/raw {} + || true
+          merged=$(find coverage/js/raw -maxdepth 1 -type f -name '*.json' | wc -l)
+          if [ "$merged" -ne "$staged" ]; then
+            echo "::error::Shards wrote clashing raw file names: $staged staged, $merged merged"
+            exit 1
+          fi
+          echo "merged $merged raw files"
+      - name: Frontend coverage report and baseline check
+        run: npm run --prefix tools/vendor coverage:merge
+      - name: Upload frontend coverage report
         if: always()
         uses: actions/upload-artifact@v7
         with:
           name: frontend-test-reports
           path: |
-            tools/vendor/test-results/
             coverage/js/
             !coverage/js/raw/**
           retention-days: 14
 
   browser-workflows:
-    name: page workflows (${{ matrix.browser }})
+    name: page workflows (${{ matrix.browser }}, shard ${{ matrix.shard }}/2)
     runs-on: ubuntu-latest
     # Layout comparisons depend on the fonts each browser falls back to, so browser jobs run in the
     # same pinned Playwright image as local runs (tag = @playwright/test in tools/vendor/package-lock.json;
@@ -75,11 +171,16 @@ jobs:
     container:
       image: mcr.microsoft.com/playwright:v1.64.1-noble
       options: --user 1001
-    timeout-minutes: 25
+    timeout-minutes: 20
+    # Public-repo ubuntu-latest runners have 4 vCPUs: three Playwright workers, each with its own test
+    # server, leave one CPU for the servers and the runner.
+    env:
+      PM_TEST_WORKERS: 3
     strategy:
       fail-fast: false
       matrix:
         browser: [firefox, webkit]
+        shard: [1, 2]
     steps:
       - uses: actions/checkout@v7
       # The image's own Python 3.12 and Node 24: setup-python/setup-node install the runner's toolcache
@@ -98,22 +199,29 @@ jobs:
           "$venv/bin/python" -m pip install --no-cache-dir -r requirements-dev.txt
           echo "$venv/bin" >> "$GITHUB_PATH"
       - run: python --version && node --version && node --version | grep -q '^v24\.'
+      # npm's download cache, keyed on the lockfile; inside the container ~ is the job's HOME (/github/home),
+      # which npm also uses. node_modules itself is still built fresh by npm ci.
+      - name: Cache npm downloads
+        uses: actions/cache@v6
+        with:
+          path: ~/.npm
+          key: npm-${{ runner.os }}-playwright-v1.64.1-${{ hashFiles('tools/vendor/package-lock.json') }}
       - run: npm ci --prefix tools/vendor --ignore-scripts
-      - name: Page workflows in ${{ matrix.browser }}
-        run: npm run --prefix tools/vendor test:workflows -- --project ${{ matrix.browser }}
+      - name: Page workflows in ${{ matrix.browser }}, shard ${{ matrix.shard }}/2
+        run: npm run --prefix tools/vendor test:workflows -- --project ${{ matrix.browser }} --shard ${{ matrix.shard }}/2
       - name: Upload browser traces
         if: always()
         uses: actions/upload-artifact@v7
         with:
-          name: browser-traces-${{ matrix.browser }}
+          name: browser-traces-${{ matrix.browser }}-shard-${{ matrix.shard }}
           path: tools/vendor/test-results/
           retention-days: 14
 
   # Branch protection requires the exact context "viewerjs". This job keeps that
-  # name and fails unless every frontend job above succeeded, including when one
-  # was cancelled or skipped.
+  # name and fails unless every frontend job above succeeded (every shard, the
+  # coverage merge and every browser shard), including when one was cancelled or skipped.
   viewerjs:
-    needs: [frontend-coverage, browser-workflows]
+    needs: [frontend-coverage, frontend-coverage-merge, browser-workflows]
     if: always()
     runs-on: ubuntu-latest
     timeout-minutes: 5
@@ -121,8 +229,10 @@ jobs:
       - name: Require every frontend job to succeed
         env:
           COVERAGE_RESULT: ${{ needs.frontend-coverage.result }}
+          MERGE_RESULT: ${{ needs.frontend-coverage-merge.result }}
           WORKFLOWS_RESULT: ${{ needs.browser-workflows.result }}
         run: |
           echo "frontend-coverage: $COVERAGE_RESULT"
+          echo "frontend-coverage-merge: $MERGE_RESULT"
           echo "browser-workflows: $WORKFLOWS_RESULT"
-          test "$COVERAGE_RESULT" = success && test "$WORKFLOWS_RESULT" = success
+          test "$COVERAGE_RESULT" = success && test "$MERGE_RESULT" = success && test "$WORKFLOWS_RESULT" = success

TESTING.md

@@ -39,9 +39,44 @@ npm run --prefix tools/vendor test:coverage
 npm run --prefix tools/vendor test:workflows -- --project firefox --project webkit
 ```
 
-`PM_TEST_PYTHON` selects the Python executable; `PM_TEST_PORT` selects the local
-port (default 8765). Set these environment variables when Python is outside PATH
-or the port is occupied. Test servers are never reused. On unsupported Linux
+`PM_TEST_PYTHON` selects the Python executable; set it when Python is outside
+PATH. Every Playwright worker starts its own `tests/browser/server.py --port 0`,
+which binds a port the OS picks and prints it as `PM_TEST_SERVER_PORT=<n>`, with
+its own temp root and database (`tests/browser/fixtures.mjs`), so
+parallel workers and concurrent runs never share state. `PM_TEST_WORKERS` sets the
+worker count (default 1 under `CI`, otherwise a quarter of the CPUs, at most 8).
+Test servers are never reused across runs. The launcher holds each server's
+stdin open and the server (`--exit-on-stdin-eof`) shuts down cleanly, removing its
+temp root, when that pipe closes, so a worker killed without teardown leaves no
+server behind. A test that fails gets the stderr its server wrote during the test
+attached to the report as `server-stderr` (an aiohttp traceback behind a 500).
+
+The coverage run can be split the way CI splits it, and each shard keeps the raw
+V8 files it writes under `coverage/js/raw`:
+
+```sh
+PM_TEST_WORKERS=8 npm run --prefix tools/vendor test:coverage            # whole run, one report
+npm run --prefix tools/vendor test:coverage:shard -- 1/3                 # node units, viewer suite, workflows 1/3
+npm run --prefix tools/vendor test:coverage:shard -- 2/3                 # workflows 2/3 (likewise 3/3)
+npm run --prefix tools/vendor coverage:merge                             # report and baseline check of every raw file
+npm run --prefix tools/vendor test:workflows -- --project webkit --shard 2/2
+```
+
+`coverage:merge` refuses an empty raw directory; it does not count shards, so run
+every shard first (CI's merge job checks all three arrived). Measured on a 32-CPU
+machine in the pinned image (2026-10-21): `test:coverage` takes about 5 minutes
+with 8 workers and 24 minutes with 1; all four projects, 8 workers, each test 3
+times (6132 tests), 24 minutes. A CI-sized run (`docker run --cpus=4`,
+`PM_TEST_WORKERS=3`, all four projects once, 2044 tests) takes 25.5 minutes.
+
+Specs must pass beside each other, not only alone. Wait for the page's own
+signal, never for time: a rendering frame (`requestAnimationFrame`) for layout
+work queued by a render or a scroll, a response or a visible state for data. A
+timer that "usually" covers a frame is a race on a loaded runner. Library
+virtualization specs settle with `libraryObservation.settled()` (two consecutive
+frames with no Library frame queued). To look for such races, repeat a spec under
+load (`--workers 8 --repeat-each 10`, and in the image with `--cpus=1` or `2`).
+On unsupported Linux
 distributions, browser system dependencies may need manual installation. For a
 quick browser run, use `test:workflows -- --project chromium`.
 
@@ -114,16 +149,39 @@ second without a browser and fails, naming the files, when a measured
 `web/js` file has no baseline entry or the baseline lists a file that no longer
 exists. It uses the gate's own `discoverSources`, and `npm test` runs the same
 check, so run it before pushing a new or deleted frontend file instead of
-waiting for the `frontend coverage` job.
+waiting for the `frontend coverage` jobs.
 
 The `viewerjs` check runs on every PR, including backend-only changes. It is an
-aggregate job that passes only when the `frontend coverage (chromium and touch)`
-job and both `page workflows (firefox)` and `page workflows (webkit)` jobs
-succeed; the browsers run on separate runners in parallel. GitHub artifacts
-named `frontend-test-reports` contain HTML coverage, a JSON summary, and the
-Chromium/touch failure screenshots/traces; `browser-traces-firefox` and
-`browser-traces-webkit` hold the other browsers' traces. Open traces with
-Playwright's `show-trace` tool.
+aggregate job that passes only when every frontend job succeeds: the three
+`frontend coverage (shard N/3)` jobs, `frontend coverage (merge)`, and the four
+`page workflows (firefox|webkit, shard N/2)` jobs. A failed, cancelled or skipped
+job fails it. Every browser job runs with `PM_TEST_WORKERS: 3` (the public
+runners have 4 vCPUs). Each coverage shard runs `test:coverage:shard -- N/3`
+(shard 1 also runs the node units and the viewer suite, shard 3 the vendor pins
+and tooling tests) and uploads its raw V8 files as `coverage-raw-N` (kept 3
+days). The merge job downloads them, fails unless all three arrived with raw
+files and without clashing names, then runs `coverage:merge` (report plus
+baseline check). `tests/vendor/ci-workflow.test.mjs` checks this shape. GitHub
+artifacts named `frontend-test-reports` contain the merged HTML coverage and
+JSON summary; `frontend-traces-shard-N` hold the Chromium/touch failure
+screenshots/traces and `browser-traces-<browser>-shard-N` the other browsers'
+traces. Open traces with Playwright's `show-trace` tool.
+
+Every `npm ci` first restores npm's download cache (`~/.npm`), keyed on the vendor
+lockfile and the job's pinned image tag (`tests/vendor/ci-workflow.test.mjs` checks
+the key names the same tag as the job's `container.image`). Without it, `npm ci` stalled for about 300 s
+(npm's fetch timeout) in most of the eight concurrent jobs; with a hit it takes
+0 to 2 s. A lockfile change misses the cache once, so expect one slow run after it.
+Measured on PR #274 (run 39725176320 at cbc8a61, cache hit): the whole workflow
+took 8m38s, against about 33 minutes with one worker and no shards. Jobs: coverage
+shards 7m34s/6m04s/5m22s (tests 6m38s/4m53s/4m04s), merge 54s, firefox shards
+7m57s/5m21s (tests 7m08s/4m34s), webkit shards 8m16s/5m03s (tests 7m26s/3m58s).
+Browser jobs keep a 20-minute limit: the slowest job is about 8.5 minutes, a cache
+miss adds up to 5 to `npm ci`, and runner speed varies (the first sharded run was
+cancelled at a 12-minute limit).
+Playwright shards by test count, whole files at a time, so shard 1 carries the
+slow concept-comparison specs (`*-concept.spec.mjs`, alphabetically first) and
+runs longer; the remedy, if it starts to matter, is more shards, not reordering.
 
 The real-host suite covers both Classic and Nodes 2.0 startup, switching both
 ways, actual database prompt selection, execution-text serialization, workflow
@@ -234,8 +292,9 @@ The shared inline and collection inspector reads the current rooted file's suppo
 embedded records, exact generation facts and recorded or inferred prompt evidence.
 It renders in two concept surfaces (#249 D20–D22, D38). The Images side panel shows
 only the concept content: cover, title, filename, the evidence badge with a category
-chip, the + and − prompt blocks with copy and send, tag chips (the linked positive's
-own tags filter Images, D21), the six facts, and a footer with "Send to ComfyUI" and
+chip, the + and − prompt blocks with copy and send, tag chips (the file's own image
+tags first, D88, then the linked tags; image tags and the linked positive's own tags
+filter Images, D21), the six facts, and a footer with "Send to ComfyUI" and
 "Prompt history" only when prompt roles resolved or a link is saved; the sticky panel
 scrolls inside the window. The viewer's details (`#viewer-details`) repeat that
 content without the cover, then "File & provenance" (the honest file facts, the
@@ -274,7 +333,30 @@ becomes record format `limited` (no raw text, graph `limited`, roles and facts
 unaffected); other text chunks are skipped unread; byte-identical duplicates are
 accepted within a bounded read budget, while differing duplicates (and any duplicate
 role record over that budget) still fail. Other read bounds are unchanged: 8 MiB
-non-PNG header, 10,000 PNG chunks and a three-second ffprobe. Resolution under
+non-PNG header, 10,000 PNG chunks and a three-second ffprobe (per probe, at most two
+probes). Video containers are read through one shared carrier selector
+(`utils/video_metadata.select_video_carrier`, used by the inspector and by the admin
+scan): carriers are, in order, the per-key tags that current savers write (`prompt`,
+`workflow`, `parameters`, `comfyassets_prompt_pairs`, each `original-text`; Matroska
+upper-cased names fold on our side), then the old VideoHelperSuite `comment` JSON
+envelope, then `description` (object values are `reserialized-json`; envelope keys are
+matched exactly, as the legacy readers did, never folded). The first carrier holding a
+role key wins, otherwise the first holding a workflow, and carriers never merge. Two
+per-key tags folding to one key are a duplicate record: byte-identical ones are one
+record, differing ones fail the read. ffprobe is asked for exactly those six tags.
+Its JSON output escapes tag text (at most doubling it), so the stdout bounds are sized
+so that only the decoded per-record budgets above can fail or limit a record, and a
+video reads exactly as the identical PNG: the first probe allows twice the role and
+workflow budgets plus 64 KiB of envelope (12 MiB + 64 KiB); past that the inspector
+probes once more without the `workflow` tag within twice the role budget plus 64 KiB
+(4 MiB + 64 KiB) and the per-key workflow becomes `limited`, while an envelope over
+that bound still fails the read because it cannot be split. The admin scan reads the
+same carrier with the default resolver profile, so a per-key video links like the same
+graph in a PNG; the live monitor processes images only (unchanged, see the follow-up
+below) (`tests/test_utils_video_metadata.py`, `tests/test_api_admin.py`
+`TestScanReadsVideos`, `tests/test_api_prompt_inspector.py` `test_video_*`,
+`tests/test_inspection_response.py` `TestVideoProbeBoundsMatchPngBudgets`; the
+real-container round trips skip without ffmpeg). Resolution under
 `INSPECTION_PROFILE` bounds work, not graph size: 2,000 visits per candidate chain,
 20,000 in total, 64 link/JSON-container levels, 100,000 resolved characters, and a
 two 20,000-node ceilings: one over the top-level entries of the API prompt, and one
@@ -688,8 +770,12 @@ Recorded shell deviations:
   "Save", not "Apply in prototype", and its toast drops the demo's word "Sample".
 - **D31**: at `max-width: 950px` the overlay caps the one-column shell at
   `minmax(0, 1fr)`, so extreme zoom cannot widen it to the sidebar's min-content.
-- **#245 D2**: no "Auto-tag images" on the Tags heading until the image-annotation
-  slice.
+  On the Images cards (#249 C2) a card's info row stacks its text above Info and Tag
+  once the row is narrower than the narrowest bounded card (120px; a list row's info
+  is what its art leaves), and the two buttons wrap when even that row cannot hold
+  them (`stacksCardInfo`, `tests/js/workspace-image-layout.test.js`).
+- **#245 D2 (retired, #249 C2)**: the Tags heading shows the concept's "Auto-tag images";
+  it opens the Tag image dialog (D115 when nothing is selected).
 - **Brand name**: not a deviation. The brand link's accessible name matches the
   concept's, mark included, because production's `.brand-mark` is not `aria-hidden`
   either. The inventory compares `brand.name` and `brand.mark.hidden`.
@@ -712,13 +798,15 @@ zoom, the 2nd card opens the viewer titled "<title> · <filename>", Escape retur
 to it (production only: the concept re-renders its grid and drops focus), Info on the
 3rd card moves `.selected`, searches for "alpine" and for one image, and Load more with
 its toast. `mapImagesConcept` applies D3, D8 (the selection bar's real actions only),
-D9, D10, D14, D15, D17, D29, D32, D35 (singular counts) and D36 (categories
+D10, D14, D15, D17, D29, D32, D35 (singular counts) and D36 (categories
 sorted alphabetically). Removed toolbar and filter-row items flow again from the
 concept's own boxes (`reflow`), checked against the concept page with those items
 removed at runtime. Masonry tops and heights are compared unrounded, within one layout
 unit per card above in the column (D3). `one footer height for the collection (D3)`
-seeds two seedless files (`seedless_images` on `/__test__/concept/reset`) and asserts
-that every row reserves the tallest footer.
+seeds two seedless files (`seedless_images` on `/__test__/concept/reset`, with
+`short_seedless` so their "96 × 72 · PNG" subtitle stays on one line beside Info and
+Tag) and asserts that every row reserves the tallest footer and that their own footer
+is shorter than the row it reserves.
 
 ### Images page controls and recorded deviations (#249 PR C)
 
@@ -804,7 +892,9 @@ Recorded deviations on this page (spec §8):
   the concept` compares the bar and the checked card with one card checked; the omitted
   actions flow again from the concept's own boxes, checked against the concept page with
   them removed at runtime.
-- **D9**: no Tag button until the image-annotations slice.
+- **D9 (retired, #249 C2)**: every card shows the concept's Tag (`tagImage`, "Tag or
+  auto-tag <filename>") after Info; it opens the Tag image dialog without selecting
+  the card (see "Tag image dialog" below).
 - **D10**: the card title is the linked positive's display title, else the filename;
   the `h3` keeps the relative path in `title`.
 - **D11**: the subtitle is "W × H · EXT · Seed N" with unknown parts omitted.
@@ -862,6 +952,8 @@ Recorded deviations on this page (spec §8):
   history appear only with resolved prompt roles or a saved link; a video file keeps
   `video.detail-cover`; and only the linked positive's own tags are `button.chip`
   filters, while recipe, recorded-run and embedded tags stay plain `span.chip`.
+  Amended by D88 (#249 PR C): the file's own current image tags lead the same row
+  and are filters too; a linked tag already shown (case-insensitive) is left out.
 - **D22**: an empty positive reads "No positive prompt (explicitly empty)" (the concept
   applies its negative wording to both roles), and a confirmed `p.prompt-text` is
   focusable (`tabindex=0`, marked focus ring) so long prompts stay keyboard-readable.
@@ -968,6 +1060,186 @@ npm run --prefix tools/vendor test:workflows -- images-concept.spec.mjs \
   --project chromium --project touch --project firefox --project webkit
 ```
 
+### Tag image dialog and recorded deviations (#249 PR C)
+
+Every Images card and every Library-embedded gallery card shows Info then the
+concept's Tag (`app.js` mediaCard). Tag opens the concept's **Tag image** dialog
+(`web/js/workspace/image-tagging.mjs` over the C1 controller
+`image-tagging-state.mjs`) in the shared modal shell with the concept's
+`tagging-modal` class, so its `.dialog-actions` stay sticky while the content
+scrolls. Focus starts on × and returns to the Tag button. The preview is the
+capture's own `preview_url` (through `ApiClient.resolveUrl`): an `img`, or a
+`video preload="metadata" muted`. The capture is released on ×, Escape, Cancel,
+Save, a route change (the shell closes its dialog) and `pagehide`, which closes the
+dialog (a page restored from the back/forward cache shows no dead dialog); the
+release `DELETE` and the Save `PUT` are sent with `keepalive` (`ApiClient`'s
+additive option), so they finish while the page unloads. A save still in flight
+when the dialog closes keeps its capture until the `PUT` settles, then releases it
+and still reports the outcome: the saved toast and every page's refresh, or the
+failure as a toast. Generate is explicit and uses the server's real defaults
+(`GET /prompt_manager/autotag/models`), and the fields show exactly those values
+(they are re-rendered when the observation lands and on a model change, never while
+typing); an invalid field reports native validity and sends nothing. A tag conflict
+shows the current set at once: the status line appends "Saved now: <tags>." and the
+manual field holds the current set ahead of the draft (repeats dropped), so the next
+Save sends what is shown against the current revision. Save writes the manual
+names, then the remaining chips, deduplicated case-insensitively (first spelling
+wins), and toasts the mock's copy verbatim, "<n> tags saved for <filename>." even
+for one tag, plus " Prompt association still awaits review." from Needs review.
+Tagging one file never changes another file (an equal-byte copy included), a
+prompt, a pair or a link. States the mock lacks (generating, busy, conflict,
+changed storage or file, expired capture, unconfirmed save, unavailable model,
+video) are text in `#tag-generation-status` only; **Retry** (D116) leads the
+actions while the controller allows it, where the other workspace dialogs place
+their Retry. The Tags
+page re-reads its lists after every confirmed save, so the Images counts follow.
+
+The image-tagging plan's deviation numbers were renumbered by +23 because the
+Inspect plan keeps D60–D66 (plan D60 → D83, D61 → D84, D65 → D88, D68 → D91).
+
+- **D83** (plan D60, ruling T4): the suggestions are removable chips instead of
+  checkbox rows: `span.chip.tag-suggestion` in a `.tags` row, in the concept's
+  checked-row tokens (accent border, `--soft-blue`), each with the name, the
+  confidence (`98%` for WD14, "Suggested" for language models) and a × "Remove
+  <tag>". A chip has the concept `.chip`'s 4px corner (the later `.tag-suggestion`
+  row rule would give 5px) around the row's own 12px text. At most the requested
+  count, all preselected. While chips exist, ruling T4's "Remove anything you do not
+  want to add." (`p.small`) leads them inside `#tag-suggestions`; removing the last
+  chip shows "All suggestions removed. Generate again or enter tags manually." The
+  opening sentence is the concept's on-screen copy, "Generate suggestions with the
+  selected model before approving.": the concept's markup carries another ("Choose a
+  model… Uncheck anything…"), but its `renderTagModelOptions` (a1062–1076) replaces it
+  as soon as the dialog opens, so the opened state is compared unmapped.
+- **D84** (plan D61, rulings T1/T2): the Tagging model options are legacy's four
+  engines in the mock's order with legacy labels and sizes (`WD14 SwinV2 · ~400
+  MB`, `WD14 ViT · ~400 MB`, `GGUF JoyCaption (Recommended) · ~4.5 GB` in the
+  mock's Ollama slot, `HuggingFace JoyCaption · ~16 GB`); the instructions start
+  from the server's `default_prompt`.
+- **D91** (plan D68, ruling T3-bulk): a **Suggestions** number field (1–100) for
+  every model, prefilled from `autotag.defaults.suggestion_limit` (10); a model
+  change keeps it.
+- **D88** (plan D65, ruling T7/C2): the inspector's chip row (side panel and viewer
+  details, after the prompt blocks) shows the file's current image tags first (the
+  B3 envelope's `image_tags`, `state: "current"`, S1), then D21's linked tags not
+  already shown (case-insensitive, the image tag's spelling wins). In the Images
+  panel every image tag is a Tags filter `button.chip` (the filter matches image
+  tags too) and linked tags keep D21's rule; in the viewer, and wherever no Tags
+  filter exists (the Library-embedded gallery, Activity and recorded runs), every
+  chip is a plain `span.chip`. `none` and `unavailable` add no chips and no other
+  chrome; a malformed facet fails the inspection like any malformed field. The
+  concept has one row of the image's own tags (`i.tags`, app.js a402), so with the
+  concept image's tags saved as its image tags production still equals the concept.
+  After a Tag save the selected file's inspection is re-read, so the panel redraws
+  the row without reselecting and the viewer, which shares that inspection, shows it.
+- **D115** (beyond the plan's list): the concept always has a selected image, so
+  its "Auto-tag images" always opens the dialog. Production opens it for the image
+  selected in Images when it was last left; with none it is `aria-disabled` and
+  pressing it toasts "Select an image in Images first." (no new chrome).
+- **D116** (ruling T5): **Retry** leads `.dialog-actions` (before Cancel and the
+  primary) while a failed observation, a failed Generate, an unconfirmed save or an
+  answer that needs a fresh observation (changed storage or file, expired capture)
+  can be retried; the concept has no failure states, so it has no Retry. After a
+  failed Generate it takes Generate's own path (an invalid field reports its
+  validity); an unconfirmed save resends the identical body; after a changed storage,
+  file or expired capture it observes again and keeps the typed draft (when the saved
+  set changed meanwhile, the current set leads the draft and the status line says
+  "The tags of this image changed since the dialog opened. Saved now: <tags>.").
+  Browser tests: "D116: Retry leads the actions after an unconfirmed save…", "D116:
+  Retry after an expired capture observes again and keeps the typed draft (M5)" and
+  the paired `production-unknown-outcome` capture.
+- **D29** (no prototype copy) in the dialog: no "Prototype · suggestions are
+  simulated…" note, the status reads "Suggestions from <model>. Approve tags to save."
+  (not "Sample suggestions from …"), and the toast has no " in this tab".
+
+After a confirmed save the selected file's inspection is re-read quietly: the panel
+and the viewer keep the current details until the new answer lands, with no "Loading
+file details…" in between (`createInspectionController().refresh()`). While a Tags
+filter or a search is active the saved file may start or stop matching, so Images
+re-reads its loaded range in place (`createImageStream().refresh()`: the same query
+from the first page until the loaded count, published once). The cards, the scroll
+position and the selection stay while the selected file still matches; when it no
+longer does, the panel follows the collection (its first file), as after a reload,
+and the selection bar keeps only the checked files still listed, so its count is what
+Set category and Export metadata apply to. A refresh never runs beside a page load
+already in flight: asked during one, it is pending and runs once that page settles
+(success or failure; a reset drops it), and Load more waits for a running refresh
+(it neither loads nor moves focus). The re-read asks for pages as large as the rest of
+the loaded range (at most 200 files, at least one batch), one after another (each
+cursor comes from the page before); a failed re-read toasts "Images could not be
+refreshed after the tag save. Refresh image collection to retry." and keeps the
+cards. A save made while Images is left (Tags › Auto-tag images) marks a saved Images
+visit with a Tags filter or a search stale: returning lists it again with its
+filters instead of restoring the old cards. Browser tests: "the post-save re-read
+keeps the panel content until the new read lands", "a save under a Tags filter
+re-reads the listing: …", "a save under a search keeps the scroll position and the
+loaded range", "a failed listing re-read after a save says so and keeps the cards
+(M2)" and "a tag save from Tags reloads a filtered Images visit instead of restoring
+it (M6)"; the pending refresh and the page sizes are unit-tested.
+
+Concept comparison (`tests/browser/tagging-concept.spec.mjs` with
+`tagging-canonical-comparison.mjs`, plan Task C4): production runs on the concept
+fixture with the concept image's own tags (`cinematic, blue_hour, architecture`)
+saved as its image tags, and the fake model answers the concept's own sample
+(`/__test__/tagging/suggestions`: mountain 0.98, observatory 0.95, snow 0.88,
+blue_hour 0.81), so the dialog's inventory (order, text, field attributes, tokens,
+preview box, chips, status, sticky actions) must equal the concept's at 1600×1100 and
+390×844, dark and light: opened, generated with WD14 and generated with a language
+model, plus the empty result (the concept above its own threshold), the review-scoped
+dialog with its scope note and the card's Info and Tag group (M1). `mapConceptTagging`
+applies only D29, D83, D84 and D91; a mutation check (the model panel's border
+removed, the primary relabelled) shows the comparison notices a change. Every
+production-versus-concept difference of PR C and its record:
+
+| Difference | Record |
+| --- | --- |
+| Removable chips with × instead of checked checkbox rows; "Remove anything…" while chips exist; the all-removed copy | D83 (ruling T4) |
+| Legacy engine labels and values, GGUF in the Ollama slot, server default instructions | D84 (rulings T1/T2) |
+| Suggestions count field | D91 (ruling T3-bulk) |
+| No prototype note, "Suggestions from …", no " in this tab" | D29 |
+| Image tags first in the inspector and viewer chip row | D88 (ruling T7/C2) |
+| Auto-tag images `aria-disabled` without a selected image | D115 |
+| Retry leads the actions when the controller allows it | D116 (ruling T5) |
+| States the concept lacks (busy, conflict, changed storage or file, unconfirmed save, unavailable model, video) | Global Constraints: `#tag-generation-status` text only |
+
+Native image tagging (#249 PR B routes, PR C UI): `POST
+/prompt_manager/image-annotations/observe` (capture one listed file, its saved set and
+`tag_revision`), `GET`/`DELETE /prompt_manager/image-annotations/capture/<token>`
+(the preview and its release), `POST /prompt_manager/image-annotations/suggest` (the
+gated top-N suggestions, at most the requested count) and `PUT
+/prompt_manager/image-annotations/tags` (re-hash, then save one byte record's set).
+Cheap reads follow spec amendment S1 (the verified `(size, mtime_ns)` fingerprint;
+residual risk above, under "Image annotations schema"). No automated test downloads
+or loads a model: the browser server injects `TestModel`, whose `download_model`
+raises, and every tagging spec asserts no `/autotag/download/` request. Capture
+limits are those of the capture store above (256 MiB per file, 8 completed captures,
+512 MiB reserved, 900 s idle, 1800 s hard). The browser fixtures release every capture
+on reset. PR C is HTML, CSS and JS only (no restart); the PR B routes it calls need a
+maintainer ComfyUI restart once.
+
+Test files: `tests/browser/workspace-tagging.spec.mjs` (the fixture is
+`tests/browser/workspace_tagging_fixture.py`: two files sharing one prompt and
+pair, an equal-byte copy, a video, an unlinked file, the fake model with `hf`, and
+no model files; every test asserts no download request),
+`tests/js/workspace-image-tagging.test.js`,
+`tests/js/workspace-image-tagging-state.test.js`,
+`tests/js/workspace-images-lifecycle.test.js`, and for D88
+`tests/js/workspace-file-inspector.test.js`,
+`tests/js/workspace-inspection-state.test.js` and `tests/browser/inspector-concept.spec.mjs`
+("image tags saved on the concept image keep the inspector equal to the concept"), and
+for the comparison `tests/browser/tagging-concept.spec.mjs` and
+`tests/js/workspace-images-state.test.js` (the in-place refresh), and on all four
+projects `tests/browser/workspace-tagging-lifecycle.spec.mjs` (leaving the page with
+the dialog open releases the capture with no console error; closing during a save
+releases it after the save and still reports it).
+
+```sh
+node --test tests/js/workspace-image-tagging*.test.js tests/js/workspace-images-lifecycle.test.js
+npm run --prefix tools/vendor test:workflows -- workspace-tagging.spec.mjs tagging-concept.spec.mjs \
+  workspace-vocabulary.spec.mjs images-concept.spec.mjs --project chromium --project touch
+npm run --prefix tools/vendor test:workflows -- workspace-tagging-lifecycle.spec.mjs tagging-concept.spec.mjs \
+  --project firefox --project webkit
+```
+
 ### Prompt library cards and recorded deviations (#249 PR E)
 
 The shelf uses the shared concept toolbar and caption (E1) and the concept's
@@ -1286,6 +1558,83 @@ concept fixture stars the concept's favorites (each other prompt's first image).
   restore. Old backups without the table restore, and backup plus restore keeps
   WAL-resident favorites and closes every handle (`tests/test_database_backup.py`).
 
+Image annotations schema (#249, `database/annotation_schema.py`): the group
+`media_annotation_metadata` (one row: schema version, `image_index_revision`),
+`media_annotation_files` (one private locator per physical file,
+`normcase(join(realpath(root), relative))`, server-side only),
+`media_annotation_bytes` (one record per observed whole-file SHA-256 with the
+`size`/`mtime_ns` verified when those bytes were last seen and the set's
+`tag_revision`) and `media_image_tags` (the ordered tag set of one byte record,
+joined into the global `tags` registry). It installs like favorites: optional in
+old backups, required once current (`annotation_schema_error`, the single
+`require_library_state` flag), validated read-only before and after a SAVEPOINT
+install and on restore, never backfilled and never stat-ing media
+(`tests/test_annotation_schema.py`). Writers (`database/annotations.py`) always
+supply the UUID ids, run in `_pair_transaction` fenced by the storage epoch, and
+replace or union one byte record's set (`tag-conflict` / `tag-limit` with the
+current set, identical retries `replayed`); vocabulary rename, merge and delete
+carry `media_image_tags` through the one `_move_tag_links` loop, and tag rows,
+impact and command receipts report `image_count` / `images` /
+`images_changed` as files (`tests/test_annotations.py`,
+`tests/test_pair_tag_maintenance.py`, `tests/test_vocabulary.py`). Cheap reads
+(spec amendment S1) use a byte record only while the file's current
+`(size, mtime_ns)` equals its verified fingerprint; an in-place rewrite that
+preserves both keeps the old set matching until the file is next observed. The
+Images Tags filter (a file matches when every tag is on its linked positive or
+its current image tag set) and `q` search stream the tagged records once and
+stat only the listed files whose records carry a filter tag or a `q` hit; the
+inspector's `image_tags` facet (`current` / `none` / `unavailable`) uses the
+inspection's own stable stat (`tests/test_api_workspace_images.py`
+`TestWorkspaceImageTags`, `tests/test_api_prompt_inspector.py`). The
+browser fixtures clear `media_image_tags`, `media_annotation_bytes` and
+`media_annotation_files` on reset and never the metadata row.
+
+Capture store (#249, `utils/media_capture.py`, `tests/test_media_capture.py`):
+one explicit action captures one rooted media file into an owner-only
+`prompt-manager-captures-*` temp directory, streaming it in 1 MiB chunks while
+hashing. Identity is compared like with like: the opened descriptor against
+the path only on the API-stable facts (`size`, `mtime_ns`); the full identity
+(`size`, `mtime_ns`, `ctime_ns`, `dev`/`ino` where meaningful) between the
+`fstat` before and after the copy and between the path stats before and after
+(native Windows CI with Python 3.12/3.13 showed `stat` and `fstat` disagreeing
+on `st_ctime_ns`: change time versus creation time; 3.10/3.11 agree). Any
+disagreement is `CaptureChanged` and the copy is removed. Error messages carry
+`strerror` only, never a path. Aliases (a symlinked
+file or directory, even inside the root), escapes, non-media names and
+directories are `CaptureDenied`; the locator is
+`normcase(join(realpath(root), relative))`, so a symlinked root shares the
+listing's identity. Fixed limits: 256 MiB per file (stat and stream), 8
+completed captures, 512 MiB reserved, 2 capture/verify workers, 900 s idle,
+1800 s hard; admission is nonblocking (`CaptureCapacity`) and lifetimes are
+tested with an injected clock. An acquired capture outlives expiry and release
+until its last use ends. Header-only dimensions (4096 px per side,
+16,777,216 px) and video mark Generate unsupported; `verify` re-hashes the
+current file without a copy. Every handle closes before unlink (a patched
+`os.unlink` asserts it, plus `assert_closed`), and files are unlinked outside
+the store lock. Known limit: on Windows before Python 3.13 `fstat` reports the
+creation time, so a same-size rewrite that restores `mtime_ns` during the copy
+window completes there with a digest of the copied bytes; POSIX and Windows on
+3.13+ detect it through the descriptor's change time (native CI, 3.13 only).
+Vocabulary rename, merge and delete bump the `tag_revision` of every image set
+they change (and `image_index_revision`), so a Tag dialog opened before the
+change gets `tag-conflict` on Save (`tests/test_pair_tag_maintenance.py`).
+A case or Unicode spelling of a listed name (`A.PNG` for `a.png`, an NFD name)
+is an alias too: each path part must be spelled as its directory lists it, so a
+case- or normalization-insensitive filesystem (macOS, Windows) cannot store a
+second locator. The capture preview answers one byte range with 206 (416 for
+an unsatisfiable or multi-range header) inside the capture's pin, and a
+cancelled preview, observe or Generate still closes its handle and ends its pin
+or capture (`tests/test_api_annotations.py` `TestCancellation`,
+`tests/test_model_gate.py`).
+
+Follow-ups (tracked, not accepted): the legacy Auto-tag batch's cleanup
+`unload_model()` in `py/api/autotag_routes.py` is not gated against the
+batch's own in-flight inference when the batch coroutine is cancelled (other
+owners are kept out; the thread may log an error); and `py/autotag.py`,
+`py/api/images.py` and `database/operations.py` are over the 800-line ceiling
+(all three were before #249 PR B), with a split of `py/autotag.py` into
+backends and service planned first.
+
 Budgets: `test_filter_work_is_bounded` always runs. It asserts that a plain first
 page makes 1 link SELECT and a plain cursor page none, that every filter shape
 makes exactly 1, the covering-index plan, and that the per-file values are built
@@ -1556,7 +1905,7 @@ bytes at the same path without changing the immutable receipt.
 The focused acceptance command retains all four projects:
 
 ```sh
-PM_TEST_PORT=8766 npm run --prefix tools/vendor test:workflows -- \
+npm run --prefix tools/vendor test:workflows -- \
   workspace-history.spec.mjs workspace-visual.spec.mjs \
   --project chromium --project touch --project firefox --project webkit
 npm run --prefix tools/vendor test:executions
@@ -1589,8 +1938,8 @@ page lifecycle recovery retain their separate behavioral tests.
 Initial production acceptance requires maintainer inspection of paired PNGs and
 geometry; never update snapshots automatically or mask content to bless a gap.
 Data-dependent text/images/counts and the clearly labeled saved/frozen/current
-distinctions are documented in the comparison helper. The Images card tag
-button (#249 D9) remains outside #209; every prompt block now renders the concept's
+distinctions are documented in the comparison helper. The gallery cards carry the
+concept's Tag button beside Info (#249 C2 retired D9); every prompt block now renders the concept's
 `section.prompt-section` (D22), so the shared role-block spacing, padding and glyph
 gap is closed, and only the production-only surfaces above keep their own rules
 (D48). The
@@ -1598,6 +1947,47 @@ responsive production Viewer keeps its toolbar in the media panel and prompt det
 mobile toolbar collision. These are explicit remaining visual differences, not
 a whole-application or full-gallery/Viewer pixel-parity claim.
 
+Follow-up: monitor video support. The live image monitor (`utils/image_monitor.py`,
+`GalleryConfig.SUPPORTED_EXTENSIONS`) processes images only; videos written by
+ComfyUI savers reach the Library through the admin scan, which reads the shared video
+carrier. Accepting `VIDEO_EXTENSIONS` in the monitor and routing them through
+`read_video_metadata` is a separate, issue-linked change.
+
+Follow-up: local Inspect parity (#249 PR L1). The browser readers
+(`web/js/png-metadata.js`, `web/js/workspace/local-video.mjs`,
+`web/js/workspace/local-projection.mjs`) answer as the server does for the same bytes
+and a matching extension (`tests/test_local_inspect_parity.py`; the real-ffprobe cases
+skip without ffmpeg). The remaining known divergences, checked against ffprobe 9.0.2,
+`read_inspection_file` and `_metadata`, are tracked here, not accepted as parity:
+
+- Media kind comes from the bytes locally and from the extension on the server: a PNG
+  named `.mp4` is read as a PNG locally (format and size shown) while the server probes
+  it with ffprobe (no format, no size); MP4 bytes named `.png` read locally as video
+  while Pillow fails them on the server (`unreadable`). Parity is stated for the same
+  bytes with a matching extension.
+- JPEG dimensions: the local header read stops at 64 KiB, so a JPEG whose SOF follows
+  more than 64 KiB of APP segments (camera, Photoshop and EXIF/XMP workflow savers)
+  shows no dimensions, while Pillow reads them within the server's 8 MiB header bound.
+  The fix is a segment-header walk up to 8 MiB, which needs a plan amendment.
+- Classic QuickTime `moov/udta` text atoms (`©cmt`/`desc` with a 16-bit size and
+  language, outside `meta`) are not read locally; ffprobe reports them as
+  `comment`/`description`, so the server can select that envelope while the local
+  read is `absent`. No ComfyUI saver writes them.
+- Tag-size bounds count raw bytes locally and ffprobe's escaped JSON on the server
+  (first probe 12 MiB + 64 KiB, role-only re-probe 4 MiB + 64 KiB): only tag sets near
+  those bounds and dense with quotes or control characters can make one side re-probe
+  or fail while the other reads.
+- More than 10,000 container elements stop the local read (`limited`, metadata
+  `unreadable`); ffprobe reads such files.
+- An EBML DocType other than `matroska`/`webm` is an invalid container locally; ffmpeg
+  only warns and reads it.
+- Envelope JSON deeper than 1,000 levels is refused locally as Python's default
+  recursion limit refuses it; the server's real limit depends on the interpreter and
+  call stack. A duplicate key in an object enclosing a nest deeper than the running
+  Python's json limit is `malformed` locally and `limited` on that Python.
+- Decimal exponent limits assume CPython's C `_decimal`; on a `_pydecimal` interpreter
+  the server would read the over-limit exponents the local projection refuses.
+
 Preserve raw logs, failed traces, paired PNGs, geometry, runtime versions and
 source hashes before later managed commands clear their output directories.
 Frontend baseline updates are additive through `coverage:update` only after
@@ -1649,7 +2039,7 @@ acceptance; a baseline is added only after the maintainer accepts them.
 Focused commands:
 
 ```sh
-PM_TEST_PORT=8766 npm run --prefix tools/vendor test:workflows -- \
+npm run --prefix tools/vendor test:workflows -- \
   workspace-library-editor.spec.mjs library-editor-visual.spec.mjs \
   --project chromium --project touch --project firefox --project webkit
 npm run --prefix tools/vendor test:comfyui   # library-sender.spec.mjs host regression
@@ -1713,7 +2103,7 @@ legacy admin page. The screen consumes only the `/prompt_manager/vocabulary/*` r
 above and adds no route, so PR B needs no ComfyUI restart: a browser refresh is enough.
 
 Screen: a `+ Create tag` heading action, a Tag vocabulary table (columns Tag, Prompts,
-Images shown as "—" with hidden "Not available yet", and Actions with Browse and
+Images counting the files that carry the tag (`image_count`, #249 C2), and Actions with Browse and
 Rename / merge), and a Categories panel whose category rows end with the Manage
 category row (Review). Lists load 100 rows at a time, append when scrolled into view
 and keep `Load more tags` / `Load more categories` while more remain. Prompts counts
@@ -1749,13 +2139,12 @@ mobile, the concept geometry/token/control comparison and its mutation check),
 
 ```sh
 node --test tests/js/workspace-vocabulary-state.test.js tests/js/workspace-state.test.js
-PM_TEST_PORT=8766 npm run --prefix tools/vendor test:workflows -- \
+npm run --prefix tools/vendor test:workflows -- \
   workspace-vocabulary.spec.mjs vocabulary-visual.spec.mjs focused-node.spec.mjs \
   --project chromium --project touch --project firefox --project webkit
 ```
 
-Recorded deviations from the concept (maintainer acceptance): no Auto-tag images
-button and an Images "—" cell until the image-annotation slice; bounded lists with
+Recorded deviations from the concept (maintainer acceptance): bounded lists with
 Load more; title "Tags"; Manage categories adds a Category select and Preview before
 Apply; "Apply" instead of "Apply in prototype"; Delete tag… inside Rename or merge;
 Browse sets the exact filter; singular "1 positive prompt"; the production list appends

__init__.py

@@ -90,6 +90,7 @@ try:
     host_app = getattr(config.server_instance, "app", None)
     if host_app is not None and hasattr(host_app, "on_cleanup"):
         host_app.on_cleanup.append(api.cleanup_history_reads)
+        host_app.on_cleanup.append(api.cleanup_annotation_captures)
 
 except Exception as e:
     # Log to internal logging system without console spam

database/annotation_schema.py

@@ -0,0 +1,290 @@
+"""Image annotation schema group (#249): file locators, byte records, image tags.
+
+The group installs like ``image_favorites`` and ``saved_views``: optional in old
+backups, required once a database is current. It owns
+
+- ``media_annotation_metadata``: one row, the group's schema version and the
+  image index revision;
+- ``media_annotation_files``: one private canonical locator per physical file
+  (``normcase(join(realpath(root), relative))``, server-side only);
+- ``media_annotation_bytes``: one record per observed whole-file SHA-256 of a
+  file, with the size and ``mtime_ns`` verified when those bytes were last seen
+  and the tag set's revision;
+- ``media_image_tags``: the ordered image tag set of one byte record, joined
+  into the global ``tags`` registry so vocabulary maintenance carries image
+  links like prompt and pair links.
+
+Installation never backfills anything from prompt or pair tags and never
+touches the filesystem. Validation reads only: the catalog objects are compared
+after whitespace compaction, columns, foreign keys and indexes through the
+PRAGMAs, and persisted values in single bounded SQL passes.
+"""
+
+import re
+import sqlite3
+from typing import Optional
+
+from .image_favorites import MAX_KEY_LENGTH
+
+SCHEMA_VERSION = 1
+METADATA_TABLE = "media_annotation_metadata"
+FILES_TABLE = "media_annotation_files"
+BYTES_TABLE = "media_annotation_bytes"
+IMAGE_TAGS_TABLE = "media_image_tags"
+TABLES = (METADATA_TABLE, FILES_TABLE, BYTES_TABLE, IMAGE_TAGS_TABLE)
+MAX_LOCATOR_LENGTH = MAX_KEY_LENGTH
+_TIMESTAMP_BYTES = 64
+
+_TABLE_DDL = {
+    METADATA_TABLE: (
+        f"CREATE TABLE {METADATA_TABLE} (singleton_id INTEGER PRIMARY KEY "
+        "CHECK(singleton_id=1), schema_version INTEGER NOT NULL "
+        f"CHECK(schema_version={SCHEMA_VERSION}), "
+        "image_index_revision INTEGER NOT NULL DEFAULT 0 "
+        "CHECK(image_index_revision>=0))"
+    ),
+    FILES_TABLE: (
+        f"CREATE TABLE {FILES_TABLE} (id TEXT PRIMARY KEY CHECK(length(id)=36), "
+        "locator TEXT NOT NULL UNIQUE, created_at TEXT NOT NULL)"
+    ),
+    BYTES_TABLE: (
+        f"CREATE TABLE {BYTES_TABLE} (id TEXT PRIMARY KEY CHECK(length(id)=36), "
+        f"file_id TEXT NOT NULL REFERENCES {FILES_TABLE}(id) ON DELETE CASCADE, "
+        "sha256 TEXT NOT NULL CHECK(length(sha256)=64 AND sha256 NOT GLOB "
+        "'*[^0-9a-f]*'), size INTEGER NOT NULL CHECK(size>=0), "
+        "mtime_ns INTEGER NOT NULL, verified_at TEXT NOT NULL, "
+        "first_seen_at TEXT NOT NULL, tag_revision INTEGER NOT NULL DEFAULT 0 "
+        "CHECK(tag_revision>=0), UNIQUE(file_id, sha256))"
+    ),
+    IMAGE_TAGS_TABLE: (
+        f"CREATE TABLE {IMAGE_TAGS_TABLE} (byte_id TEXT NOT NULL REFERENCES "
+        f"{BYTES_TABLE}(id) ON DELETE CASCADE, tag_id INTEGER NOT NULL "
+        "REFERENCES tags(id) ON DELETE CASCADE, position INTEGER NOT NULL "
+        "CHECK(position>=0), PRIMARY KEY(byte_id, tag_id))"
+    ),
+}
+_INDEX_DDL = {
+    "idx_media_annotation_bytes_fingerprint": (
+        "CREATE INDEX idx_media_annotation_bytes_fingerprint ON "
+        f"{BYTES_TABLE}(file_id, size, mtime_ns)"
+    ),
+    "idx_media_image_tags_tag": (
+        f"CREATE INDEX idx_media_image_tags_tag ON {IMAGE_TAGS_TABLE}(tag_id)"
+    ),
+    "idx_media_image_tags_position": (
+        "CREATE UNIQUE INDEX idx_media_image_tags_position ON "
+        f"{IMAGE_TAGS_TABLE}(byte_id, position)"
+    ),
+}
+# SQLite names the indexes behind TEXT primary keys and UNIQUE constraints
+# ``sqlite_autoindex_<table>_<n>``; their catalog rows carry no SQL.
+_AUTOINDEXES = {FILES_TABLE: 2, BYTES_TABLE: 2, IMAGE_TAGS_TABLE: 1}
+# (name, declared type, not null, default, primary key position)
+_COLUMNS = {
+    METADATA_TABLE: (
+        ("singleton_id", "INTEGER", 0, None, 1),
+        ("schema_version", "INTEGER", 1, None, 0),
+        ("image_index_revision", "INTEGER", 1, "0", 0),
+    ),
+    FILES_TABLE: (
+        ("id", "TEXT", 0, None, 1),
+        ("locator", "TEXT", 1, None, 0),
+        ("created_at", "TEXT", 1, None, 0),
+    ),
+    BYTES_TABLE: (
+        ("id", "TEXT", 0, None, 1),
+        ("file_id", "TEXT", 1, None, 0),
+        ("sha256", "TEXT", 1, None, 0),
+        ("size", "INTEGER", 1, None, 0),
+        ("mtime_ns", "INTEGER", 1, None, 0),
+        ("verified_at", "TEXT", 1, None, 0),
+        ("first_seen_at", "TEXT", 1, None, 0),
+        ("tag_revision", "INTEGER", 1, "0", 0),
+    ),
+    IMAGE_TAGS_TABLE: (
+        ("byte_id", "TEXT", 1, None, 1),
+        ("tag_id", "INTEGER", 1, None, 2),
+        ("position", "INTEGER", 1, None, 0),
+    ),
+}
+# (column, parent table, parent column, on delete)
+_FOREIGN_KEYS = {
+    METADATA_TABLE: set(),
+    FILES_TABLE: set(),
+    BYTES_TABLE: {("file_id", FILES_TABLE, "id", "CASCADE")},
+    IMAGE_TAGS_TABLE: {
+        ("byte_id", BYTES_TABLE, "id", "CASCADE"),
+        ("tag_id", "tags", "id", "CASCADE"),
+    },
+}
+
+
+def _compact(sql: str) -> str:
+    return re.sub(r"\s+", "", sql).lower()
+
+
+def _uuid(column: str) -> str:
+    """SQL: *column* is a lowercase hyphenated 36-character UUID string."""
+    return (
+        f"typeof({column}) = 'text' AND length({column}) = 36 "
+        f"AND substr({column}, 9, 1) = '-' AND substr({column}, 14, 1) = '-' "
+        f"AND substr({column}, 19, 1) = '-' AND substr({column}, 24, 1) = '-' "
+        f"AND length(replace({column}, '-', '')) = 32 "
+        f"AND replace({column}, '-', '') NOT GLOB '*[^0-9a-f]*'"
+    )
+
+
+def _timestamp(column: str) -> str:
+    return (
+        f"typeof({column}) = 'text' "
+        f"AND length(CAST({column} AS BLOB)) <= {_TIMESTAMP_BYTES}"
+    )
+
+
+def _non_negative(column: str) -> str:
+    return f"typeof({column}) = 'integer' AND {column} >= 0"
+
+
+# Every row of a table must satisfy its predicate; one pass each, LIMIT 1.
+_ROW_PREDICATES = {
+    FILES_TABLE: " AND ".join(
+        (
+            _uuid("id"),
+            "typeof(locator) = 'text'",
+            # UTF-8 is at most 4 bytes per character; the decoded length of
+            # the few candidates above the byte bound is checked in Python.
+            f"length(CAST(locator AS BLOB)) <= {MAX_LOCATOR_LENGTH * 4}",
+            _timestamp("created_at"),
+        )
+    ),
+    BYTES_TABLE: " AND ".join(
+        (
+            _uuid("id"),
+            _uuid("file_id"),
+            "typeof(sha256) = 'text' AND length(sha256) = 64 "
+            "AND sha256 NOT GLOB '*[^0-9a-f]*'",
+            _non_negative("size"),
+            "typeof(mtime_ns) = 'integer'",
+            _timestamp("verified_at"),
+            _timestamp("first_seen_at"),
+            _non_negative("tag_revision"),
+        )
+    ),
+    IMAGE_TAGS_TABLE: " AND ".join(
+        (_uuid("byte_id"), "typeof(tag_id) = 'integer'", _non_negative("position"))
+    ),
+}
+
+
+def _expected_catalog() -> set:
+    objects = {("table", name, _compact(sql)) for name, sql in _TABLE_DDL.items()}
+    objects |= {("index", name, _compact(sql)) for name, sql in _INDEX_DDL.items()}
+    objects |= {
+        ("index", f"sqlite_autoindex_{table}_{n}", "")
+        for table, count in _AUTOINDEXES.items()
+        for n in range(1, count + 1)
+    }
+    return objects
+
+
+def _catalog(conn: sqlite3.Connection) -> list:
+    names = TABLES + tuple(_INDEX_DDL)
+    marks = ",".join("?" * len(names))
+    table_marks = ",".join("?" * len(TABLES))
+    return conn.execute(  # only "?" placeholders are interpolated
+        "SELECT type,name,sql FROM sqlite_master "
+        f"WHERE name IN ({marks}) OR tbl_name IN ({table_marks})",  # nosec B608
+        names + TABLES,
+    ).fetchall()
+
+
+def _structure_error(conn: sqlite3.Connection) -> Optional[str]:
+    for table, expected in _COLUMNS.items():
+        columns = tuple(
+            (row[1], row[2].upper(), row[3], row[4], row[5])
+            for row in conn.execute(f"PRAGMA table_info({table})")
+        )
+        if columns != expected:
+            return f"Image annotations table {table} has invalid columns"
+        foreign_keys = {
+            (row[3], row[2], row[4], row[6])
+            for row in conn.execute(f"PRAGMA foreign_key_list({table})")
+        }
+        if foreign_keys != _FOREIGN_KEYS[table]:
+            return f"Image annotations table {table} has invalid foreign keys"
+        if conn.execute(f"PRAGMA foreign_key_check({table})").fetchone():
+            return f"Image annotations table {table} contains foreign key violations"
+    return None
+
+
+def _value_error(conn: sqlite3.Connection) -> Optional[str]:
+    # Only this module's constants are interpolated into the SQL below.
+    rows = conn.execute(
+        "SELECT COUNT(*), MIN(typeof(singleton_id) = 'integer' "  # nosec B608
+        "AND singleton_id = 1 AND typeof(schema_version) = 'integer' "
+        f"AND schema_version = {SCHEMA_VERSION} "
+        f"AND {_non_negative('image_index_revision')}) FROM {METADATA_TABLE}"
+    ).fetchone()
+    if rows[0] != 1 or rows[1] != 1:
+        return "Image annotations metadata is invalid"
+    for table, predicate in _ROW_PREDICATES.items():
+        if conn.execute(
+            f"SELECT 1 FROM {table} WHERE NOT ({predicate}) LIMIT 1"  # nosec B608
+        ).fetchone():
+            return f"Image annotations table {table} contains invalid values"
+    # length() stops at an embedded NUL, so the SQL pass bounds bytes and the
+    # few candidates over the character bound are measured decoded here.
+    for (raw,) in conn.execute(
+        f"SELECT CAST(locator AS BLOB) FROM {FILES_TABLE} "  # nosec B608
+        f"WHERE length(CAST(locator AS BLOB)) > {MAX_LOCATOR_LENGTH}"
+    ):
+        if len(bytes(raw).decode("utf-8", "replace")) > MAX_LOCATOR_LENGTH:
+            return f"Image annotations table {FILES_TABLE} contains invalid values"
+    return None
+
+
+def annotation_schema_error(
+    conn: sqlite3.Connection, *, allow_missing: bool = True
+) -> Optional[str]:
+    """Reject partial or unsupported annotation schemas and invalid rows.
+
+    Reads only: nothing is ever repaired. A database without any of the
+    group's objects is acceptable for backups that predate it unless
+    ``allow_missing`` is false.
+    """
+    objects = _catalog(conn)
+    if not objects:
+        return None if allow_missing else "Missing image annotation schema"
+    present = {name for kind, name, _sql in objects if kind == "table"}
+    if present != set(TABLES):
+        return "Incomplete image annotation schema"
+    actual = {(kind, name, _compact(sql or "")) for kind, name, sql in objects}
+    if actual != _expected_catalog():
+        return "Image annotations has an invalid schema definition"
+    return _structure_error(conn) or _value_error(conn)
+
+
+def migrate_annotation_state(conn: sqlite3.Connection) -> None:
+    """Install the group atomically; validate before and after creating it."""
+    conn.execute("SAVEPOINT media_annotation_schema")
+    try:
+        error = annotation_schema_error(conn)
+        if error:
+            raise sqlite3.DatabaseError(error)
+        for sql in _TABLE_DDL.values():
+            conn.execute(sql.replace("CREATE TABLE", "CREATE TABLE IF NOT EXISTS", 1))
+        for sql in _INDEX_DDL.values():
+            conn.execute(sql.replace("INDEX", "INDEX IF NOT EXISTS", 1))
+        conn.execute(
+            f"INSERT OR IGNORE INTO {METADATA_TABLE} "
+            "(singleton_id, schema_version, image_index_revision) VALUES (1, ?, 0)",
+            (SCHEMA_VERSION,),
+        )
+        error = annotation_schema_error(conn, allow_missing=False)
+        if error:
+            raise sqlite3.DatabaseError(error)
+        conn.execute("RELEASE media_annotation_schema")
+    except BaseException:
+        conn.execute("ROLLBACK TO media_annotation_schema")
+        conn.execute("RELEASE media_annotation_schema")
+        raise

database/annotations.py

@@ -0,0 +1,461 @@
+"""Per-image tag sets (#249): byte-exact storage in the shared vocabulary.
+
+A tag set belongs to one byte record (``media_annotation_bytes``): the
+whole-file SHA-256 of one physical file, owned by the file's private locator
+(``media_annotation_files``). Callers pass verified facts (digest, size,
+``mtime_ns``); nothing here reads the filesystem or resolves an owner through
+prompt links, so tagging one image never changes another image, a prompt or a
+pair.
+
+Commands run in one ``_pair_transaction`` fenced by the storage epoch, like the
+native Library and vocabulary commands. A save is a full replacement guarded
+by the set's ``tag_revision``; the identical set sent again (a lost-response
+retry, or the current set at its own revision) answers ``replayed: true`` and
+writes nothing. ``add_image_tags`` unions into the existing set, keeping the
+stored spelling. Names enter the global ``tags`` registry exactly as spelled,
+so vocabulary rename, merge and delete carry image links with prompt and pair
+links (``PromptDatabase._move_tag_links``).
+
+Cheap reads (spec amendment S1) treat a byte record as *current* only while
+the file's present ``(size, mtime_ns)`` equals the record's verified
+fingerprint; a file without a current record shows no image tags.
+"""
+
+import re
+import sqlite3
+from contextlib import contextmanager
+from typing import Iterator
+from uuid import uuid4
+
+from .annotation_schema import (
+    BYTES_TABLE,
+    FILES_TABLE,
+    IMAGE_TAGS_TABLE,
+    MAX_LOCATOR_LENGTH,
+    METADATA_TABLE,
+)
+from .library_commands import _UUID, LibraryCommandError, _validate_epoch
+from .models import utc_now_iso
+from .vocabulary import validate_vocabulary_tag_name
+
+MAX_IMAGE_TAGS = 100
+_LOOKUP_CHUNK = 500
+_DIGEST = re.compile(r"[0-9a-f]{64}")
+_MAX_INT = 2**63 - 1
+_MIN_INT = -(2**63)
+# Only this module's table constants are interpolated into the SQL below.
+_OWNER = f"SELECT id FROM {FILES_TABLE} WHERE locator=?"  # nosec B608
+_INSERT_OWNER = (
+    f"INSERT INTO {FILES_TABLE}(id,locator,created_at) VALUES(?,?,?)"  # nosec B608
+)
+_BYTE = (
+    f"SELECT id,tag_revision,size,mtime_ns FROM {BYTES_TABLE} "  # nosec B608
+    "WHERE file_id=? AND sha256=?"
+)
+_INSERT_BYTE = (
+    f"INSERT INTO {BYTES_TABLE}(id,file_id,sha256,size,mtime_ns,"  # nosec B608
+    "verified_at,first_seen_at,tag_revision) VALUES(?,?,?,?,?,?,?,0)"
+)
+_REFRESH_BYTE = (
+    f"UPDATE {BYTES_TABLE} SET size=?,mtime_ns=?,verified_at=? WHERE id=?"  # nosec B608
+)
+_REFRESH_BY_LOCATOR = (
+    f"UPDATE {BYTES_TABLE} SET size=?,mtime_ns=?,verified_at=? "  # nosec B608
+    f"WHERE sha256=? AND file_id=({_OWNER})"
+)
+_BYTE_TAGS = (
+    f"SELECT t.name FROM {IMAGE_TAGS_TABLE} mit "  # nosec B608
+    "JOIN tags t ON t.id=mit.tag_id WHERE mit.byte_id=? ORDER BY mit.position"
+)
+_CLEAR_TAGS = f"DELETE FROM {IMAGE_TAGS_TABLE} WHERE byte_id=?"  # nosec B608
+_INSERT_TAG = (
+    f"INSERT INTO {IMAGE_TAGS_TABLE}(byte_id,tag_id,position) "  # nosec B608
+    "VALUES(?,?,?)"
+)
+_BUMP_REVISION = (
+    f"UPDATE {BYTES_TABLE} SET tag_revision=tag_revision+1 WHERE id=?"  # nosec B608
+)
+_REVISION = f"SELECT tag_revision FROM {BYTES_TABLE} WHERE id=?"  # nosec B608
+_BUMP_INDEX = (
+    f"UPDATE {METADATA_TABLE} SET "  # nosec B608
+    "image_index_revision=image_index_revision+1 WHERE singleton_id=1"
+)
+_READ = (
+    "SELECT b.id,b.tag_revision,"  # nosec B608
+    f"(SELECT COUNT(*) FROM {BYTES_TABLE} o WHERE o.file_id=f.id AND o.sha256!=?) "
+    f"FROM {FILES_TABLE} f LEFT JOIN {BYTES_TABLE} b "
+    "ON b.file_id=f.id AND b.sha256=? WHERE f.locator=?"
+)
+_FINGERPRINTS = (
+    "SELECT f.locator,b.id,b.size,b.mtime_ns,b.verified_at "  # nosec B608
+    f"FROM {FILES_TABLE} f JOIN {BYTES_TABLE} b ON b.file_id=f.id "
+    "WHERE f.locator IN ({marks})"
+)
+_TAGS_OF = (
+    f"SELECT mit.byte_id,t.name FROM {IMAGE_TAGS_TABLE} mit "  # nosec B608
+    "JOIN tags t ON t.id=mit.tag_id WHERE mit.byte_id IN ({marks}) "
+    "ORDER BY mit.byte_id,mit.position"
+)
+_RECORDS = (
+    "SELECT f.locator,b.id,b.size,b.mtime_ns,t.name "  # nosec B608
+    f"FROM {FILES_TABLE} f JOIN {BYTES_TABLE} b ON b.file_id=f.id "
+    f"LEFT JOIN {IMAGE_TAGS_TABLE} mit ON mit.byte_id=b.id "
+    "LEFT JOIN tags t ON t.id=mit.tag_id "
+    "ORDER BY f.locator,b.first_seen_at,b.id,mit.position"
+)
+
+
+class AnnotationCommandError(LibraryCommandError):
+    """An image tag command lost a race or hit the set's bound.
+
+    ``code`` is ``tag-conflict`` (the set changed since it was read) or
+    ``tag-limit`` (the union would exceed ``MAX_IMAGE_TAGS``); ``current`` is
+    the stored set ``{sha256, tags, tag_revision}`` read inside the
+    transaction, so the caller can show it.
+    """
+
+    def __init__(self, code: str, epoch: str, current: dict | None = None):
+        super().__init__(code, epoch)
+        self.current = current
+
+
+class ImageTagLimitError(ValueError):
+    """More than ``MAX_IMAGE_TAGS`` distinct names in one set."""
+
+
+def normalize_image_tags(names: object) -> list[str]:
+    """Trim and validate each name, fold case-insensitive repeats, cap at 100.
+
+    The first occurrence keeps its spelling and position. Raises ``ValueError``
+    for anything but a list of valid vocabulary names, and its subclass
+    ``ImageTagLimitError`` for more than ``MAX_IMAGE_TAGS`` distinct names.
+    """
+    if not isinstance(names, list):
+        raise ValueError("tags must be a list of strings")
+    result: list[str] = []
+    seen: set[str] = set()
+    for value in names:
+        name = validate_vocabulary_tag_name(value)
+        key = name.casefold()
+        if key in seen:
+            continue
+        seen.add(key)
+        result.append(name)
+    if len(result) > MAX_IMAGE_TAGS:
+        raise ImageTagLimitError(f"An image can carry at most {MAX_IMAGE_TAGS} tags")
+    return result
+
+
+def _validate_locator(locator: object) -> None:
+    if not isinstance(locator, str) or not 0 < len(locator) <= MAX_LOCATOR_LENGTH:
+        raise ValueError(
+            f"locator must be a nonblank path of at most {MAX_LOCATOR_LENGTH} chars"
+        )
+
+
+def _validate_digest(sha256: object) -> None:
+    if not isinstance(sha256, str) or not _DIGEST.fullmatch(sha256):
+        raise ValueError("sha256 must be 64 lowercase hexadecimal characters")
+
+
+def _validate_int(value: object, name: str, minimum: int) -> None:
+    if type(value) is not int or not minimum <= value <= _MAX_INT:
+        raise ValueError(f"{name} must be an integer not below {minimum}")
+
+
+def _validate_target(locator: object, sha256: object, size: object, mtime_ns: object):
+    _validate_locator(locator)
+    _validate_digest(sha256)
+    _validate_int(size, "size", 0)
+    _validate_int(mtime_ns, "mtime_ns", _MIN_INT)
+
+
+def _validate_fingerprint(locator: object, fingerprint: object) -> None:
+    _validate_locator(locator)
+    if (
+        not isinstance(fingerprint, tuple)
+        or len(fingerprint) != 2
+        or any(type(value) is not int for value in fingerprint)
+    ):
+        raise ValueError("stats must map locators to (size, mtime_ns) integer pairs")
+
+
+def _new_id() -> str:
+    """A canonical UUID; the schema's CHECK lets NULL through, this does not."""
+    value = str(uuid4())
+    if not _UUID.fullmatch(value):
+        raise ValueError("Annotation ids must be canonical UUIDs")
+    return value
+
+
+def _receipt(epoch: str, sha256: str, tags: list, revision: int, replayed: bool):
+    return {
+        "library_epoch": epoch,
+        "sha256": sha256,
+        "tags": list(tags),
+        "tag_revision": revision,
+        "replayed": replayed,
+    }
+
+
+class AnnotationOperations:
+    """PromptDatabase mixin: image tag sets, fingerprints and cheap reads.
+
+    ``_`` helpers take the open ``_pair_transaction`` connection.
+    """
+
+    @contextmanager
+    def _annotation_read(self) -> Iterator[sqlite3.Connection]:
+        """One deferred read transaction, so paired SELECTs see one snapshot."""
+        conn = self.model.get_connection()
+        if conn.in_transaction:
+            raise sqlite3.ProgrammingError(
+                "Annotation reads require their own snapshot"
+            )
+        with conn:
+            conn.execute("BEGIN")
+            yield conn
+
+    @staticmethod
+    def _annotation_owner(conn: sqlite3.Connection, locator: str, now: str) -> str:
+        row = conn.execute(_OWNER, (locator,)).fetchone()
+        if row is not None:
+            return row[0]
+        file_id = _new_id()
+        conn.execute(_INSERT_OWNER, (file_id, locator, now))
+        return file_id
+
+    @staticmethod
+    def _annotation_byte(
+        conn: sqlite3.Connection,
+        file_id: str,
+        sha256: str,
+        size: int,
+        mtime_ns: int,
+        now: str,
+    ) -> tuple[str, int, bool, bool]:
+        """The byte record for these bytes, refreshed or created.
+
+        Returns ``(id, tag_revision, created, refingerprinted)``; the last is
+        True when an existing record's verified ``(size, mtime_ns)`` changed,
+        which moves its currency for cheap reads (the caller bumps the index
+        revision when nothing else does).
+        """
+        row = conn.execute(_BYTE, (file_id, sha256)).fetchone()
+        if row is None:
+            byte_id = _new_id()
+            conn.execute(
+                _INSERT_BYTE, (byte_id, file_id, sha256, size, mtime_ns, now, now)
+            )
+            return byte_id, 0, True, False
+        conn.execute(_REFRESH_BYTE, (size, mtime_ns, now, row[0]))
+        return row[0], row[1], False, (row[2], row[3]) != (size, mtime_ns)
+
+    @staticmethod
+    def _byte_tags(conn: sqlite3.Connection, byte_id: str) -> list[str]:
+        return [row[0] for row in conn.execute(_BYTE_TAGS, (byte_id,))]
+
+    def _write_image_tags(
+        self, conn: sqlite3.Connection, byte_id: str, tags: list[str]
+    ) -> int:
+        """Rewrite the junction in order, bump both revisions; return the new one."""
+        tag_ids = self._ensure_tags(conn, tags)
+        conn.execute(_CLEAR_TAGS, (byte_id,))
+        conn.executemany(
+            _INSERT_TAG,
+            [(byte_id, tag_ids[name], position) for position, name in enumerate(tags)],
+        )
+        conn.execute(_BUMP_REVISION, (byte_id,))
+        conn.execute(_BUMP_INDEX)
+        return conn.execute(_REVISION, (byte_id,)).fetchone()[0]
+
+    def _current_set(self, conn, sha256, byte_id, revision) -> dict:
+        return {
+            "sha256": sha256,
+            "tags": self._byte_tags(conn, byte_id),
+            "tag_revision": revision,
+        }
+
+    def read_image_tags(self, locator: str, sha256: str) -> dict:
+        """The stored set of these exact bytes, or an empty set at revision 0.
+
+        ``prior_count`` is the number of other byte records at the locator
+        (earlier or later bytes of the same file).
+        """
+        _validate_locator(locator)
+        _validate_digest(sha256)
+        with self._annotation_read() as conn:
+            row = conn.execute(_READ, (sha256, sha256, locator)).fetchone()
+            if row is None:
+                return {"tags": [], "tag_revision": 0, "prior_count": 0}
+            tags = self._byte_tags(conn, row[0]) if row[0] is not None else []
+            return {"tags": tags, "tag_revision": row[1] or 0, "prior_count": row[2]}
+
+    def replace_image_tags(
+        self,
+        *,
+        locator: str,
+        sha256: str,
+        size: int,
+        mtime_ns: int,
+        expected_epoch: str,
+        expected_tag_revision: int,
+        tags: list,
+    ) -> dict:
+        """Replace the set of these exact bytes, guarded by epoch and revision.
+
+        Creates the owner and byte record when the bytes are new (an explicit
+        empty set is a real save). The identical set at the current revision,
+        or one behind it (a lost-response retry), is ``replayed`` and writes
+        nothing; any other revision mismatch is ``tag-conflict`` with the
+        current set and no writes.
+        """
+        _validate_target(locator, sha256, size, mtime_ns)
+        _validate_epoch(expected_epoch)
+        _validate_int(expected_tag_revision, "expected_tag_revision", 0)
+        tags = normalize_image_tags(tags)
+        with self._pair_transaction() as conn:
+            epoch = self._library_epoch(conn, expected_epoch)
+            now = utc_now_iso()
+            file_id = self._annotation_owner(conn, locator, now)
+            byte_id, revision, created, refingerprinted = self._annotation_byte(
+                conn, file_id, sha256, size, mtime_ns, now
+            )
+            current = self._byte_tags(conn, byte_id)
+            if (
+                not created
+                and current == tags
+                and expected_tag_revision in (revision, revision - 1)
+            ):
+                if refingerprinted:
+                    conn.execute(_BUMP_INDEX)
+                return _receipt(epoch, sha256, current, revision, True)
+            if expected_tag_revision != revision:
+                raise AnnotationCommandError(
+                    "tag-conflict",
+                    epoch,
+                    {"sha256": sha256, "tags": current, "tag_revision": revision},
+                )
+            revision = self._write_image_tags(conn, byte_id, tags)
+            return _receipt(epoch, sha256, tags, revision, False)
+
+    def add_image_tags(
+        self,
+        *,
+        locator: str,
+        sha256: str,
+        size: int,
+        mtime_ns: int,
+        expected_epoch: str,
+        tags: list,
+    ) -> dict:
+        """Union ``tags`` into the set of these exact bytes, keeping stored spelling.
+
+        ``added`` counts the names that were new (case-insensitively). A union
+        over ``MAX_IMAGE_TAGS`` is ``tag-limit`` with the current set and no
+        writes; a union that adds nothing to an existing record is ``replayed``.
+        """
+        _validate_target(locator, sha256, size, mtime_ns)
+        _validate_epoch(expected_epoch)
+        tags = normalize_image_tags(tags)
+        with self._pair_transaction() as conn:
+            epoch = self._library_epoch(conn, expected_epoch)
+            now = utc_now_iso()
+            file_id = self._annotation_owner(conn, locator, now)
+            byte_id, revision, created, refingerprinted = self._annotation_byte(
+                conn, file_id, sha256, size, mtime_ns, now
+            )
+            current = self._byte_tags(conn, byte_id)
+            present = {name.casefold() for name in current}
+            added = [name for name in tags if name.casefold() not in present]
+            merged = current + added
+            if len(merged) > MAX_IMAGE_TAGS:
+                raise AnnotationCommandError(
+                    "tag-limit",
+                    epoch,
+                    {"sha256": sha256, "tags": current, "tag_revision": revision},
+                )
+            if not created and not added:
+                if refingerprinted:
+                    conn.execute(_BUMP_INDEX)
+                return {**_receipt(epoch, sha256, current, revision, True), "added": 0}
+            revision = self._write_image_tags(conn, byte_id, merged)
+            return {
+                **_receipt(epoch, sha256, merged, revision, False),
+                "added": len(added),
+            }
+
+    def refresh_image_fingerprint(
+        self, locator: str, sha256: str, size: int, mtime_ns: int
+    ) -> bool:
+        """Record that these exact bytes were just seen with this stat.
+
+        True when a byte record was refreshed; nothing is created and no tag
+        or tag revision changes. The image index revision moves because the
+        record's currency for cheap reads may have changed.
+        """
+        _validate_target(locator, sha256, size, mtime_ns)
+        with self.model.get_connection() as conn:
+            cursor = conn.execute(
+                _REFRESH_BY_LOCATOR, (size, mtime_ns, utc_now_iso(), sha256, locator)
+            )
+            if cursor.rowcount > 0:
+                conn.execute(_BUMP_INDEX)
+            return cursor.rowcount > 0
+
+    def current_image_tags(self, stats: dict) -> dict:
+        """Image tags of the files whose current ``(size, mtime_ns)`` is verified.
+
+        ``stats`` maps locators to their present fingerprint. A locator is
+        answered (possibly with an empty set) only when a byte record carries
+        exactly that verified fingerprint; the most recently verified record
+        wins when several do. Lookups go in chunks of 500 locators.
+        """
+        if not isinstance(stats, dict):
+            raise ValueError(
+                "stats must map locators to (size, mtime_ns) integer pairs"
+            )
+        for locator, fingerprint in stats.items():
+            _validate_fingerprint(locator, fingerprint)
+        locators = list(stats)
+        chosen: dict[str, tuple[tuple[str, str], str]] = {}
+        result: dict[str, list[str]] = {}
+        with self._annotation_read() as conn:
+            for start in range(0, len(locators), _LOOKUP_CHUNK):
+                chunk = locators[start : start + _LOOKUP_CHUNK]
+                marks = ",".join("?" * len(chunk))
+                for locator, byte_id, size, mtime_ns, verified_at in conn.execute(
+                    _FINGERPRINTS.format(marks=marks), chunk
+                ):
+                    if (size, mtime_ns) != stats[locator]:
+                        continue
+                    key = (verified_at, byte_id)
+                    if locator not in chosen or key > chosen[locator][0]:
+                        chosen[locator] = (key, byte_id)
+            owners = {byte_id: locator for locator, (_key, byte_id) in chosen.items()}
+            result = {locator: [] for locator in chosen}
+            byte_ids = list(owners)
+            for start in range(0, len(byte_ids), _LOOKUP_CHUNK):
+                chunk = byte_ids[start : start + _LOOKUP_CHUNK]
+                marks = ",".join("?" * len(chunk))
+                for byte_id, name in conn.execute(_TAGS_OF.format(marks=marks), chunk):
+                    result[owners[byte_id]].append(name)
+        return result
+
+    def iter_image_tag_records(self) -> Iterator[tuple[str, int, int, list[str]]]:
+        """Stream every byte record as ``(locator, size, mtime_ns, tags)``.
+
+        One SELECT ordered by locator, then by the bytes' first sighting, with
+        each set's names in position order.
+        """
+        conn = self.model.get_connection()
+        current: tuple | None = None
+        for locator, byte_id, size, mtime_ns, name in conn.execute(_RECORDS):
+            if current is None or current[0] != byte_id:
+                if current is not None:
+                    yield current[1]
+                current = (byte_id, (locator, size, mtime_ns, []))
+            if name is not None:
+                current[1][3].append(name)
+        if current is not None:
+            yield current[1]

database/models.py

@@ -17,6 +17,7 @@ from typing import Optional, Set, Tuple
 
 from .library_schema import library_schema_error, migrate_library_state
 from .image_favorites import image_favorites_schema_error, migrate_image_favorites
+from .annotation_schema import annotation_schema_error, migrate_annotation_state
 from .saved_views import migrate_saved_views, saved_views_schema_error
 from .execution_history_schema import (
     history_schema_error,
@@ -302,6 +303,7 @@ class PromptModel:
         migrate_library_state(conn)
         migrate_image_favorites(conn)
         migrate_saved_views(conn)
+        migrate_annotation_state(conn)
 
     def _migrate_prompt_pairs(self, conn: sqlite3.Connection) -> None:
         conn.execute("SAVEPOINT prompt_pair_schema")
@@ -1021,9 +1023,9 @@ class PromptModel:
         Args:
             path: File to inspect
             require_library_state: Also require the tables every current
-                database has (Library state, image favorites and saved
-                views), as a migrated restore stage must; old backups may
-                lack them.
+                database has (Library state, image favorites, saved views
+                and image annotations), as a migrated restore stage must;
+                old backups may lack them.
 
         Returns:
             (ok, reason): reason is "ok" when the file is acceptable, otherwise
@@ -1088,6 +1090,11 @@ class PromptModel:
             )
             if views_error:
                 return False, views_error
+            annotation_error = annotation_schema_error(
+                conn, allow_missing=not require_library_state
+            )
+            if annotation_error:
+                return False, annotation_error
             return True, "ok"
         except sqlite3.Error as e:
             return False, f"Invalid SQLite database: {e}"

database/operations.py

@@ -12,6 +12,7 @@ from .models import PromptModel, normalize_image_path, utc_now_iso
 from .pairs import PairDependencyError, PairOperations
 from .library import LibraryOperations, titled
 from .image_favorites import ImageFavoriteOperations
+from .annotations import AnnotationOperations
 from .saved_views import SavedViewOperations
 from .library_commands import LibraryCommandOperations
 from .vocabulary import VocabularyOperations
@@ -231,6 +232,23 @@ JOIN prompts p ON p.id = COALESCE(pp.positive_prompt_id, gi.prompt_id)
 WHERE gi.file_path IS NOT NULL AND CASE WHEN {admit} THEN NOT EXISTS (
     SELECT 1 FROM generated_images other
     WHERE other.file_path = gi.file_path AND other.id <> gi.id) ELSE 0 END"""
+# The linked positive of each given path (one unambiguous link, pair scope as
+# in _LINK_FILTER_SQL) with its category; {hit} as in link_filter_query. Used
+# for the few files whose current image tags meet a Tags filter (#249 B3).
+_LINK_FACTS_SQL = """SELECT gi.file_path, p.rating, {hit}, p.id, p.category
+FROM generated_images gi
+LEFT JOIN prompt_pairs pp ON pp.id = (
+    SELECT CASE WHEN COUNT(DISTINCT ip.pair_id) = 1 THEN MIN(ip.pair_id) END
+    FROM image_prompt_pairs ip WHERE ip.image_id = gi.id)
+JOIN prompts p ON p.id = COALESCE(pp.positive_prompt_id, gi.prompt_id)
+WHERE gi.file_path IN ({marks}) AND NOT EXISTS (
+    SELECT 1 FROM generated_images other
+    WHERE other.file_path = gi.file_path AND other.id <> gi.id)"""
+_POSITIVE_TAG_NAMES_SQL = (
+    "SELECT pt.prompt_id, t.name FROM prompt_tags pt JOIN tags t "
+    "ON t.id = pt.tag_id WHERE pt.prompt_id IN ({marks})"
+)
+_FACTS_CHUNK = 500
 # The positive's text, its Library title or one of its tag names contains
 # :pattern (LIKE, so ASCII case is ignored as in Library search). A pair's own
 # title names its negative variant and is never searched (#249 D12).
@@ -322,6 +340,7 @@ class PromptDatabase(
     ExecutionOutputOperations,
     HistoryReadOperations,
     ImageFavoriteOperations,
+    AnnotationOperations,
     SavedViewOperations,
 ):
     """Database operations class for managing prompts."""
@@ -1084,16 +1103,39 @@ class PromptDatabase(
             }
 
     @staticmethod
+    def _bump_image_tag_bytes(conn: sqlite3.Connection, tag_id: int) -> int:
+        """Move the tag revision of every image set that carries *tag_id*.
+
+        Vocabulary rename, merge and delete change what an image's set reads
+        as, so a Tag dialog opened before the change must see ``tag-conflict``
+        on Save instead of recreating the old name. Bumps the image index
+        revision once when any set was affected; returns that count.
+        """
+        cursor = conn.execute(
+            "UPDATE media_annotation_bytes SET tag_revision = tag_revision + 1 "
+            "WHERE id IN (SELECT byte_id FROM media_image_tags WHERE tag_id = ?)",
+            (tag_id,),
+        )
+        if cursor.rowcount > 0:
+            conn.execute(
+                "UPDATE media_annotation_metadata SET "
+                "image_index_revision = image_index_revision + 1 WHERE singleton_id = 1"
+            )
+        return cursor.rowcount
+
+    @classmethod
     def _move_tag_links(
-        conn: sqlite3.Connection, source_id: int, target_id: int
+        cls, conn: sqlite3.Connection, source_id: int, target_id: int
     ) -> None:
-        """Repoint component and pair links from source to target.
+        """Repoint component, pair and image links from source to target.
 
         Owners that already carry the target keep that one link; the source's
-        leftover duplicates are removed. The source registry row is left for the
-        caller to delete.
+        leftover duplicates are removed (an image set keeps the target's
+        position). The source registry row is left for the caller to delete.
+        Image sets that carried the source get a new tag revision first.
         """
-        for junction in ("prompt_tags", "prompt_pair_tags"):
+        cls._bump_image_tag_bytes(conn, source_id)
+        for junction in ("prompt_tags", "prompt_pair_tags", "media_image_tags"):
             conn.execute(
                 f"UPDATE OR IGNORE {junction} SET tag_id = ? WHERE tag_id = ?",
                 (target_id, source_id),
@@ -1146,6 +1188,7 @@ class PromptDatabase(
                 conn.execute("DELETE FROM tags WHERE id = ?", (old_tag_id,))
             else:
                 # Simple rename
+                self._bump_image_tag_bytes(conn, old_tag_id)
                 conn.execute(
                     "UPDATE tags SET name = ? WHERE id = ?", (new_name, old_tag_id)
                 )
@@ -1187,6 +1230,7 @@ class PromptDatabase(
             affected = cursor.fetchone()["c"]
 
             # CASCADE delete handles prompt_tags entries
+            self._bump_image_tag_bytes(conn, tag_id)
             conn.execute("DELETE FROM tags WHERE id = ?", (tag_id,))
             conn.commit()
 
@@ -1239,7 +1283,10 @@ class PromptDatabase(
                 pair_count = conn.execute(
                     "SELECT COUNT(*) FROM prompt_pair_tags WHERE tag_id = ?", (src_id,)
                 ).fetchone()[0]
-                if src_count or pair_count:
+                image_count = conn.execute(
+                    "SELECT COUNT(*) FROM media_image_tags WHERE tag_id = ?", (src_id,)
+                ).fetchone()[0]
+                if src_count or pair_count or image_count:
                     self._move_tag_links(conn, src_id, target_id)
                     tags_merged += 1
                     affected += src_count
@@ -2420,6 +2467,46 @@ class PromptDatabase(
                     matches[path] = (bool(matched), rating, positive)
         return matches
 
+    def get_image_link_facts(
+        self, paths: List[str], q: str = ""
+    ) -> Dict[str, Tuple[bool, Optional[int], int, Optional[str], Tuple[str, ...]]]:
+        """Map each normalized path with one unambiguous link to its positive's
+        (hit, rating, positive id, category, tag names).
+
+        The link is resolved as in get_image_link_matches, and ``hit`` says
+        whether *q* occurs in the positive's text, Library title or a tag
+        name (always true without *q*). Paths go in chunks of 500 through the
+        ``file_path`` index; the tag names are read for those positives only.
+        """
+        hit, params = "1", {}
+        if q:
+            hit = _LINK_FILTER_HIT
+            params["pattern"] = f"%{escape_like(q)}%"
+        rows: Dict[str, Tuple[bool, Optional[int], int, Optional[str]]] = {}
+        names: Dict[int, List[str]] = {}
+        unique = list(dict.fromkeys(paths))
+        with self.model.get_connection() as conn:
+            for start in range(0, len(unique), _FACTS_CHUNK):
+                chunk = unique[start : start + _FACTS_CHUNK]
+                bound = {f"path{i}": path for i, path in enumerate(chunk)}
+                marks = ",".join(f":{key}" for key in bound)
+                sql = _LINK_FACTS_SQL.format(hit=hit, marks=marks)
+                for path, rating, matched, positive, category in conn.execute(
+                    sql, {**params, **bound}
+                ):
+                    rows[path] = (bool(matched), rating, positive, category)
+            positives = sorted({row[2] for row in rows.values()})
+            for start in range(0, len(positives), _FACTS_CHUNK):
+                chunk = positives[start : start + _FACTS_CHUNK]
+                marks = ",".join("?" * len(chunk))
+                for positive, name in conn.execute(
+                    _POSITIVE_TAG_NAMES_SQL.format(marks=marks), chunk
+                ):
+                    names.setdefault(positive, []).append(name)
+        return {
+            path: (*row, tuple(names.get(row[2], ()))) for path, row in rows.items()
+        }
+
     def get_image_paths_matching(
         self, q: str, category: Optional[str], tags: List[str]
     ) -> Set[str]:

database/vocabulary.py

@@ -3,6 +3,9 @@
 A count is the number of eligible Library parents that carry the value
 directly or through one active saved pair, which is exactly what
 ``list_library(tags=[name])`` / ``list_library(category=value)`` return.
+A tag row's ``image_count`` is the number of files (private locators) with any
+byte record carrying the tag (#249), so a file counts once however many of its
+byte versions carry it, and equal bytes at two paths count twice.
 
 Tag commands run in one ``_pair_transaction``: the storage epoch is checked
 first, then the exact ``{id, name}`` identity, and the receipt is read on the
@@ -26,6 +29,11 @@ except ImportError:
     from utils.validators import validate_category, validate_tags
 
 _CATEGORY_DOMAIN = "prompt-manager-vocabulary-category:v1\n"
+# Files carrying a tag through any of their byte records (#249 image tags).
+_IMAGE_FILES = (
+    "SELECT mit.tag_id,COUNT(DISTINCT b.file_id) FROM media_image_tags mit "
+    "JOIN media_annotation_bytes b ON b.id=mit.byte_id"
+)
 
 
 def _tag_rows(page: str) -> str:
@@ -40,9 +48,13 @@ def _tag_rows(page: str) -> str:
         "WHERE pt.tag_id IN (SELECT id FROM page) AND " + _ELIGIBLE + " "
         "UNION SELECT ppt.tag_id,pp.positive_prompt_id FROM prompt_pairs pp "
         "JOIN prompt_pair_tags ppt ON ppt.pair_id=pp.id "
-        "WHERE pp.archived=0 AND ppt.tag_id IN (SELECT id FROM page)) "
-        "SELECT page.id,page.name,COUNT(DISTINCT owners.pid) AS prompt_count "
+        "WHERE pp.archived=0 AND ppt.tag_id IN (SELECT id FROM page)),"
+        "images(tag_id,n) AS (" + _IMAGE_FILES + " "
+        "WHERE mit.tag_id IN (SELECT id FROM page) GROUP BY mit.tag_id) "
+        "SELECT page.id,page.name,COUNT(DISTINCT owners.pid) AS prompt_count,"
+        "COALESCE(MAX(images.n),0) AS image_count "
         "FROM page LEFT JOIN owners ON owners.tag_id=page.id "
+        "LEFT JOIN images ON images.tag_id=page.id "
         "GROUP BY page.id ORDER BY page.id"
     )
 
@@ -61,7 +73,10 @@ _TAG_IMPACT = (
     "AS negative_variants,"
     "(SELECT COUNT(*) FROM prompt_pairs pp JOIN prompt_pair_tags ppt "
     "ON ppt.pair_id=pp.id WHERE ppt.tag_id=:tag AND pp.archived!=0) "
-    "AS archived_negative_variants"
+    "AS archived_negative_variants,"
+    "(SELECT COUNT(DISTINCT b.file_id) FROM media_image_tags mit "
+    "JOIN media_annotation_bytes b ON b.id=mit.byte_id WHERE mit.tag_id=:tag) "
+    "AS images"
 )
 _MAX_ID = 2**63 - 1
 
@@ -205,7 +220,16 @@ class VocabularyOperations:
             "SELECT COUNT(DISTINCT pair_id) FROM prompt_pair_tags WHERE tag_id=?",
             (tag_id,),
         ).fetchone()[0]
-        return {"components_changed": components, "pairs_changed": pairs}
+        images = conn.execute(
+            "SELECT COUNT(DISTINCT b.file_id) FROM media_image_tags mit "
+            "JOIN media_annotation_bytes b ON b.id=mit.byte_id WHERE mit.tag_id=?",
+            (tag_id,),
+        ).fetchone()[0]
+        return {
+            "components_changed": components,
+            "pairs_changed": pairs,
+            "images_changed": images,
+        }
 
     def create_vocabulary_tag(self, name: str, *, expected_epoch: str) -> dict:
         """Insert an unused registry row, or reuse the exact existing name."""
@@ -250,6 +274,7 @@ class VocabularyOperations:
                 "SELECT id FROM tags WHERE name=?", (new_name,)
             ).fetchone()
             if target is None:
+                self._bump_image_tag_bytes(conn, tag_id)
                 conn.execute("UPDATE tags SET name=? WHERE id=?", (new_name, tag_id))
                 result_id = tag_id
             else:
@@ -275,6 +300,7 @@ class VocabularyOperations:
             epoch = self._library_epoch(conn, expected_epoch)
             self._vocabulary_identity(conn, epoch, tag_id, name)
             counts = self._tag_owner_counts(conn, tag_id)
+            self._bump_image_tag_bytes(conn, tag_id)
             conn.execute("DELETE FROM tags WHERE id=?", (tag_id,))
             return {"library_epoch": epoch, "deleted": True, **counts}
 

docs/prompt-inspector.md

@@ -29,6 +29,12 @@ Each role has `status` (`resolved`, `absent`, `unresolved`, or `ambiguous`), `te
 (string only when resolved, otherwise null), and `source`. Inspection does not
 index files, alter saved prompts, run scans, or decode image pixels. PNG text and
 video probe output are bounded; unsupported or malformed evidence remains
-unresolved rather than becoming sendable placeholder text.
+unresolved rather than becoming sendable placeholder text. Videos are read from
+one container carrier: the per-key tags written by core SaveVideo/SaveWEBM and
+current VideoHelperSuite (`prompt`, `workflow`, `parameters`,
+`comfyassets_prompt_pairs`), else the old VideoHelperSuite `comment` envelope,
+else `description`. The admin scan reads the same carrier, so Images and
+ingestion agree on a video's prompts (see TESTING.md for the order and budgets;
+the live monitor processes images only).
 
 Python route changes require restarting ComfyUI. Browser assets require a refresh.

py/api/__init__.py

@@ -26,6 +26,7 @@ from .prompts import PromptRoutesMixin
 from .pairs import PairRoutesMixin
 from .library import LibraryRoutesMixin
 from .vocabulary import VocabularyRoutesMixin
+from .annotations import AnnotationRoutesMixin
 from .saved_views import SavedViewRoutesMixin
 from .history import HistoryRoutesMixin
 from .images import ImageRoutesMixin
@@ -252,6 +253,7 @@ class PromptManagerAPI(
     PairRoutesMixin,
     LibraryRoutesMixin,
     VocabularyRoutesMixin,
+    AnnotationRoutesMixin,
     SavedViewRoutesMixin,
     HistoryRoutesMixin,
     PromptRoutesMixin,
@@ -285,6 +287,10 @@ class PromptManagerAPI(
         self.db = db if db is not None else PromptDatabase()
         self._history_read_fence = None
         self._history_reads_closed = False
+        # Private media captures of the annotation routes (py/api/annotations.py),
+        # created on first use with their restore subscription.
+        self._captures = None
+        self._captures_unsubscribe = None
         self._cached_output_dir = None  # Lazy-cached by _find_comfyui_output_dir()
         self._html_cache = {}  # Cached HTML file contents keyed by path
         self._gallery_cache = {}  # dict: path_str -> (files, timestamp)
@@ -441,6 +447,7 @@ class PromptManagerAPI(
         self._register_pair_routes(routes)
         self._register_library_routes(routes)
         self._register_vocabulary_routes(routes)
+        self._register_annotation_routes(routes)
         self._register_saved_view_routes(routes)
         self._register_history_routes(routes)
         self._register_prompt_routes(routes)

py/api/annotation_suggest.py

@@ -0,0 +1,121 @@
+"""Request validation and the worker of the Tag dialog's Generate (#249 B2).
+
+``POST /prompt_manager/image-annotations/suggest`` asks one of legacy's four
+engines for the top N tags of a captured image. Options are validated
+strictly, per engine: WD14 classifiers take two thresholds, the language
+models (GGUF and HuggingFace JoyCaption) take the instructions; every engine
+takes the suggestion count and the GPU switch. Nothing here touches the
+filesystem or the database.
+"""
+
+import math
+
+WD14_MODELS = frozenset({"wd14-swinv2", "wd14-vit"})
+LLM_MODELS = frozenset({"gguf", "hf"})
+MODEL_TYPES = WD14_MODELS | LLM_MODELS
+
+MAX_INSTRUCTION_CHARS = 4000
+MAX_INSTRUCTION_BYTES = 16 * 1024
+
+SUGGEST_FIELDS = frozenset(
+    {"target", "capture_token", "expected_epoch", "model_type", "options"}
+)
+_COMMON = frozenset({"suggestion_limit", "use_gpu"})
+_WD14_REQUIRED = frozenset({"general_threshold", "character_threshold"})
+_LLM_REQUIRED = frozenset({"instructions"})
+
+UNAVAILABLE_HINT = "Download it from the legacy Auto-tag page."
+
+
+class OptionError(ValueError):
+    """A suggest request whose model type or options are not acceptable."""
+
+
+def _threshold(name: str, value) -> float:
+    if type(value) not in (int, float) or not math.isfinite(value):
+        raise OptionError(f"{name} must be a finite number")
+    if not 0 <= value <= 1:
+        raise OptionError(f"{name} must be between 0 and 1")
+    return float(value)
+
+
+def _instructions(value) -> str:
+    if not isinstance(value, str) or not value.strip():
+        raise OptionError("instructions must be non-empty text")
+    if len(value) > MAX_INSTRUCTION_CHARS:
+        raise OptionError(
+            f"instructions must be at most {MAX_INSTRUCTION_CHARS} characters"
+        )
+    try:
+        encoded = value.encode("utf-8")
+    except UnicodeEncodeError:
+        raise OptionError("instructions must be valid Unicode text") from None
+    if len(encoded) > MAX_INSTRUCTION_BYTES:
+        raise OptionError("instructions are too long")
+    return value
+
+
+def validate_options(model_type, options, *, default_limit: int, max_limit: int):
+    """Normalised ``(model_type, options)`` or :class:`OptionError`.
+
+    The answer echoes every option the request will run with, the defaults
+    (suggestion count from the settings, GPU on) filled in.
+    """
+    if not isinstance(model_type, str) or model_type not in MODEL_TYPES:
+        raise OptionError("model_type must be one of " + ", ".join(sorted(MODEL_TYPES)))
+    if not isinstance(options, dict):
+        raise OptionError("options must be an object")
+    required = _WD14_REQUIRED if model_type in WD14_MODELS else _LLM_REQUIRED
+    keys = set(options)
+    if not required <= keys or not keys <= required | _COMMON:
+        raise OptionError(
+            f"options for {model_type} take exactly "
+            + ", ".join(sorted(required))
+            + " and optionally "
+            + ", ".join(sorted(_COMMON))
+        )
+    limit = options.get("suggestion_limit", default_limit)
+    if type(limit) is not int or not 1 <= limit <= max_limit:
+        raise OptionError(f"suggestion_limit must be an integer from 1 to {max_limit}")
+    use_gpu = options.get("use_gpu", True)
+    if not isinstance(use_gpu, bool):
+        raise OptionError("use_gpu must be true or false")
+    normalised = {"use_gpu": use_gpu, "suggestion_limit": limit}
+    if model_type in WD14_MODELS:
+        for name in sorted(_WD14_REQUIRED):
+            normalised[name] = _threshold(name, options[name])
+    else:
+        normalised["instructions"] = _instructions(options["instructions"])
+    return model_type, normalised
+
+
+def unavailable_message(model_name: str) -> str:
+    return f"{model_name} is not downloaded. {UNAVAILABLE_HINT}"
+
+
+def run_suggestions(service, capture_path: str, model_type: str, options: dict):
+    """Blocking: load (or swap to) *model_type*, then suggest for the capture.
+
+    The caller holds the model gate for the whole call; nothing is ever
+    downloaded (``load_model`` refuses a model that is not on disk).
+    """
+    if not service.is_model_loaded() or service.get_loaded_model_type() != model_type:
+        service.load_model(model_type, options["use_gpu"])
+    return service.generate_tag_suggestions(
+        capture_path,
+        limit=options["suggestion_limit"],
+        prompt=options.get("instructions"),
+        general_threshold=options.get("general_threshold"),
+        character_threshold=options.get("character_threshold"),
+    )
+
+
+def suggestion_rows(pairs) -> list:
+    """``[{name, confidence}]``; a non-finite score is reported as unknown."""
+    rows = []
+    for name, score in pairs:
+        confidence = None
+        if score is not None and math.isfinite(score):
+            confidence = float(score)
+        rows.append({"name": str(name), "confidence": confidence})
+    return rows

py/api/annotations.py

@@ -0,0 +1,759 @@
+"""Image annotation routes: observe, capture preview and byte-exact tag saves.
+
+Every route names its file through the listing identity ``{root_index,
+root_scope, path}`` and checks the scope against the current gallery roots
+before any file access. Observe copies the file into a private capture (the
+token the preview and Save refer to), Save re-hashes the current file and
+writes only when the bytes are still the captured ones. Responses carry no
+server path and no locator.
+"""
+
+import asyncio
+import datetime
+import functools
+import os
+import re
+from pathlib import Path
+
+from aiohttp import web
+
+from .annotation_suggest import (
+    SUGGEST_FIELDS,
+    OptionError,
+    run_suggestions,
+    suggestion_rows,
+    unavailable_message,
+    validate_options,
+)
+from .autotag_routes import model_busy_response
+from .images import _forbidden, _has_traversal, _unavailable_link, _verified_title
+from .prompts import bad_request
+from .. import model_gate
+
+try:
+    from ...database.annotations import (
+        MAX_IMAGE_TAGS,
+        AnnotationCommandError,
+        ImageTagLimitError,
+        normalize_image_tags,
+    )
+    from ...database.execution_history import subscribe_history_restore
+    from ...database.library_commands import LibraryCommandError
+    from ...utils.media_capture import CAPTURE_EXTENSIONS, CaptureError, CaptureStore
+except ImportError:
+    from database.annotations import (
+        MAX_IMAGE_TAGS,
+        AnnotationCommandError,
+        ImageTagLimitError,
+        normalize_image_tags,
+    )
+    from database.execution_history import subscribe_history_restore
+    from database.library_commands import LibraryCommandError
+    from utils.media_capture import CAPTURE_EXTENSIONS, CaptureError, CaptureStore
+
+# Request bodies of the annotation routes are small (a target plus a tag set).
+MAX_BODY_BYTES = 64 * 1024
+MAX_TARGET_PATH = 1024
+CAPTURE_URL = "/prompt_manager/image-annotations/capture/"
+_PREVIEW_CHUNK = 64 * 1024
+
+_TARGET_FIELDS = frozenset({"root_index", "root_scope", "path"})
+_OBSERVE_FIELDS = frozenset({"target"})
+_SAVE_FIELDS = frozenset(
+    {
+        "target",
+        "capture_token",
+        "sha256",
+        "expected_epoch",
+        "expected_tag_revision",
+        "tags",
+    }
+)
+_HEX64 = re.compile(r"[0-9a-f]{64}")
+_UUID = re.compile(r"[0-9a-f]{8}(?:-[0-9a-f]{4}){3}-[0-9a-f]{12}")
+# secrets.token_urlsafe(32) is 43 characters; the bound only rejects junk.
+_TOKEN = re.compile(r"[A-Za-z0-9_-]{1,128}")
+
+_PREVIEW_TYPES = {
+    ".png": "image/png",
+    ".jpg": "image/jpeg",
+    ".jpeg": "image/jpeg",
+    ".webp": "image/webp",
+    ".gif": "image/gif",
+    ".bmp": "image/bmp",
+    ".tiff": "image/tiff",
+    ".tif": "image/tiff",
+    ".mp4": "video/mp4",
+    ".m4v": "video/mp4",
+    ".webm": "video/webm",
+    ".mov": "video/quicktime",
+    ".mkv": "video/x-matroska",
+    ".avi": "video/x-msvideo",
+    ".wmv": "video/x-ms-wmv",
+}
+
+_STATUS = {
+    "review-capacity": 429,
+    "unsupported-media": 415,
+    "file-changed": 409,
+    "source-changed": 409,
+    "storage-changed": 409,
+    "tag-conflict": 409,
+    "capture-expired": 409,
+    "model-unavailable": 409,
+    "tag-limit": 400,
+    "metadata-conflict": 409,
+    "vocabulary-conflict": 409,
+}
+_MESSAGES = {
+    "source-changed": (
+        "The source changed. Refresh the image collection before tagging this file."
+    ),
+    "file-changed": "The file changed since it was previewed. Reopen the Tag dialog.",
+    "review-capacity": "Too many previews are open. Close a Tag dialog and try again.",
+    "storage-changed": "Prompt storage was restored; reopen the Tag dialog.",
+    "tag-conflict": (
+        "The tags of this image changed since the dialog opened; review them first."
+    ),
+    "capture-expired": "The preview expired. Open the Tag dialog again.",
+    "tag-limit": f"An image can carry at most {MAX_IMAGE_TAGS} tags.",
+}
+# Any other storage code (a conflict a later command adds) still answers a
+# coded envelope rather than a bare 500.
+_DEFAULT_STATUS = 409
+_DEFAULT_MESSAGE = "The tags could not be saved. Reopen the Tag dialog and try again."
+_OTHER_FILE = "The preview belongs to another file. Open the Tag dialog again."
+
+
+def preview_content_type(extension: str) -> str:
+    """The media type the preview is served as; never sniffed from bytes."""
+    return _PREVIEW_TYPES.get(extension.lower(), "application/octet-stream")
+
+
+def _error(status: int, code: str, message: str, **extra) -> web.Response:
+    return web.json_response(
+        {"success": False, "code": code, "error": message, **extra},
+        status=status,
+        headers={"Cache-Control": "no-store"},
+    )
+
+
+def _coded(code: str, message: str | None = None, **extra) -> web.Response:
+    status = _STATUS.get(code, _DEFAULT_STATUS)
+    return _error(
+        status, code, message or _MESSAGES.get(code, _DEFAULT_MESSAGE), **extra
+    )
+
+
+def _open_capture(path: str):
+    """Worker thread: the capture file opened for reading."""
+    return open(path, "rb")
+
+
+async def _settled(future, cleanup):
+    """Await an executor *future*; a cancelled caller still cleans up after it.
+
+    The future is shielded, so the worker call always runs to its end. When the
+    caller is cancelled first, ``cleanup(result)`` runs once the call is done
+    (``cleanup(None)`` when it failed): a handle is closed, a capture released.
+    """
+    try:
+        return await asyncio.shield(future)
+    except asyncio.CancelledError:
+        future.add_done_callback(functools.partial(_cleanup_result, cleanup))
+        raise
+
+
+def _cleanup_result(cleanup, future) -> None:
+    if future.cancelled() or future.exception() is not None:
+        cleanup(None)
+    else:
+        cleanup(future.result())
+
+
+def _submit(func, *args, **kwargs):
+    return asyncio.get_running_loop().run_in_executor(
+        None, functools.partial(func, *args, **kwargs)
+    )
+
+
+def _capture_error(exc: CaptureError) -> web.Response:
+    """The envelope for a capture outcome; messages never name a path."""
+    if exc.code not in _STATUS:
+        return _error(500, exc.code, "The file could not be captured.")
+    return _coded(exc.code, _MESSAGES.get(exc.code) or str(exc))
+
+
+def _target(data: object):
+    """``((index, scope, relative), None)`` for a target, else ``(None, error)``.
+
+    Only the listing identity is accepted: an int root index, the root's
+    64-hex scope and a bounded root-relative POSIX media path, never an
+    absolute, drive-qualified, backslash or ``..`` path. Nothing here touches
+    the filesystem or the database.
+    """
+    if not isinstance(data, dict) or set(data) != _TARGET_FIELDS:
+        return None, bad_request("target must name root_index, root_scope and path")
+    index, scope, relative = data["root_index"], data["root_scope"], data["path"]
+    if type(index) is not int or index < 0:
+        return None, bad_request("Invalid gallery root")
+    if not isinstance(scope, str) or _HEX64.fullmatch(scope) is None:
+        return None, bad_request("Invalid gallery root scope")
+    if not isinstance(relative, str) or not relative or len(relative) > MAX_TARGET_PATH:
+        return None, bad_request("path must be a relative media path")
+    parts = relative.split("/")
+    if (
+        "\0" in relative
+        or "\\" in relative
+        or relative.startswith("/")
+        or _has_traversal(relative)
+        or any(part in ("", ".", "..") or ":" in part for part in parts)
+    ):
+        return None, _forbidden()
+    if os.path.splitext(parts[-1])[1].lower() not in CAPTURE_EXTENSIONS:
+        return None, _forbidden("Only media files can be tagged")
+    return (index, scope, relative), None
+
+
+def _expires_at(seconds: float | None) -> str:
+    moment = datetime.datetime.now(datetime.timezone.utc) + datetime.timedelta(
+        seconds=max(0.0, seconds or 0.0)
+    )
+    return moment.replace(microsecond=0).isoformat().replace("+00:00", "Z")
+
+
+def _capture_info(store: CaptureStore, capture) -> dict:
+    return {
+        "token": capture.token,
+        "expires_at": _expires_at(store.expires_in(capture.token)),
+        "preview_url": CAPTURE_URL + capture.token,
+        "sha256": capture.sha256,
+        "size": capture.size,
+        "width": capture.width,
+        "height": capture.height,
+        "media_type": capture.media_type,
+        "generate": {
+            "supported": capture.inferable,
+            "reason": capture.inferable_reason,
+        },
+    }
+
+
+class AnnotationRoutesMixin:
+    """Observe, capture preview and tag save routes over PR A's storage."""
+
+    def _register_annotation_routes(self, routes: web.RouteTableDef) -> None:
+        routes.post("/prompt_manager/image-annotations/observe")(self.observe_image)
+        routes.get(CAPTURE_URL + "{token}")(self.get_capture_preview)
+        routes.delete(CAPTURE_URL + "{token}")(self.release_capture)
+        routes.put("/prompt_manager/image-annotations/tags")(self.save_image_tags)
+        routes.post("/prompt_manager/image-annotations/suggest")(
+            self.suggest_image_tags
+        )
+
+    # -- capture store lifetime ------------------------------------------------
+
+    def _capture_store(self) -> CaptureStore:
+        """The process's capture store, created on first use.
+
+        Created lazily so an API object that never observes a file (most tests,
+        a server that never opens the Tag dialog) owns no temp directory. A
+        restore of the database dooms every capture (``invalidate_all``).
+        """
+        if self._captures is None:
+            self._captures = CaptureStore()
+        if self._captures_unsubscribe is None:
+            self._captures_unsubscribe = subscribe_history_restore(
+                self.db.model.db_path, self._invalidate_captures
+            )
+        return self._captures
+
+    def _invalidate_captures(self) -> None:
+        store = self._captures
+        if store is not None:
+            store.invalidate_all()
+
+    def close_annotation_captures(self) -> None:
+        """Remove every capture file and the owned directory. Idempotent."""
+        unsubscribe, self._captures_unsubscribe = self._captures_unsubscribe, None
+        if unsubscribe is not None:
+            unsubscribe()
+        store, self._captures = self._captures, None
+        if store is not None:
+            store.close()
+
+    async def cleanup_annotation_captures(self, app: web.Application) -> None:
+        self.close_annotation_captures()
+
+    # -- shared request steps --------------------------------------------------
+
+    async def _annotation_body(self, request: web.Request, fields: frozenset):
+        data, error = await self._read_json_body(request, max_bytes=MAX_BODY_BYTES)
+        if error is not None:
+            return None, error
+        if set(data) != fields:
+            return None, bad_request(
+                "Request must contain exactly " + ", ".join(sorted(fields))
+            )
+        return data, None
+
+    def _scoped_root(self, index: int, scope: str):
+        """``(root, None)`` when the scope names a current root, else an error."""
+        roots = list(self._get_all_output_dirs())
+        error = self._gallery_scope_error(
+            {"root_scope": scope, "root": str(index)}, roots
+        )
+        if error is not None:
+            if error.status == 409:
+                return None, _coded("source-changed")
+            return None, error
+        return roots[index], None
+
+    # -- observe ---------------------------------------------------------------
+
+    async def observe_image(self, request: web.Request) -> web.Response:
+        """POST ``{target}``: capture the file and answer its saved tag set.
+
+        The capture is the byte-exact copy the preview and Save refer to; the
+        tags are those of exactly these bytes (an empty set at revision 0 for
+        unseen bytes). Observing never writes tags; equal bytes refresh the
+        record's verified fingerprint (S1).
+        """
+        data, error = await self._annotation_body(request, _OBSERVE_FIELDS)
+        if error is not None:
+            return error
+        target, error = _target(data["target"])
+        if error is not None:
+            return error
+        index, scope, relative = target
+        root, error = self._scoped_root(index, scope)
+        if error is not None:
+            return error
+        store = self._capture_store()
+        try:
+            epoch = await self._run_in_executor(self.db.get_execution_epoch)
+            capture = await _settled(
+                _submit(
+                    store.capture,
+                    root,
+                    relative,
+                    epoch=epoch,
+                    scope=scope,
+                    root_index=index,
+                ),
+                # A cancelled observe never hands its capture out: release it.
+                lambda made: made is not None and store.release(made.token),
+            )
+        except CaptureError as exc:
+            return _capture_error(exc)
+        except Exception:
+            self.logger.exception("Image observation failed")
+            return _error(500, "observe-failed", "The image could not be observed.")
+        try:
+            stored, title = await self._run_in_executor(
+                self._observe_capture, store, capture, root, relative
+            )
+        except asyncio.CancelledError:
+            store.release(capture.token)
+            raise
+        except Exception:
+            store.release(capture.token)
+            self.logger.exception("Image observation failed")
+            return _error(500, "observe-failed", "The image could not be observed.")
+        return web.json_response(
+            {
+                "success": True,
+                "library_epoch": capture.epoch,
+                "target": {"root_index": index, "root_scope": scope, "path": relative},
+                "capture": _capture_info(store, capture),
+                "title": title,
+                "filename": relative.rsplit("/", 1)[-1],
+                "tags": stored["tags"],
+                "tag_revision": stored["tag_revision"],
+            },
+            headers={"Cache-Control": "no-store"},
+        )
+
+    def _observe_capture(self, store: CaptureStore, capture, root, relative: str):
+        """Blocking part of observe: fingerprint refresh, tag read, title."""
+        self.db.refresh_image_fingerprint(
+            capture.locator, capture.sha256, capture.size, capture.mtime_ns
+        )
+        stored = self.db.read_image_tags(capture.locator, capture.sha256)
+        return stored, self._observed_title(store, capture, root, relative)
+
+    def _observed_title(self, store: CaptureStore, capture, root, relative: str):
+        """The dialog's title under the card and inspector rule (``_verified_title``).
+
+        A stored Library title always shows; a first-line title only when the
+        captured bytes' own records confirm it. An ambiguous association has no
+        title. Videos read no card metadata, so only a stored title shows.
+        """
+        path = str(Path(root) / relative)
+        summary = self._workspace_link_summaries([path]).get(path)
+        if summary is None:
+            summary = _unavailable_link()
+        positive_title = None
+        if (
+            capture.media_type != "video"
+            and summary.get("title_source") == "text"
+            and store.acquire(capture.token) is not None
+        ):
+            try:
+                facts = self._read_card_facts(Path(capture.path), parse=True)
+            finally:
+                store.release_use(capture.token)
+            positive_title = facts.positive_title if facts is not None else None
+        return _verified_title(summary, positive_title)
+
+    # -- capture preview -------------------------------------------------------
+
+    async def get_capture_preview(self, request: web.Request) -> web.StreamResponse:
+        """GET the captured bytes; the token is pinned while they stream."""
+        token = request.match_info["token"]
+        store = self._captures
+        if store is None or _TOKEN.fullmatch(token) is None:
+            return _error(404, "capture-expired", _MESSAGES["capture-expired"])
+        capture = store.acquire(token)
+        if capture is None:
+            return _error(404, "capture-expired", _MESSAGES["capture-expired"])
+        pinned = True
+        try:
+            _root, error = self._scoped_root(capture.root_index, capture.scope)
+            if error is not None:
+                # The roots changed under the token: it is no longer usable.
+                store.release(token)
+                return error
+            try:
+                handle = await _settled(
+                    _submit(_open_capture, capture.path),
+                    functools.partial(_close_then_release, store, token),
+                )
+            except asyncio.CancelledError:
+                # Cancelled while the file opens: the pin ends only after the
+                # handle closes (Windows cannot unlink an open file).
+                pinned = False
+                raise
+            except OSError:
+                store.release(token)
+                return _error(404, "capture-expired", _MESSAGES["capture-expired"])
+            try:
+                return await self._stream_capture(request, capture, handle)
+            finally:
+                handle.close()
+        finally:
+            if pinned:
+                store.release_use(token)
+
+    @staticmethod
+    async def _stream_capture(request, capture, handle) -> web.StreamResponse:
+        """The capture's bytes, or the one byte range a ``Range`` header asks for.
+
+        The response is written here, inside the caller's pin, so the handle
+        is closed and the pin ended only after the last byte is sent.
+        """
+        size = capture.size
+        headers = {
+            "Content-Type": preview_content_type(Path(capture.path).suffix),
+            "Accept-Ranges": "bytes",
+            "Cache-Control": "no-store",
+            "X-Content-Type-Options": "nosniff",
+        }
+        status, start, end = 200, 0, size
+        if "Range" in request.headers:
+            wanted = _byte_range(request, size)
+            if wanted is None:
+                return web.Response(
+                    status=416,
+                    headers={**headers, "Content-Range": f"bytes */{size}"},
+                )
+            status, (start, end) = 206, wanted
+            headers["Content-Range"] = f"bytes {start}-{end - 1}/{size}"
+        headers["Content-Length"] = str(end - start)
+        response = web.StreamResponse(status=status, headers=headers)
+        await response.prepare(request)
+        loop = asyncio.get_running_loop()
+        if start:
+            await loop.run_in_executor(None, handle.seek, start)
+        remaining = end - start
+        while remaining > 0 and request.method != "HEAD":
+            chunk = await loop.run_in_executor(
+                None, handle.read, min(_PREVIEW_CHUNK, remaining)
+            )
+            if not chunk:
+                break
+            remaining -= len(chunk)
+            await response.write(chunk)
+        await response.write_eof()
+        return response
+
+    async def release_capture(self, request: web.Request) -> web.Response:
+        """DELETE a capture token; releasing an unknown token is a success."""
+        token = request.match_info["token"]
+        store = self._captures
+        if store is not None and _TOKEN.fullmatch(token) is not None:
+            store.release(token)
+        return web.json_response(
+            {"success": True}, headers={"Cache-Control": "no-store"}
+        )
+
+    # -- save ------------------------------------------------------------------
+
+    async def save_image_tags(self, request: web.Request) -> web.Response:
+        """PUT the tag set of the previewed bytes, re-hashing the current file.
+
+        Writes only when the file still holds the captured bytes
+        (``file-changed`` with zero writes otherwise), under the storage epoch
+        and the tag revision the dialog read; an identical retry is
+        ``replayed``. The hash never runs inside a database transaction.
+        """
+        data, error = await self._annotation_body(request, _SAVE_FIELDS)
+        if error is not None:
+            return error
+        target, error = _target(data["target"])
+        if error is not None:
+            return error
+        guards, error = _save_guards(data)
+        if error is not None:
+            return error
+        token, sha256, expected_epoch, revision, tags = guards
+        index, scope, relative = target
+        root, error = self._scoped_root(index, scope)
+        if error is not None:
+            return error
+        store = self._captures
+        capture = store.get(token) if store is not None else None
+        if capture is None:
+            return _coded("capture-expired")
+        if capture.root_index != index or capture.scope != scope:
+            return _coded("capture-expired", _OTHER_FILE)
+        if sha256 != capture.sha256:
+            return bad_request("sha256 does not match the preview")
+        if expected_epoch != capture.epoch:
+            current = await self._run_in_executor(self.db.get_execution_epoch)
+            return _coded("storage-changed", library_epoch=current)
+        try:
+            facts = await self._run_in_executor(
+                _verify_captured_file, store, capture, root, relative
+            )
+        except CaptureError as exc:
+            return _capture_error(exc)
+        if facts is None or facts.locator != capture.locator:
+            return _coded("capture-expired", _OTHER_FILE)
+        if facts.sha256 != capture.sha256:
+            return _coded("file-changed")
+        try:
+            receipt = await self._run_in_executor(
+                self.db.replace_image_tags,
+                locator=facts.locator,
+                sha256=facts.sha256,
+                size=facts.size,
+                mtime_ns=facts.mtime_ns,
+                expected_epoch=expected_epoch,
+                expected_tag_revision=revision,
+                tags=tags,
+            )
+        except AnnotationCommandError as exc:
+            current = exc.current or {}
+            return _coded(
+                exc.code,
+                library_epoch=exc.epoch,
+                tags=current.get("tags", []),
+                tag_revision=current.get("tag_revision", 0),
+            )
+        except LibraryCommandError as exc:
+            return _coded(exc.code, library_epoch=exc.epoch)
+        except ValueError as exc:
+            return bad_request(str(exc))
+        except Exception:
+            self.logger.exception("Image tag save failed")
+            return _error(500, "save-failed", "The tags could not be saved.")
+        return web.json_response(
+            {"success": True, **receipt}, headers={"Cache-Control": "no-store"}
+        )
+
+    # -- suggest ---------------------------------------------------------------
+
+    async def suggest_image_tags(self, request: web.Request) -> web.Response:
+        """POST: the top N tags of the captured bytes from one legacy engine.
+
+        Runs only on the private capture (pinned for the whole inference),
+        under the shared model gate (``model-busy`` while anything else holds
+        it, including a legacy batch), never downloads (``model-unavailable``)
+        and never writes.
+        """
+        from ..autotag import MODELS, get_autotag_service
+        from ..config import MAX_SUGGESTION_LIMIT, AutotagConfig
+
+        data, error = await self._annotation_body(request, SUGGEST_FIELDS)
+        if error is not None:
+            return error
+        target, error = _target(data["target"])
+        if error is not None:
+            return error
+        try:
+            model_type, options = validate_options(
+                data["model_type"],
+                data["options"],
+                default_limit=AutotagConfig.suggestion_limit(),
+                max_limit=MAX_SUGGESTION_LIMIT,
+            )
+        except OptionError as exc:
+            return bad_request(str(exc))
+        token, epoch = data["capture_token"], data["expected_epoch"]
+        if not isinstance(token, str) or _TOKEN.fullmatch(token) is None:
+            return bad_request("Invalid capture token")
+        if not isinstance(epoch, str) or _UUID.fullmatch(epoch) is None:
+            return bad_request("expected_epoch must be a canonical UUID")
+        index, scope, relative = target
+        root, error = self._scoped_root(index, scope)
+        if error is not None:
+            return error
+        store = self._captures
+        capture = store.acquire(token) if store is not None else None
+        if capture is None:
+            return _coded("capture-expired")
+        started = False
+        try:
+            error = await self._suggest_guards(capture, target, root, epoch)
+            if error is not None:
+                return error
+            service = get_autotag_service()
+            status = await self._run_in_executor(service.get_models_status)
+            if not status.get(model_type, {}).get("downloaded"):
+                name = MODELS.get(model_type, {}).get("name", model_type)
+                return _coded("model-unavailable", unavailable_message(name))
+            owner = model_gate.new_owner("image-annotations")
+            if not model_gate.try_acquire(owner):
+                return model_busy_response()
+            try:
+                # From here the pin is the gate's to end: the worker ends it,
+                # or ``on_skipped`` does when the worker never starts.
+                started = True
+                pairs = await model_gate.run(
+                    owner,
+                    _pinned_suggestions,
+                    store,
+                    token,
+                    service,
+                    capture,
+                    model_type,
+                    options,
+                    on_skipped=functools.partial(store.release_use, token),
+                )
+            except Exception:
+                self.logger.exception("Tag suggestion failed")
+                return _error(
+                    500,
+                    "suggest-failed",
+                    "Auto-tagging failed. Check server logs for details.",
+                )
+            finally:
+                model_gate.release(owner)
+        finally:
+            if not started:
+                store.release_use(token)
+        return web.json_response(
+            {
+                "success": True,
+                "model_type": model_type,
+                "options": options,
+                "suggestions": suggestion_rows(pairs),
+            },
+            headers={"Cache-Control": "no-store"},
+        )
+
+    async def _suggest_guards(self, capture, target, root, epoch):
+        """The capture must be this target's, of this epoch, and inferable."""
+        index, scope, relative = target
+        expected = os.path.normcase(
+            os.path.join(os.path.realpath(root), *relative.split("/"))
+        )
+        if (
+            capture.root_index != index
+            or capture.scope != scope
+            or capture.locator != expected
+        ):
+            return _coded("capture-expired", _OTHER_FILE)
+        if epoch != capture.epoch:
+            current = await self._run_in_executor(self.db.get_execution_epoch)
+            return _coded("storage-changed", library_epoch=current)
+        if not capture.inferable:
+            return _coded("unsupported-media", capture.inferable_reason)
+        return None
+
+
+def _close_then_release(store: CaptureStore, token: str, handle) -> None:
+    """Close a preview handle a cancelled request opened, then end its pin."""
+    try:
+        if handle is not None:
+            handle.close()
+    finally:
+        store.release_use(token)
+
+
+def _byte_range(request: web.Request, size: int):
+    """``(start, end)`` of the one satisfiable byte range asked for, else None.
+
+    A malformed, multi-range or unsatisfiable header is None (416), as
+    aiohttp's own file responses answer it.
+    """
+    try:
+        wanted = request.http_range
+    except ValueError:
+        return None
+    start, stop = wanted.start, wanted.stop
+    if start is None:
+        return None
+    if start < 0:  # a suffix: the last -start bytes
+        start = max(0, size + start)
+        stop = size
+    stop = size if stop is None else min(stop, size)
+    if start >= size or stop <= start:
+        return None
+    return start, stop
+
+
+def _pinned_suggestions(store, token, service, capture, model_type, options):
+    """Worker thread: infer on the capture, then end its pin (even on error)."""
+    try:
+        return run_suggestions(service, capture.path, model_type, options)
+    finally:
+        store.release_use(token)
+
+
+def _verify_captured_file(store: CaptureStore, capture, root, relative: str):
+    """Re-hash the current file, or None when the token names another file.
+
+    The cheap locator comparison runs first, so a token sent for another
+    target never costs a whole-file hash.
+    """
+    expected = os.path.normcase(
+        os.path.join(os.path.realpath(root), *relative.split("/"))
+    )
+    if expected != capture.locator:
+        return None
+    return store.verify(root, relative)
+
+
+def _save_guards(data: dict):
+    """Validated ``(token, sha256, epoch, revision, tags)`` or ``(None, error)``."""
+    token = data["capture_token"]
+    if not isinstance(token, str) or _TOKEN.fullmatch(token) is None:
+        return None, bad_request("Invalid capture token")
+    sha256 = data["sha256"]
+    if not isinstance(sha256, str) or _HEX64.fullmatch(sha256) is None:
+        return None, bad_request("sha256 must be 64 lowercase hexadecimal characters")
+    epoch = data["expected_epoch"]
+    if not isinstance(epoch, str) or _UUID.fullmatch(epoch) is None:
+        return None, bad_request("expected_epoch must be a canonical UUID")
+    revision = data["expected_tag_revision"]
+    if type(revision) is not int or revision < 0:
+        return None, bad_request("expected_tag_revision must be a non-negative integer")
+    try:
+        tags = normalize_image_tags(data["tags"])
+    except ImageTagLimitError as exc:
+        return None, _error(400, "tag-limit", str(exc))
+    except ValueError as exc:
+        return None, bad_request(str(exc))
+    return (token, sha256, epoch, revision, tags), None

py/api/autotag_routes.py

@@ -9,11 +9,24 @@ from typing import Optional
 from urllib.parse import quote
 
 from .prompts import thumbnail_url_for
+from .. import model_gate
 
 from aiohttp import web
 
 _INTERNAL_ERROR = "An internal error occurred. Check server logs for details."
 _BUSY_ERROR = "Auto-tagging is already running"
+MODEL_BUSY_ERROR = (
+    "A tagging model is busy with another operation. Try again when it finishes."
+)
+
+
+def model_busy_response() -> web.Response:
+    """``409 model-busy``: another operation holds the shared model gate."""
+    return web.json_response(
+        {"success": False, "code": "model-busy", "error": MODEL_BUSY_ERROR},
+        status=409,
+    )
+
 
 # Only these may be handed to the tagging engine or listed by the scanner.
 _MEDIA_EXTENSIONS = frozenset(
@@ -86,6 +99,7 @@ class AutotagRoutesMixin:
         """Get status of available AutoTag models."""
         try:
             from ..autotag import get_autotag_service
+            from ..config import AutotagConfig
 
             service = get_autotag_service()
             models_status = service.get_models_status()
@@ -99,6 +113,7 @@ class AutotagRoutesMixin:
                     "loaded_model_type": service.get_loaded_model_type(),
                     "wd14_general_threshold": service.wd14_general_threshold,
                     "wd14_character_threshold": service.wd14_character_threshold,
+                    "suggestion_limit": AutotagConfig.suggestion_limit(),
                 }
             )
 
@@ -138,6 +153,7 @@ class AutotagRoutesMixin:
     async def download_autotag_model(self, request):
         """Download an AutoTag model with streaming progress (POST only)."""
         model_type = request.match_info.get("model_type")
+        owner = model_gate.new_owner("legacy-download")
 
         async def events():
             try:
@@ -160,8 +176,8 @@ class AutotagRoutesMixin:
                 def progress_callback(status: str, progress: float):
                     pass
 
-                success = await self._run_in_executor(
-                    service.download_model, model_type, progress_callback
+                success = await model_gate.run(
+                    owner, service.download_model, model_type, progress_callback
                 )
                 if success:
                     yield {
@@ -178,7 +194,12 @@ class AutotagRoutesMixin:
                     "message": _INTERNAL_ERROR,
                 }
 
-        return await self._stream_sse(request, events())
+        if not model_gate.try_acquire(owner):
+            return model_busy_response()
+        try:
+            return await self._stream_sse(request, events())
+        finally:
+            model_gate.release(owner)
 
     async def _read_start_params(self, request) -> Optional[dict]:
         """Merge query-string and optional JSON-body parameters for autotag/start.
@@ -239,8 +260,9 @@ class AutotagRoutesMixin:
             return
 
         yield {"type": "progress", "progress": 0, "status": "Loading model..."}
+        run = self._gated_runner(opts.get("gate_owner"))
         try:
-            await self._run_in_executor(service.load_model, model_type, True)
+            await run(service.load_model, model_type, True)
         except Exception:
             self.logger.exception("Failed to load model")
             yield {
@@ -338,7 +360,7 @@ class AutotagRoutesMixin:
                     continue
 
                 try:
-                    tags = await self._run_in_executor(
+                    tags = await run(
                         service.generate_tags,
                         allowed_path,
                         general_threshold=opts["general_threshold"],
@@ -423,10 +445,26 @@ class AutotagRoutesMixin:
                 {"success": False, "error": _BUSY_ERROR}, status=409
             )
         try:
-            return await self._stream_sse(request, events())
+            opts["gate_owner"] = model_gate.new_owner("legacy-batch")
+            if not model_gate.try_acquire(opts["gate_owner"]):
+                return model_busy_response()
+            try:
+                return await self._stream_sse(request, events())
+            finally:
+                model_gate.release(opts["gate_owner"])
         finally:
             _autotag_batch_lock.release()
 
+    def _gated_runner(self, owner):
+        """Executor calls under the model gate *owner* holds (plain when None)."""
+        if owner is None:
+            return self._run_in_executor
+
+        async def run(func, *args, **kwargs):
+            return await model_gate.run(owner, func, *args, **kwargs)
+
+        return run
+
     def _resolve_allowed_image_path(self, raw_path: str) -> Optional[str]:
         """Return the real path when it is a media file inside an output directory."""
         if Path(str(raw_path)).suffix.lower() not in _MEDIA_EXTENSIONS:
@@ -549,22 +587,28 @@ class AutotagRoutesMixin:
             from ..autotag import get_autotag_service
 
             service = get_autotag_service()
+            owner = model_gate.new_owner("legacy-single")
+            if not model_gate.try_acquire(owner):
+                return model_busy_response()
+            try:
+                if (
+                    not service.is_model_loaded()
+                    or service.get_loaded_model_type() != model_type
+                ):
+                    await model_gate.run(owner, service.load_model, model_type, use_gpu)
 
-            if (
-                not service.is_model_loaded()
-                or service.get_loaded_model_type() != model_type
-            ):
-                await self._run_in_executor(service.load_model, model_type, use_gpu)
+                if custom_prompt:
+                    service.custom_prompt = custom_prompt
 
-            if custom_prompt:
-                service.custom_prompt = custom_prompt
-
-            tags = await self._run_in_executor(
-                service.generate_tags,
-                image_path,
-                general_threshold=general_threshold,
-                character_threshold=character_threshold,
-            )
+                tags = await model_gate.run(
+                    owner,
+                    service.generate_tags,
+                    image_path,
+                    general_threshold=general_threshold,
+                    character_threshold=character_threshold,
+                )
+            finally:
+                model_gate.release(owner)
 
             return web.json_response(
                 {
@@ -636,14 +680,19 @@ class AutotagRoutesMixin:
             from ..autotag import get_autotag_service
 
             service = get_autotag_service()
+            owner = model_gate.new_owner("legacy-unload")
+            if not model_gate.try_acquire(owner):
+                return model_busy_response()
+            try:
+                if not service.is_model_loaded():
+                    return web.json_response(
+                        {"success": True, "message": "No model was loaded"}
+                    )
 
-            if not service.is_model_loaded():
-                return web.json_response(
-                    {"success": True, "message": "No model was loaded"}
-                )
-
-            model_type = service.get_loaded_model_type()
-            service.unload_model()
+                model_type = service.get_loaded_model_type()
+                service.unload_model()
+            finally:
+                model_gate.release(owner)
 
             return web.json_response(
                 {

py/api/images.py

@@ -14,7 +14,7 @@ import urllib.parse
 from bisect import bisect_left, bisect_right
 from collections.abc import Mapping
 from pathlib import Path
-from typing import NamedTuple
+from typing import Callable, NamedTuple
 
 from aiohttp import web
 from PIL import Image, UnidentifiedImageError
@@ -74,6 +74,7 @@ IMAGE_EXTENSIONS = _IMAGE_EXTENSIONS
 # The header records the card parses (see utils.card_seed.header_params).
 _CARD_RECORDS = ("prompt", "workflow")
 _PNG_SIGNATURE = b"\x89PNG\r\n\x1a\n"
+_LISTING_FAILED = "The image list could not be filtered. Check server logs for details."
 
 
 def _card_header(path: Path) -> tuple:
@@ -418,6 +419,29 @@ def _listed(path: Path, mtime: float, root: Path, index: int) -> _Listed:
     )
 
 
+class _ImageTagMatch(NamedTuple):
+    """How one listed file's current image tags meet the filters (#249 B3)."""
+
+    tags: bool  # every filter tag is in its image tags or linked positive tags
+    query: bool  # ``q`` is a case-folded substring of one image tag name
+    link: tuple | None  # (hit, rating, positive id) of its unambiguous link
+
+
+def _current_fingerprint(path: Path) -> tuple[int, int] | None:
+    """The file's present ``(st_size, st_mtime_ns)``; None when it is gone."""
+    try:
+        observed = os.stat(path)
+    except OSError:
+        return None
+    return observed.st_size, observed.st_mtime_ns
+
+
+def _image_tags_relevant(names: list, wanted: set, query: str) -> bool:
+    if wanted and not wanted.isdisjoint(names):
+        return True
+    return bool(query) and any(query in name.casefold() for name in names)
+
+
 def _workspace_key(entry: tuple, order: str) -> tuple:
     """The keyset sort key of *entry* (a _Listed or a 4-tuple entry)."""
     if not isinstance(entry, _Listed):
@@ -752,6 +776,7 @@ def _filtered_page(
     filters: dict,
     links: dict | None,
     favorites: frozenset | None = None,
+    image_tags: dict | None = None,
 ) -> tuple[int, list, int | None]:
     """Count the *entries* the filters admit and take *size* from *start* on.
 
@@ -761,6 +786,9 @@ def _filtered_page(
     or a linked-positive hit. Returns the total, the page and the number of
     distinct linked positives among all admitted entries (None without links).
     *favorites*, when given, admits only those listed paths (``favorites=1``).
+    *image_tags* maps normalized paths to their _ImageTagMatch: a tag filter
+    also admits a file whose current image tags (with its linked positive's
+    tags) carry every tag, and ``q`` also admits an image tag name hit.
     """
     query, media, folder = filters["query"], filters["media"], filters["folder"]
     members_only = filters["category"] is not None or bool(filters["tags"])
@@ -780,9 +808,15 @@ def _filtered_page(
         if favorites is not None and entry.path not in favorites:
             continue
         match = None if links is None else links.get(entry.normalized)
-        if members_only and match is None:
+        image = None if image_tags is None else image_tags.get(entry.normalized)
+        if members_only and match is None and not (image and image.tags):
             continue
-        if query and query not in entry.folded and not (match and match[0]):
+        if (
+            query
+            and query not in entry.folded
+            and not (match and match[0])
+            and not (image and image.query)
+        ):
             continue
         total += 1
         if match is not None:
@@ -1140,7 +1174,7 @@ class ImageRoutesMixin:
                 }
             )
         except Exception as e:
-            self.logger.error(f"Output images error: {e}")
+            self.logger.error(f"Output images error: {e}", exc_info=True)
             return web.json_response(
                 {"success": False, "error": _safe_error(e)}, status=500
             )
@@ -1184,11 +1218,27 @@ class ImageRoutesMixin:
             for index, root in enumerate(output_dirs)
             if root_index is None or index == root_index
         ]
-        page, total, has_more, next_cursor, positive_count = (
-            await self._run_in_executor(
-                self._select_workspace_page, listings, filters, after, limit, scope
+        try:
+            page, total, has_more, next_cursor, positive_count = (
+                await self._run_in_executor(
+                    self._select_workspace_page, listings, filters, after, limit, scope
+                )
+            )
+        except Exception:
+            # A filter that needs the prompt links or the image tags cannot be
+            # served without them. Never echo the exception: it can carry SQL
+            # or database text. The traceback goes to the server log.
+            self.logger.error(
+                "Workspace image list could not be filtered", exc_info=True
+            )
+            return web.json_response(
+                {
+                    "success": False,
+                    "code": "listing-failed",
+                    "error": _LISTING_FAILED,
+                },
+                status=500,
             )
-        )
         images = await self._run_in_executor(self._format_workspace_page, page)
         result = {
             "success": True,
@@ -1276,6 +1326,80 @@ class ImageRoutesMixin:
                 )
         return orders[order]
 
+    def _workspace_locators(self, listings: list[tuple]) -> dict:
+        """The selected listing's private locators -> _Listed (cached).
+
+        A locator is ``normcase(join(realpath(root), relative))``, the key
+        image tags are stored under (database/annotations.py); the listing
+        never follows symlinks, so it equals the listed file's own locator.
+        """
+        orders = self._workspace_orders(listings)
+        reals = tuple(os.path.realpath(root) for _index, root, _files in listings)
+        key = ("locators", reals)
+        if key not in orders:
+            located = {}
+            for (index, root, files), real in zip(listings, reals):
+                prefix = os.path.join(str(root), "")
+                for entry in self._workspace_listed(index, root, files):
+                    text = str(entry.path)
+                    relative = (
+                        text[len(prefix) :]
+                        if text.startswith(prefix)
+                        else str(entry.path.relative_to(root))
+                    )
+                    located[os.path.normcase(os.path.join(real, relative))] = entry
+            orders[key] = located
+        return orders[key]
+
+    def _image_tag_matches(self, listings: list[tuple], filters: dict) -> dict:
+        """Normalized path -> _ImageTagMatch for files whose image tags count.
+
+        Streams the tagged byte records once and keeps only listed locators
+        carrying a filter tag or a ``q`` hit; only those candidates are
+        stat'ed (never every listed file), and a candidate's tags count only
+        while a byte record is verified at its present ``(size, mtime_ns)``
+        (spec amendment S1). Under a Tags filter the candidates' linked
+        positives are read too, so their tags and category combine.
+        """
+        tags, query, category = filters["tags"], filters["query"], filters["category"]
+        wanted = set(tags)
+        located = self._workspace_locators(listings)
+        candidates = {}
+        records = self.db.iter_image_tag_records()
+        try:
+            for locator, _size, _mtime_ns, names in records:
+                entry = located.get(locator)
+                if entry is not None and locator not in candidates:
+                    if _image_tags_relevant(names, wanted, query):
+                        candidates[locator] = entry
+        finally:
+            records.close()
+        stats = {}
+        for locator, entry in candidates.items():
+            fingerprint = _current_fingerprint(entry.path)
+            if fingerprint is not None:
+                stats[locator] = fingerprint
+        current = self.db.current_image_tags(stats) if stats else {}
+        facts = {}
+        if wanted and current:
+            facts = self.db.get_image_link_facts(
+                [candidates[locator].normalized for locator in current],
+                filters["text"],
+            )
+        matches = {}
+        for locator, names in current.items():
+            entry = candidates[locator]
+            link = facts.get(entry.normalized)
+            carried = set(names).union(link[4] if link else ())
+            matches[entry.normalized] = _ImageTagMatch(
+                bool(wanted)
+                and wanted <= carried
+                and (category is None or (link is not None and link[3] == category)),
+                bool(query) and any(query in name.casefold() for name in names),
+                link[:3] if link else None,
+            )
+        return matches
+
     def _select_workspace_page(
         self,
         listings: list[tuple],
@@ -1328,6 +1452,30 @@ class ImageRoutesMixin:
                     ),
                     exc_info=True,
                 )
+        image_tags = None
+        if filters["query"] or tags:
+            try:
+                image_tags = self._image_tag_matches(listings, filters)
+            except Exception:
+                # A Tags filter cannot be served without them (as without the
+                # links); a search then matches without image tags.
+                if tags:
+                    raise
+                self.logger.warning(
+                    "Workspace image tags could not be read; searching without "
+                    "image tags",
+                    exc_info=True,
+                )
+        if image_tags and links is not None:
+            # A linked file admitted through its image tags keeps its rating
+            # and counts its positive.
+            admitted = {
+                path: match.link
+                for path, match in image_tags.items()
+                if match.tags and match.link is not None and path not in links
+            }
+            if admitted:
+                links = {**links, **admitted}
         if order == "rating":
 
             def rank(entry):
@@ -1360,7 +1508,7 @@ class ImageRoutesMixin:
             self._workspace_favorite_paths(listings) if filters["favorites"] else None
         )
         total, page, positive_count = _filtered_page(
-            entries, start, limit + 1, filters, links, favorites
+            entries, start, limit + 1, filters, links, favorites, image_tags
         )
         if not needed and after is not None:
             positive_count = _COUNT_OMITTED
@@ -2467,6 +2615,9 @@ class ImageRoutesMixin:
                         library=self._inspection_library(
                             saved, unavailable=unavailable
                         ),
+                        image_tags=self._inspection_image_tags(
+                            roots[index], Path(relative).parts
+                        ),
                     )
                 except CurrentFileChanged:
                     return None
@@ -2505,6 +2656,30 @@ class ImageRoutesMixin:
                 {"success": False, "error": "Prompt inspection failed"}, status=500
             )
 
+    def _inspection_image_tags(self, root: Path, parts: tuple) -> Callable:
+        """The inspector's ``image_tags`` reader for one rooted file (S1).
+
+        The returned function takes the inspection's own stable stat and
+        answers ``current`` only for a byte record verified at exactly that
+        ``(size, mtime_ns)``; a failed read is ``unavailable``, never an error.
+        """
+        locator = os.path.normcase(os.path.join(os.path.realpath(root), *parts))
+
+        def read(size: int, mtime_ns: int) -> dict:
+            try:
+                found = self.db.current_image_tags({locator: (size, mtime_ns)})
+            except Exception:
+                self.logger.warning(
+                    "Inspection image tags could not be read", exc_info=True
+                )
+                return {"state": "unavailable", "tags": []}
+            tags = found.get(locator)
+            if tags is None:
+                return {"state": "none", "tags": []}
+            return {"state": "current", "tags": list(tags)}
+
+        return read
+
     def _inspection_library(self, saved: dict | None, *, unavailable: bool) -> dict:
         """The inspector's Library facts; the route sets the title afterwards."""
         result = {

py/api/inspection_service.py

@@ -3,6 +3,7 @@
 import os
 import stat
 from pathlib import Path
+from typing import Callable
 
 try:
     from ...utils.prompt_inspector import (
@@ -93,8 +94,19 @@ def _in_host_temp(path: Path) -> bool:
 
 
 def inspect_current_file(
-    path: Path, *, source_label: str, saved: dict | None, library: dict
+    path: Path,
+    *,
+    source_label: str,
+    saved: dict | None,
+    library: dict,
+    image_tags: Callable[[int, int], dict] | None = None,
 ) -> dict:
+    """Current-file evidence read between two equal stats of *path*.
+
+    *image_tags*, when given, is called once with that stable
+    ``(st_size, st_mtime_ns)`` and its facet is returned as ``image_tags``
+    (spec amendment S1: tags of a byte record verified at exactly that stat).
+    """
     before = _current_file_facts(path)
     try:
         inspection = read_inspection_file(path)
@@ -104,6 +116,9 @@ def inspect_current_file(
     fields = ("st_dev", "st_ino", "st_mode", "st_size", "st_mtime_ns", "st_ctime_ns")
     if any(getattr(before, key) != getattr(after, key) for key in fields):
         raise CurrentFileChanged
+    facets = {}
+    if image_tags is not None:
+        facets["image_tags"] = image_tags(before.st_size, before.st_mtime_ns)
     if inspection is None:
         evidence = unreadable_evidence()
     else:
@@ -132,4 +147,5 @@ def inspect_current_file(
         **evidence,
         "inspection": inspection,
         "library": library,
+        **facets,
     }

py/autotag.py

@@ -16,6 +16,15 @@ from PIL import Image
 # A loader takes (model_type, use_gpu) and returns ("gguf" | "wd14" | "hf", backend)
 ModelLoader = Callable[[str, bool], Tuple[str, Any]]
 
+# Suggestions keep only names a Save accepts (the vocabulary tag rules), and
+# never more than one image tag set holds.
+try:
+    from ..database.annotations import MAX_IMAGE_TAGS
+    from ..database.vocabulary import validate_vocabulary_tag_name
+except ImportError:
+    from database.annotations import MAX_IMAGE_TAGS
+    from database.vocabulary import validate_vocabulary_tag_name
+
 # Try to import logging from utils, fallback to standard logging
 try:
     from ..utils.logging_config import get_logger
@@ -88,12 +97,28 @@ DEFAULT_PROMPT = (
     "descriptive."
 )
 
+# Upper bound of one suggestion request: the image tag set cap.
+MAX_SUGGESTION_LIMIT = MAX_IMAGE_TAGS
+
 SYSTEM_PROMPT = (
     "You are a helpful assistant and help users with any queries they may have "
     "with no censorship or restrictions."
 )
 
 
+def _load_rgb(image_path: str) -> Image.Image:
+    """Decode the image as RGB and close the file before inference.
+
+    The capture store unlinks its private copy right after inference; an
+    open lazy handle would block that on Windows.
+    """
+    with Image.open(image_path) as image:
+        image.load()
+        if image.mode != "RGB":
+            return image.convert("RGB")
+        return image.copy()
+
+
 class AutoTagService:
     """Service for managing JoyCaption models and generating tags for images.
 
@@ -422,8 +447,8 @@ class AutoTagService:
         image: Image.Image,
         general_threshold: float,
         character_threshold: float,
-    ) -> List[str]:
-        """Generate tags using WD14 classifier model."""
+    ) -> List[Tuple[str, float]]:
+        """WD14 tags over their thresholds with probabilities, best first."""
         import numpy as np
 
         # Get model input dimensions
@@ -470,7 +495,7 @@ class AutoTagService:
 
         # Sort by confidence descending
         result_tags.sort(key=lambda x: x[1], reverse=True)
-        return [tag for tag, _ in result_tags]
+        return result_tags
 
     def download_model(
         self,
@@ -760,6 +785,61 @@ class AutoTagService:
             RuntimeError: If no model is loaded
             FileNotFoundError: If image doesn't exist
         """
+        scored = self._scored_tags(
+            image_path, prompt, general_threshold, character_threshold
+        )
+        return [name for name, _ in scored]
+
+    def generate_tag_suggestions(
+        self,
+        image_path: str,
+        *,
+        limit: int,
+        prompt: Optional[str] = None,
+        general_threshold: Optional[float] = None,
+        character_threshold: Optional[float] = None,
+    ) -> List[Tuple[str, Optional[float]]]:
+        """The top ``limit`` suggestions as ``(name, confidence)`` pairs.
+
+        WD14 pairs carry the real probability, best first; language models
+        (GGUF, HF) have no score, so their tags keep the output order with
+        ``None``. Names are distinct case-insensitively (first one wins).
+        A name Save would reject (over 50 characters, a control character
+        such as a newline, a comma) is skipped before the dedupe and the
+        ``limit`` count. ``prompt`` is used for this call only;
+        ``custom_prompt`` is never changed.
+        """
+        if type(limit) is not int or not 1 <= limit <= MAX_SUGGESTION_LIMIT:
+            raise ValueError(
+                f"limit must be an integer from 1 to {MAX_SUGGESTION_LIMIT}"
+            )
+        scored = self._scored_tags(
+            image_path, prompt, general_threshold, character_threshold
+        )
+        suggestions: List[Tuple[str, Optional[float]]] = []
+        seen = set()
+        for raw, score in scored:
+            try:
+                name = validate_vocabulary_tag_name(raw)
+            except ValueError:
+                continue
+            key = name.casefold()
+            if key in seen:
+                continue
+            seen.add(key)
+            suggestions.append((name, score))
+            if len(suggestions) == limit:
+                break
+        return suggestions
+
+    def _scored_tags(
+        self,
+        image_path: str,
+        prompt: Optional[str],
+        general_threshold: Optional[float],
+        character_threshold: Optional[float],
+    ) -> List[Tuple[str, Optional[float]]]:
+        """Run the loaded model: WD14 ``(name, probability)``, LLM ``(name, None)``."""
         if self._tagger is None:
             raise RuntimeError("No model loaded. Call load_model() first.")
 
@@ -767,18 +847,14 @@ class AutoTagService:
             raise FileNotFoundError(errno.ENOENT, "Image not found", str(image_path))
 
         use_prompt = prompt or self._custom_prompt
-
-        # Load image
-        image = Image.open(image_path)
-        if image.mode != "RGB":
-            image = image.convert("RGB")
+        image = _load_rgb(image_path)
 
         # Generate based on model type
         model_type, tagger_obj = self._tagger
 
         if model_type == "gguf":
             raw_tags = self._generate_gguf(tagger_obj, image, use_prompt)
-            return self._parse_tags(raw_tags)
+            return [(tag, None) for tag in self._parse_tags(raw_tags)]
         elif model_type == "wd14":
             session, tags_list = tagger_obj
             gen_thresh = (
@@ -799,7 +875,7 @@ class AutoTagService:
             raw_tags = self._generate_hf(
                 model, processor, device, compute_dtype, image, use_prompt
             )
-            return self._parse_tags(raw_tags)
+            return [(tag, None) for tag in self._parse_tags(raw_tags)]
 
     def _generate_gguf(self, model, image: Image.Image, prompt: str) -> str:
         """Generate tags using GGUF model."""

py/config.py

@@ -465,6 +465,59 @@ class IntegrationConfig:
             cls.CIVITAI_API_KEY = lora["civitai_api_key"]
 
 
+# Number of tag suggestions the Tag dialogs request by default (D68). The cap
+# is the image tag set limit: more suggestions could never all be kept.
+DEFAULT_SUGGESTION_LIMIT = 10
+MAX_SUGGESTION_LIMIT = 100  # pinned to MAX_IMAGE_TAGS by tests/test_config.py
+
+
+class AutotagConfig:
+    """Persisted tagging defaults under ``autotag.defaults`` in config.json.
+
+    ``suggestion_limit`` is how many suggestions Generate returns (the top N
+    by score), an integer in ``1..MAX_SUGGESTION_LIMIT``. A hand-edited value
+    outside that range, or of another type, falls back to the default with a
+    warning rather than failing the load.
+    """
+
+    SUGGESTION_LIMIT = DEFAULT_SUGGESTION_LIMIT
+
+    @classmethod
+    def suggestion_limit(cls) -> int:
+        return cls.SUGGESTION_LIMIT
+
+    @staticmethod
+    def valid_suggestion_limit(value: Any) -> bool:
+        return (
+            type(value) is int and 1 <= value <= MAX_SUGGESTION_LIMIT
+        )  # bool, float and str are rejected
+
+    @classmethod
+    def get_config(cls) -> Dict[str, Any]:
+        return {"defaults": {"suggestion_limit": cls.SUGGESTION_LIMIT}}
+
+    @classmethod
+    def update_config(cls, new_config: Any) -> None:
+        if not isinstance(new_config, dict):
+            config_logger.warning("Ignoring autotag settings that are not an object")
+            return
+        defaults = new_config.get("defaults", {})
+        if not isinstance(defaults, dict):
+            config_logger.warning("Ignoring autotag.defaults that is not an object")
+            return
+        if "suggestion_limit" not in defaults:
+            return
+        value = defaults["suggestion_limit"]
+        if cls.valid_suggestion_limit(value):
+            cls.SUGGESTION_LIMIT = value
+            return
+        config_logger.warning(
+            "autotag.defaults.suggestion_limit must be an integer from 1 to "
+            f"{MAX_SUGGESTION_LIMIT}; using {DEFAULT_SUGGESTION_LIMIT}"
+        )
+        cls.SUGGESTION_LIMIT = DEFAULT_SUGGESTION_LIMIT
+
+
 class PromptManagerConfig:
     """Main configuration class for PromptManager core functionality.
 
@@ -566,6 +619,7 @@ class PromptManagerConfig:
             },
             "gallery": GalleryConfig.get_config(),
             "integrations": IntegrationConfig.get_config(),
+            "autotag": AutotagConfig.get_config(),
         }
 
     @classmethod
@@ -706,6 +760,9 @@ class PromptManagerConfig:
         if "integrations" in new_config:
             IntegrationConfig.update_config(new_config["integrations"])
 
+        if "autotag" in new_config:
+            AutotagConfig.update_config(new_config["autotag"])
+
 
 # Load configuration on import
 try:

py/model_gate.py

@@ -0,0 +1,177 @@
+"""One process-wide, non-blocking gate around the tagging models (#249 B2).
+
+Loading, swapping, running, downloading and unloading a tagging model all
+happen under this gate, whether the caller is a legacy Auto-tag route, the
+Tag dialog's Generate or (later) a bulk job. A caller that cannot take the
+gate is refused at once (``409 model-busy``), never queued.
+
+Work that runs in a worker thread is admitted with :meth:`ModelGate.run`. A
+release requested while such work still runs (a client disconnect cancels
+the waiting coroutine, not the thread) takes effect only when the last
+admitted call really exits, so nobody can swap the model under a running
+inference.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import threading
+import uuid
+from typing import Any, Callable, Optional
+
+
+class GateNotHeld(RuntimeError):
+    """Raised when work is admitted for an owner that does not hold the gate."""
+
+
+class ModelGate:
+    """A single-owner gate whose release waits for admitted worker calls."""
+
+    def __init__(self) -> None:
+        self._lock = threading.Lock()
+        self._holder: Optional[str] = None
+        self._inflight = 0
+        self._releasing = False
+
+    def try_acquire(self, owner: str) -> bool:
+        """Take the gate for *owner*; False at once when anyone holds it."""
+        with self._lock:
+            if self._holder is not None:
+                return False
+            self._holder = owner
+            self._inflight = 0
+            self._releasing = False
+            return True
+
+    def release(self, owner: str) -> None:
+        """Let go; deferred until admitted work of *owner* has exited."""
+        with self._lock:
+            if self._holder != owner:
+                return
+            if self._inflight:
+                self._releasing = True
+            else:
+                self._clear_locked()
+
+    def holder(self) -> Optional[str]:
+        with self._lock:
+            return self._holder
+
+    def begin(self, owner: str) -> None:
+        """Admit one worker call; the holder must still hold the gate."""
+        with self._lock:
+            if self._holder != owner or self._releasing:
+                raise GateNotHeld(f"{owner} does not hold the model gate")
+            self._inflight += 1
+
+    def end(self, owner: str) -> None:
+        """One admitted call exited; completes a deferred release."""
+        with self._lock:
+            self._end_locked(owner)
+
+    async def run(
+        self,
+        owner: str,
+        func: Callable[..., Any],
+        *args,
+        on_skipped: Optional[Callable[[], None]] = None,
+        **kwargs,
+    ):
+        """Run ``func`` in the default executor while *owner* holds the gate.
+
+        The executor future is shielded: cancelling the caller never cancels
+        a call that has not started yet (which would leave it admitted
+        forever) and never ends the admission of a call that is running.
+        ``on_skipped`` runs once when ``func`` never starts (the gate is not
+        held, the executor refuses the call, or a loop shutdown drops it), so
+        a caller can free what ``func`` would have freed.
+        """
+        try:
+            self.begin(owner)
+        except GateNotHeld:
+            if on_skipped is not None:
+                on_skipped()
+            raise
+        state = {"started": False, "finished": False}
+
+        def finish() -> None:
+            with self._lock:
+                if not state["finished"]:
+                    state["finished"] = True
+                    self._end_locked(owner)
+
+        def work():
+            with self._lock:
+                if state["finished"]:
+                    # Already declared skipped (the future was cancelled as
+                    # the thread picked the call up): never run it twice over.
+                    return None
+                state["started"] = True
+            try:
+                return func(*args, **kwargs)
+            finally:
+                finish()
+
+        def never_ran(future) -> None:
+            # Only a loop shutdown cancels the shielded future; a call that
+            # never started must still give its admission back.
+            if future.cancelled():
+                with self._lock:
+                    skipped = not state["started"] and not state["finished"]
+                    if skipped:
+                        state["finished"] = True
+                        self._end_locked(owner)
+                if skipped and on_skipped is not None:
+                    on_skipped()
+
+        try:
+            future = asyncio.get_running_loop().run_in_executor(None, work)
+        except BaseException:
+            finish()
+            if on_skipped is not None:
+                on_skipped()
+            raise
+        future.add_done_callback(never_ran)
+        return await asyncio.shield(future)
+
+    def _end_locked(self, owner: str) -> None:
+        if self._holder != owner or self._inflight == 0:
+            return
+        self._inflight -= 1
+        if self._releasing and self._inflight == 0:
+            self._clear_locked()
+
+    def _clear_locked(self) -> None:
+        self._holder = None
+        self._inflight = 0
+        self._releasing = False
+
+
+_GATE = ModelGate()
+
+
+def new_owner(kind: str) -> str:
+    """A unique owner name such as ``legacy-single:<uuid>``."""
+    return f"{kind}:{uuid.uuid4()}"
+
+
+def try_acquire(owner: str) -> bool:
+    return _GATE.try_acquire(owner)
+
+
+def release(owner: str) -> None:
+    _GATE.release(owner)
+
+
+def holder() -> Optional[str]:
+    return _GATE.holder()
+
+
+async def run(
+    owner: str,
+    func: Callable[..., Any],
+    *args,
+    on_skipped: Optional[Callable[[], None]] = None,
+    **kwargs,
+):
+    return await _GATE.run(owner, func, *args, on_skipped=on_skipped, **kwargs)

tests/browser/activity-concept.spec.mjs

@@ -1,5 +1,5 @@
 import { startCoverage, saveCoverage } from '../../tools/vendor/browser-coverage.mjs';
-import { test, expect } from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import { test, expect } from './fixtures.mjs';
 import { THEMES, seedConcept, openConcept, openProduction, diffInventory, mutate } from './concept-inventory.mjs';
 
 // Activity-1 Task 3. The review dialog is compared with the concept's own `scan-review` ("Review metadata scan",

tests/browser/classic-node.spec.mjs

@@ -1,5 +1,5 @@
 import { readFileSync } from 'node:fs';
-import { test, expect } from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import { test, expect } from './fixtures.mjs';
 import { startCoverage, saveCoverage } from '../../tools/vendor/browser-coverage.mjs';
 
 test.beforeEach(async ({page, browserName}) => {
@@ -57,7 +57,8 @@ test('Classic keeps immediate selection, search query and metadata serialization
 });
 
 test('Classic detaches native text synchronization on removal',async({page})=>{
-    await page.waitForTimeout(60);
+    // The page installs the text sync (serializeValue and its listeners) in a 50 ms timer; wait for it, not for time.
+    await expect.poll(()=>page.evaluate(()=>typeof node.widgets.find(item=>item.name==='text').serializeValue)).toBe('function');
     const values=await page.evaluate(()=>{
         const widget=node.widgets.find(item=>item.name==='text');
         widget.inputEl.value='Before removal';widget.inputEl.dispatchEvent(new Event('input'));

tests/browser/concept-inventory.mjs

@@ -2,7 +2,7 @@
 // same concept selectors in the immutable concept (/__test__/canonical/workspace) and in production, which now renders
 // the concept's classes. The concept is never rewritten: each page's map<Page>Concept(reference) applies only the
 // recorded spec-§8 deviations, each commented with its D-number, and diffInventory must then report nothing.
-import {expect} from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import {expect} from './fixtures.mjs';
 export {diffInventory} from './library-editor-canonical-comparison.mjs';
 
 export const CONCEPT_URL='/__test__/canonical/workspace';
@@ -78,6 +78,23 @@ export async function seedConcept(request,data={}) {
     return response.json();
 }
 
+/**
+ * Save `tags` as the current image tag set of the listed file at `path` (root 0) through the real observe and tag
+ * routes (#249 PR C), releasing the capture afterwards; returns the receipt's saved names.
+ */
+export async function saveImageTags(request,path,tags) {
+    const listing=await (await request.get('/prompt_manager/images/output?mode=workspace&limit=40')).json();
+    const target={root_index:0,root_scope:listing.images.find(item=>item.root_index===0).root_scope,path};
+    const observed=await (await request.post('/prompt_manager/image-annotations/observe',{data:{target}})).json();
+    expect(observed.success).toBe(true);
+    try {
+        const receipt=await (await request.put('/prompt_manager/image-annotations/tags',{data:{target,capture_token:observed.capture.token,
+            sha256:observed.capture.sha256,expected_epoch:observed.library_epoch,expected_tag_revision:observed.tag_revision,tags}})).json();
+        expect(receipt.success).toBe(true);
+        return receipt.tags;
+    } finally {await request.delete(`/prompt_manager/image-annotations/capture/${observed.capture.token}`);}
+}
+
 async function selectTheme(page,theme) {
     await page.getByRole('combobox',{name:'Theme',exact:true}).selectOption(theme);
 }

tests/browser/fixtures.mjs

@@ -0,0 +1,33 @@
+// Shared `test` for the page workflows: every Playwright worker owns one
+// tests/browser/server.py process on a port the OS chose for it, with its own temp root
+// and database. `baseURL` (and so `page`, `request` and `browser.newContext`)
+// points at that worker's server. A server that dies fails the running test
+// with its stderr tail, and the next test in the worker gets a fresh server.
+// A test that fails gets the stderr its server wrote during it attached as
+// `server-stderr`.
+import { test as base, expect } from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import { attachServerStderr, createServerSupervisor, startBrowserServer } from '../../tools/vendor/start-browser-server.mjs';
+
+const SERVER_FIXTURE_TIMEOUT_MS = 90_000;
+
+export const test = base.extend({
+    testServer: [async ({}, use) => {
+        const supervisor = createServerSupervisor({ start: () => startBrowserServer() });
+        try {
+            await use(supervisor);
+        } finally {
+            await supervisor.stop();
+        }
+    }, { scope: 'worker', timeout: SERVER_FIXTURE_TIMEOUT_MS }],
+    workerServer: [async ({ testServer }, use, testInfo) => {
+        const server = await testServer.ensure();
+        const mark = server.stderrMark();
+        await use(server);
+        await attachServerStderr(testInfo, server, mark);
+        const crash = testServer.crash(server);
+        if (crash) throw crash;
+    }, { auto: true, timeout: SERVER_FIXTURE_TIMEOUT_MS }],
+    baseURL: async ({ workerServer }, use) => use(workerServer.url),
+});
+
+export { expect };

tests/browser/focused-node.spec.mjs

@@ -1,5 +1,5 @@
 import {readFileSync} from 'node:fs';
-import {test,expect} from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import {test,expect} from './fixtures.mjs';
 import {startCoverage,saveCoverage} from '../../tools/vendor/browser-coverage.mjs';
 
 for (const variant of ['Focused', 'Split']) {

tests/browser/full-inspector.spec.mjs

@@ -1,5 +1,5 @@
 import {startCoverage,saveCoverage} from '../../tools/vendor/browser-coverage.mjs';
-import { test, expect } from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import { test, expect } from './fixtures.mjs';
 import { seedConcept, openConcept, openProduction, SIZES } from './concept-inventory.mjs';
 // D10/D20: the heading is the linked positive's title (its confirmed first line here), never the pair's
 // "Inspector recipe", which names the negative variant.
@@ -92,7 +92,7 @@ test('inspection wire preserves exact rooted envelope', async ({ request }) => {
 const inspector=page=>page.locator('.detail-panel');
 /** D19: the selected card's image opens the viewer (the side panel has no full-preview cover). */
 const openSelected=page=>page.locator('.media-card.selected .image-open').click();
-const selectFile=async(page,name,root=0)=>page.locator(`article[data-image-key="${root}:inspector/${name}"] .image-card-actions button`).click();
+const selectFile=async(page,name,root=0)=>page.locator(`article[data-image-key="${root}:inspector/${name}"] .image-card-actions button:first-child`).click();
 async function fullFixture(page,request) {
     await expect(await request.post('/__test__/reset',{data:{full_inspector:true}})).toBeOK();
     await page.goto('/prompt_manager/workspace#/images');await selectFile(page,'complete.png');
@@ -460,7 +460,7 @@ test('full inspector labels container representations and releases action URLs',
 test('full inspector distinguishes missing, limited and legacy unknown generation facts',async({page,request})=>{
     await fullFixture(page,request);const panel=inspector(page);
     await selectFile(page,'empty.png');await expect(panel.locator('.meta-grid dd')).toHaveText(Array(6).fill('Not recorded'));
-    await page.locator('article[data-image-key="0:architecture/study-01.png"] .image-card-actions button').click();await expect(panel.locator('.meta-grid dd')).toHaveText(Array(6).fill('Not recorded'));
+    await page.locator('article[data-image-key="0:architecture/study-01.png"] .image-card-actions button:first-child').click();await expect(panel.locator('.meta-grid dd')).toHaveText(Array(6).fill('Not recorded'));
     await selectFile(page,'control-heavy.png');
     const explanation='Configuration exceeds inspection limits; download the original file to inspect it locally.';
     await expect(panel.getByText(explanation,{exact:true})).toHaveCount(1);await expect(panel.getByText(explanation,{exact:true})).toBeVisible();
@@ -637,7 +637,7 @@ test('collection viewer follows filtered loaded order beyond mounted cards',asyn
     // #249 D26: the whole positive's images are in the All images dialog's own embedded gallery.
     await page.getByRole('button',{name:'All images for this positive',exact:true}).click();
     const listing=page.getByRole('dialog',{name:'All images for this positive',exact:true});
-    await listing.locator('article[data-image-key="0:inspector/complete.png"] .image-card-actions button').click();
+    await listing.locator('article[data-image-key="0:inspector/complete.png"] .image-card-actions button:first-child').click();
     await listing.locator('.media-card.selected .image-open').click();await expect(mediaDialog(page).locator('.viewer-canvas img')).toBeVisible();
     await expect(position(page)).toContainText('of 2 loaded:');
 });
@@ -804,6 +804,7 @@ test('slideshow preference failures leave native playback usable',async({page,re
 });
 
 test('full inspector attaches matched theme and mobile evidence with reachable content',async({page,request},testInfo)=>{
+    test.setTimeout(120_000); // 32 evidence captures across four viewports, like the concept comparisons
     await expect(await request.post('/__test__/reset',{data:{full_inspector:true,include_video:true}})).toBeOK();
     await page.goto('/prompt_manager/workspace#/images');
     const measurements=[];
@@ -929,7 +930,7 @@ test('collection viewer matches approved responsive frame',async({page,request},
     await page.screenshot({path:testInfo.outputPath('viewer-mobile-wider-font-2x.png')});
     await modal.getByRole('button',{name:'Close image viewer',exact:true}).click();await page.evaluate(()=>{document.documentElement.style.zoom='1';});await settleViewport(390);
     const longFiles=await(await request.get('/prompt_manager/images/output?mode=workspace&q=a-very-long&limit=40')).json();const long=longFiles.images[0];
-    await page.locator(`article[data-image-key="${long.root_index}:${long.path}"] .image-card-actions button`).click();await openCollection(page);
+    await page.locator(`article[data-image-key="${long.root_index}:${long.path}"] .image-card-actions button:first-child`).click();await openCollection(page);
     await page.evaluate(()=>{document.documentElement.style.zoom='2';});await settleViewport(390);await modal.evaluate(node=>{node.scrollTop=0;});
     await expect(modal.getByRole('button',{name:'Close image viewer',exact:true})).toBeInViewport();
 });

tests/browser/history-canonical-comparison.mjs

@@ -1,5 +1,5 @@
 import { writeFile } from 'node:fs/promises';
-import { expect } from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import { expect } from './fixtures.mjs';
 
 // Content, row counts, and image bytes come from independent synthetic datasets.
 // Saved choices and actual executions are separate production sections. The concept's
@@ -17,7 +17,6 @@ export const canonicalDifferences = {
     // Closed by #249: the gallery's checkbox is Images-only (#259) and the role blocks are the concept's (PR D, F); the
     // exact history comparison is tests/browser/history-concept.spec.mjs.
     remainingApplicationPolish: [
-        'Gallery card tag button (#249 D9, image-annotations slice) remains outside #209',
         'Production keeps mobile toolbar inside media panel and readable scrolling details; the concept toolbar collision is not reproduced',
     ],
 };

tests/browser/history-concept.spec.mjs

@@ -1,5 +1,5 @@
 import {writeFile} from 'node:fs/promises';
-import {test,expect} from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import {test,expect} from './fixtures.mjs';
 import {captureCanonicalState} from './history-canonical-comparison.mjs';
 import {SIZES,THEMES,MOTIONS,seedConcept,openConcept,openProduction,inventoryOf,diffInventory,mutate,stripExact} from './concept-inventory.mjs';
 

tests/browser/images-concept.spec.mjs

@@ -1,5 +1,5 @@
 import {writeFile} from 'node:fs/promises';
-import {test,expect} from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import {test,expect} from './fixtures.mjs';
 import {captureCanonicalState} from './history-canonical-comparison.mjs';
 import {SIZES,THEMES,MOTIONS,seedConcept,openConcept,openProduction,inventoryOf,diffInventory,mutate,stripExact,countCopy,reflow,withoutItems} from './concept-inventory.mjs';
 
@@ -59,12 +59,12 @@ function mapBulkbar(bulkbar,rise) {
 }
 
 /**
- * A list row's info once D9 narrows its actions to `actionsWidth`, from the concept's boxes:
+ * A list row's info with its actions `actionsWidth` wide, from the concept's boxes:
  * - D3 (the measurable list row, `.list-grid .media-info` in the overlay): the info fills the row beside the art and may
  *   shrink below a long nowrap title (`min-width: 0`), so it is the card's inner width less the art, and the actions
  *   end at its right padding; it keeps its own height (text or actions plus padding), centred in the row the concept
  *   stretches it across, so the text, centred in either, moves only by half the lines it lost;
- * - D9: the wider text column fits a subtitle the concept wrapped when its one-line width (`maxContent`) now fits.
+ * - the text column fits a subtitle the concept wrapped when its one-line width (`maxContent`) now fits.
  */
 function listInfo(card,actionsWidth) {
     const {info}=card,{copy,p}=info;
@@ -109,21 +109,16 @@ export function mapImagesConcept(reference,{manifest,now=new Date()}) {
     const cards=reference.cards.map(card=>{
         const filename=card.open.label.replace(/^Open /,'');
         const actions=card.info.actions;
-        const tag=actions.buttons.find(button=>button.label?.startsWith('Tag or auto-tag '));
-        // D9: no Tag button until the image-annotations slice; the actions group narrows by its box and gap.
-        const freed=tag?tag.width+parseFloat(actions.columnGap):0;
         // D3: a list row's info keeps its own height (text or actions plus its padding) so bounded rendering can measure
         // it, centred in the row the concept stretches it across; the text and the actions keep their place.
-        const actionsWidth=actions.width-freed;
-        const listRow=list?listInfo(card,actionsWidth):{};
+        const listRow=list?listInfo(card,actions.width):{};
         return {...card,
             // D29: the concept's alt adds its demo-only " · sample illustration N" to the card title.
             img:{...card.img,alt:card.img.alt.replace(/ · sample illustration \d+$/,'')},
             info:{...card.info,...listRow.info,copy:{...card.info.copy,...listRow.copy},p:{...card.info.p,...listRow.p},
                 // D10: the title keeps the file's relative path (the fixture manifest's) in `title`.
                 h3:{...card.info.h3,title:paths.get(filename)??null},
-                actions:{...actions,width:actionsWidth,left:listRow.actionsLeft??actions.left+freed,
-                    buttons:actions.buttons.filter(button=>button!==tag)}}};
+                actions:{...actions,left:listRow.actionsLeft??actions.left}}};
     });
     // D17: the right-hand date is the newest result's day under Recently added, here the fixture's newest mtime.
     // D35: the count line uses the singular for one.
@@ -538,8 +533,9 @@ test('one image appended by scrolling reads in the singular',async({page,request
 });
 
 // D3 (one footer height for the collection): bounded rendering cannot measure unrendered cards, so every row reserves
-// the tallest measured footer. Two seedless files (D11 "W × H · PNG", one line) form the first row at 390px, where
-// every concept subtitle wraps: their row is as tall as the wrapped rows, and every concept row moves down by one.
+// the tallest measured footer. Two seedless files (D11 "W × H · PNG", one line: a short "96 × 72" beside Info and Tag)
+// form the first row at 390px, where every concept subtitle wraps: their row is as tall as the wrapped rows, and every
+// concept row moves down by one.
 test('one footer height for the collection (D3)',async({page,request})=>{
     // A second fixture seed (dozens of PNGs and their thumbnails) runs inside the test.
     test.setTimeout(120_000);
@@ -548,7 +544,7 @@ test('one footer height for the collection (D3)',async({page,request})=>{
     const reference=await inventoryOf(page,'#main');
     expect(reference.grid.columns).toBe(2);
     expect(reference.cards.every(card=>card.info.p.height>parseFloat(card.info.p.lineHeight))).toBe(true);
-    const manifest=await seedConcept(request,{seedless_images:2});
+    const manifest=await seedConcept(request,{seedless_images:2,short_seedless:true});
     await openProductionLoaded(page,'dark','no-preference',FIXTURE_IMAGES+2);
     const actual=await inventoryOf(page,'#workspace-main');
     const box=card=>({left:card.left,top:card.top,width:card.width,height:card.height});
@@ -557,7 +553,10 @@ test('one footer height for the collection (D3)',async({page,request})=>{
         ...reference.cards.slice(0,4).map(card=>({...box(card),top:card.top+pitch}))]);
     const seedless=actual.cards.slice(0,2);
     expect(seedless.map(card=>card.open.label)).toEqual(manifest.seedless.map(image=>`Open ${image.filename}`));
-    expect(seedless.map(card=>card.info.p.text)).toEqual(['1536 × 1152 · PNG','1536 × 1152 · PNG']);
+    expect(seedless.map(card=>card.info.p.text)).toEqual(['96 × 72 · PNG','96 × 72 · PNG']);
     // Their own footer is one line shorter than the row it reserves.
-    for(const card of seedless)expect(card.info.height).toBeLessThan(reference.cards[0].info.height);
+    for(const card of seedless) {
+        expect(card.info.p.height).toBeLessThanOrEqual(parseFloat(card.info.p.lineHeight)+0.5);
+        expect(card.info.height).toBeLessThan(reference.cards[0].info.height);
+    }
 });

tests/browser/inspector-concept.spec.mjs

@@ -1,7 +1,7 @@
 import {writeFile} from 'node:fs/promises';
-import {test,expect} from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import {test,expect} from './fixtures.mjs';
 import {captureCanonicalState} from './history-canonical-comparison.mjs';
-import {SIZES,THEMES,seedConcept,openConcept,openProduction,inspectorOf,diffInventory,mutate,countCopy} from './concept-inventory.mjs';
+import {SIZES,THEMES,seedConcept,openConcept,openProduction,inspectorOf,diffInventory,mutate,countCopy,saveImageTags} from './concept-inventory.mjs';
 
 // Task D4 of #249: the Images side panel (app.js imageDetails a400–414) and the viewer's full details (imageDetails(i,
 // true)) against the immutable concept, on the concept-shaped real fixture, at 1600x1100 and 390x844 in both themes.
@@ -319,6 +319,27 @@ test('a recipe-linked image without recipe metadata shows its positive prompt me
     await expect(panelOf(page).locator('.detail-footer button')).toHaveText(['Send to ComfyUI','Prompt history']);
 });
 
+// Task C3 (D88, #249 PR C): the concept's chip row is the image's own tags (`i.tags`, app.js a402). With those names
+// saved as the file's current image tags, production draws the row from them (Tags filter buttons in the panel, plain
+// in the viewer, the linked positive's equal tags not repeated) and still equals the concept, in both sizes.
+for(const [width,height] of SIZES)test(`image tags saved on the concept image keep the inspector equal to the concept ${width}x${height} (D88)`,async({page,request},testInfo)=>{
+    test.setTimeout(120_000);
+    const viewport={width,height};await page.setViewportSize(viewport);
+    await openConceptPanel(page,'dark');
+    const names=await panelOf(page).locator('.detail-inner > .tags').last().locator('button.chip').allTextContents();
+    expect(names).toEqual(['cinematic','blue_hour','architecture']);
+    await captureCanonicalState(page,testInfo,`concept-image-tags-${width}`,'#image-inspector');
+    const reference=await inventory(page,true);
+    const {path}=testInfo.manifest.images.find(item=>item.filename==='alpine_00001.png');
+    expect(await saveImageTags(request,path,names)).toEqual(names);
+    const read=page.waitForResponse(item=>new URL(item.url()).pathname===`/prompt_manager/images/inspect/${path}`);
+    await openProductionPanel(page,'dark');
+    expect((await (await read).json()).image_tags).toEqual({state:'current',tags:names});
+    await expect(panelOf(page).locator('.detail-inner > .tags').last().locator('.chip')).toHaveText(names);
+    await captureCanonicalState(page,testInfo,`production-image-tags-${width}`,'#image-inspector');
+    assertInspectorMatch(reference,await inventory(page),viewport);
+});
+
 // The comparison is only worth something if it notices a change (spec §7 mutation check): on the panel one CSS
 // mutation (the role label's letter spacing) and one DOM mutation (the footer removed); in the viewer (final review m5)
 // its pair actions removed and D38's Show in folder removed; each applied after an exact match.

tests/browser/library-concept.spec.mjs

@@ -1,5 +1,5 @@
 import {writeFile} from 'node:fs/promises';
-import {test,expect} from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import {test,expect} from './fixtures.mjs';
 import {captureCanonicalState} from './history-canonical-comparison.mjs';
 import {SIZES,THEMES,MOTIONS,seedConcept,openConcept,openProduction,inventoryOf,diffInventory,mutate,stripExact,countCopy,reflow} from './concept-inventory.mjs';
 

tests/browser/library-editor-visual.spec.mjs

@@ -1,4 +1,4 @@
-import {test,expect} from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import {test,expect} from './fixtures.mjs';
 import {captureCanonicalState} from './history-canonical-comparison.mjs';
 
 // Artifact-only captures of the editor states that have no concept counterpart (archive confirmation, Show archived,

tests/browser/pair-node.spec.mjs

@@ -1,5 +1,5 @@
 import {readFileSync} from 'node:fs';
-import {test,expect} from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import {test,expect} from './fixtures.mjs';
 import {startCoverage,saveCoverage} from '../../tools/vendor/browser-coverage.mjs';
 
 const pair={id:4,positive_text:'Library positive',negative_text:'Library negative',category:'Portrait',tags:['studio']};

tests/browser/prompt-bridge.spec.mjs

@@ -1,4 +1,4 @@
-import { test, expect } from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import { test, expect } from './fixtures.mjs';
 import { startCoverage, saveCoverage } from '../../tools/vendor/browser-coverage.mjs';
 
 const moduleUrl = '/prompt_manager/js/shared/toolkit/prompt-bridge.js';

tests/browser/prompt-inspector.spec.mjs

@@ -1,5 +1,5 @@
 import {startCoverage,saveCoverage} from '../../tools/vendor/browser-coverage.mjs';
-import {test,expect} from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import {test,expect} from './fixtures.mjs';
 
 const workspace='/prompt_manager/workspace#/images';
 const details=page=>page.getByRole('complementary',{name:'Selected image details',exact:true});

tests/browser/server.py

@@ -1,16 +1,20 @@
 """Serve production pages and API routes with disposable state for browser tests."""
 
 import argparse
+import asyncio
 import json
 import os
 import shutil
+import socket
 from pathlib import Path
 import sys
 import tempfile
+import threading
 import time
 import types
 
 from aiohttp import web
+from aiohttp.web_runner import GracefulExit
 from PIL import Image
 from PIL.PngImagePlugin import PngInfo
 
@@ -31,6 +35,7 @@ from tests.browser.workspace_library_editor_fixture import (
     register_library_editor_commands,
 )
 from tests.browser.workspace_vocabulary_fixture import register_vocabulary_commands
+from tests.browser.workspace_tagging_fixture import register_tagging_commands
 
 CANONICAL_ROOT = Path(__file__).resolve().parents[2]
 CANONICAL_ASSETS = {
@@ -53,11 +58,19 @@ class TestModel:
     wd14_character_threshold = 0.85
     custom_prompt = ""
     loaded = None
+    # Models the Tag image dialog workflows mark as not downloaded.
+    unavailable: set = set()
+    # Scored names the concept comparisons answer instead of the fixed list
+    # (``/__test__/tagging/suggestions``); None keeps the fixed list.
+    suggestions: list | None = None
 
     def get_models_status(self) -> dict:
         return {
-            name: {"downloaded": True, "loaded": self.loaded == name}
-            for name in ("gguf", "wd14-vit", "wd14-convnext", "wd14-swinv2")
+            name: {
+                "downloaded": name not in self.unavailable,
+                "loaded": self.loaded == name,
+            }
+            for name in ("gguf", "hf", "wd14-vit", "wd14-convnext", "wd14-swinv2")
         }
 
     def is_model_loaded(self) -> bool:
@@ -76,6 +89,26 @@ class TestModel:
     def generate_tags(self, image_path: str, **kwargs: object) -> list[str]:
         return ["generated", Path(image_path).parent.parent.name]
 
+    def generate_tag_suggestions(
+        self, image_path: str, *, limit: int, **kwargs: object
+    ) -> list[tuple[str, float | None]]:
+        """Fixed scored suggestions (WD14) or unscored ones (language models)."""
+        scored = self.loaded is not None and self.loaded.startswith("wd14")
+        if self.suggestions is not None:
+            return [
+                (name, confidence if scored else None)
+                for name, confidence in self.suggestions[:limit]
+            ]
+        names = ["sandstone", "desert", "arch", "sky", "rock", "dune", "canyon"]
+        names += [f"tag_{index:02d}" for index in range(100)]
+        return [
+            (name, round(0.98 - index * 0.008, 3) if scored else None)
+            for index, name in enumerate(names[:limit])
+        ]
+
+    def download_model(self, model_type: str, progress_callback=None) -> bool:
+        raise AssertionError("browser tests never download a model")
+
 
 REVEAL_OUTCOMES = ("opened", "unavailable", "fail")
 
@@ -159,6 +192,7 @@ def create_app(root: Path) -> web.Application:
     history_commands = register_workspace_history_commands(routes, api, root)
     library_editor_commands = register_library_editor_commands(routes, api, root)
     register_vocabulary_commands(routes, api, root)
+    register_tagging_commands(routes, api, root, model)
     reveal_stub = register_reveal_stub(routes, api, root)
     register_activity_commands(app, routes, api)
 
@@ -185,6 +219,9 @@ def create_app(root: Path) -> web.Application:
             for table in (
                 "saved_views",
                 "image_favorites",
+                "media_image_tags",
+                "media_annotation_bytes",
+                "media_annotation_files",
                 "image_prompt_pairs",
                 "prompt_pair_tags",
                 "prompt_pairs",
@@ -449,7 +486,8 @@ def create_app(root: Path) -> web.Application:
         if (
             request.query
             or not isinstance(data, dict)
-            or set(data) - set(limits)
+            or set(data) - set(limits) - {"short_seedless"}
+            or not isinstance(data.get("short_seedless", False), bool)
             or any(
                 isinstance(data.get(key, 0), bool)
                 or not isinstance(data.get(key, 0), int)
@@ -468,6 +506,7 @@ def create_app(root: Path) -> web.Application:
                 root,
                 extra_images=data.get("extra_images", 0),
                 seedless_images=data.get("seedless_images", 0),
+                short_seedless=data.get("short_seedless", False),
             )
         )
         return web.json_response({"success": True, **manifest})
@@ -515,14 +554,77 @@ def create_app(root: Path) -> web.Application:
     return app
 
 
-def main() -> None:
+HOST = "127.0.0.1"
+PORT_ANNOUNCEMENT = "PM_TEST_SERVER_PORT="
+
+
+def bind_listener(port: int) -> socket.socket:
+    """Bind the server's socket before the app is built.
+
+    ``port`` 0 lets the OS choose a free port, which this process then holds,
+    so parallel launchers never race for one between choosing and binding.
+    """
+    listener = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
+    try:
+        if port and os.name != "nt":
+            # aiohttp's own default on POSIX; on Windows it would let a
+            # second process share the port.
+            listener.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
+        listener.bind((HOST, port))
+        listener.listen(128)
+    except OSError:
+        listener.close()
+        raise
+    return listener
+
+
+def _graceful_exit() -> None:
+    raise GracefulExit()
+
+
+def exit_when_stdin_closes(loop: asyncio.AbstractEventLoop) -> threading.Thread:
+    """Stop the app gracefully once stdin reaches end of file.
+
+    The launcher keeps the write end of this process's stdin pipe, and the OS
+    closes it however the launcher dies, SIGKILL included. GracefulExit runs
+    aiohttp's cleanup (closing the database) and lets the temp root be removed,
+    on every OS, where a signal from a dead parent would not.
+    """
+
+    def watch() -> None:
+        while sys.stdin.buffer.read(65536):
+            pass
+        loop.call_soon_threadsafe(_graceful_exit)
+
+    thread = threading.Thread(target=watch, name="stdin-eof", daemon=True)
+    thread.start()
+    return thread
+
+
+def main(argv: list[str] | None = None) -> None:
     parser = argparse.ArgumentParser()
-    parser.add_argument("--port", type=int, default=8765)
-    args = parser.parse_args()
+    parser.add_argument(
+        "--port", type=int, default=8765, help="0 binds an OS-chosen free port"
+    )
+    parser.add_argument(
+        "--exit-on-stdin-eof",
+        action="store_true",
+        help="shut down when stdin closes (the launcher's parent-death guard)",
+    )
+    args = parser.parse_args(argv)
+    listener = bind_listener(args.port)
+    # One machine-readable line tells the launcher which port this process holds.
+    print(f"{PORT_ANNOUNCEMENT}{listener.getsockname()[1]}", flush=True)
+    loop = asyncio.new_event_loop()
+    if args.exit_on_stdin_eof:
+        exit_when_stdin_closes(loop)
+    original_cwd = os.getcwd()
     with tempfile.TemporaryDirectory(prefix="pm-browser-") as directory:
-        web.run_app(
-            create_app(Path(directory).resolve()), host="127.0.0.1", port=args.port
-        )
+        try:
+            web.run_app(create_app(Path(directory).resolve()), sock=listener, loop=loop)
+        finally:
+            # create_app works inside the root; Windows cannot remove a process's cwd.
+            os.chdir(original_cwd)
 
 
 if __name__ == "__main__":

tests/browser/shell-concept.spec.mjs

@@ -1,5 +1,5 @@
 import {writeFile} from 'node:fs/promises';
-import {test,expect} from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import {test,expect} from './fixtures.mjs';
 import {captureCanonicalState} from './history-canonical-comparison.mjs';
 import {SIZES,THEMES,MOTIONS,seedConcept,openConcept as openConceptPage,openProduction,inventoryOf,diffInventory,focusRingOf,hoverStateOf,mutate} from './concept-inventory.mjs';
 
@@ -119,10 +119,7 @@ function mapShellConcept(reference) {
             const spans=chrome.footer.spans.map(span=>({...span,text:'ComfyAssets · Local tools. Your library.'}));
             return {...chrome.footer,spans,buttons:[],height:remainingFooter(chrome.footer,spans)};
         })(),
-    },
-    heading:reference.heading&&{...reference.heading,actions:reference.heading.actions&&{...reference.heading.actions,
-        // #245 vocabulary ruling D2 (mapVocabularyConcept): Auto-tag images is omitted until the image-annotation slice.
-        buttons:reference.heading.actions.buttons.filter(item=>item.text!=='Auto-tag images')}}};
+    }};
 }
 
 /** The heading without its page-owned description copy and actions; the h1 text is the shell's and stays compared. */

tests/browser/tagging-canonical-comparison.mjs

@@ -0,0 +1,142 @@
+// Concept-vs-production inventory of the Tag image dialog (plan Mock inventory M2: docs/concepts/workspace/app.js
+// tagReview a1037–1061, renderTagModelOptions a1062–1076, generateTags a1077–1135; style.css s1598–1694) and of the
+// card's Info and Tag group (M1: app.js mediaCard a398, style.css s1589–1597). The *Of functions run in the page.
+// The concept is never rewritten: mapConceptTagging applies only the recorded deviations, each named with its D-number
+// or ruling; everything else must be equal.
+export {diffInventory as diffTaggingInventory} from './library-editor-canonical-comparison.mjs';
+
+/** The open Tag image dialog (`dialog#modal`), its controls in order with their concept tokens. */
+export function taggingInventory(dialog) {
+    const css=(node,...keys)=>{const style=getComputedStyle(node);return Object.fromEntries(keys.map(key=>[key,style[key]]));};
+    const px=value=>Math.round(value*2)/2;
+    const box=node=>{const rect=node.getBoundingClientRect();return {width:px(rect.width),height:px(rect.height)};};
+    const font=node=>css(node,'fontSize','fontWeight','lineHeight','color','marginTop','marginBottom');
+    const edges=node=>css(node,'paddingTop','paddingRight','paddingBottom','paddingLeft','borderTopWidth','borderTopStyle','borderTopColor',
+        'borderTopLeftRadius','backgroundColor');
+    const buttonOf=node=>({text:node.textContent.trim(),label:node.getAttribute('aria-label'),primary:node.classList.contains('ca-primary'),
+        icon:node.querySelector('svg')!==null,disabled:node.getAttribute('aria-disabled')==='true'||node.disabled,
+        ...css(node,'minHeight','fontSize','color','backgroundColor','borderTopColor','borderTopLeftRadius')});
+    const controlOf=node=>({tag:node.tagName.toLowerCase(),id:node.id,type:node.type??null,value:node.value,
+        min:node.getAttribute('min'),max:node.getAttribute('max'),step:node.getAttribute('step'),
+        describedBy:node.getAttribute('aria-describedby'),
+        options:node.tagName==='SELECT'?[...node.options].map(option=>[option.value,option.textContent]):null,
+        ...css(node,'fontSize','minHeight','backgroundColor','color','borderTopColor','borderTopLeftRadius','paddingLeft')});
+    const fieldOf=field=>{const label=field.querySelector('label'),control=field.querySelector('input,textarea,select');
+        return {label:label.textContent,labelled:label.htmlFor===control.id,...css(field,'display','rowGap','marginTop','marginBottom'),control:controlOf(control)};};
+    const header=dialog.querySelector('.dialog-header'),title=header.querySelector('h2'),close=header.querySelector('button');
+    const body=dialog.querySelector('#modal-body');
+    const heading=body.querySelector('.tag-image-heading');
+    const preview=heading.querySelector('.tag-image-preview'),copy=heading.querySelector(':scope > div');
+    const sizeChip=copy.querySelector('.chip');
+    const note=body.querySelector('.tag-scope-note');
+    const manual=body.querySelector('#manual-image-tags'),manualField=manual.closest('.field'),help=body.querySelector('#manual-tags-help');
+    const panel=body.querySelector('.tag-model-panel'),options=panel.querySelector('#image-tag-model-options');
+    const generate=[...panel.querySelectorAll('button')].at(-1);
+    const suggestions=body.querySelector('section.tag-suggestion-panel');
+    const list=suggestions.querySelector('#tag-suggestions'),status=suggestions.querySelector('#tag-generation-status');
+    const count=suggestions.querySelector('#suggestion-count');
+    const actions=body.querySelector('.dialog-actions');
+    // A suggestion is a concept checkbox row (`label.tag-suggestion`) or a production chip (`span.chip.tag-suggestion`).
+    const rows=[...list.querySelectorAll('.tag-suggestion')].map(row=>{
+        const name=row.querySelector('span:not(.tag-confidence)'),confidence=row.querySelector('.tag-confidence');
+        const checkbox=row.querySelector('input[type=checkbox]'),remove=row.querySelector('button');
+        return {name:name.textContent,confidence:confidence.textContent,
+            control:checkbox?{kind:'checkbox',checked:checkbox.checked}:remove?{kind:'remove',label:remove.getAttribute('aria-label')}:null,
+            ...css(row,'display','alignItems','borderTopColor','backgroundColor','borderTopWidth','borderTopStyle','borderTopLeftRadius','fontSize','paddingTop','paddingLeft'),
+            confidenceToken:css(confidence,'color','fontSize')};
+    });
+    return {
+        dialog:{...css(dialog,'paddingTop','paddingRight','paddingBottom','paddingLeft','borderTopLeftRadius','backgroundColor','color','overflowY'),
+            width:px(dialog.getBoundingClientRect().width),modal:dialog.matches(':modal'),tagging:dialog.classList.contains('tagging-modal'),
+            labelledBy:document.getElementById(dialog.getAttribute('aria-labelledby'))===title,
+            noHorizontalOverflow:dialog.scrollWidth<=dialog.clientWidth+1},
+        header:{title:{text:title.textContent,...font(title)},close:{label:close.getAttribute('aria-label'),text:close.textContent.trim()}},
+        demo:body.querySelector('.demo-note')?.textContent??null,
+        order:[...body.children].map(node=>node.className),
+        heading:{...css(heading,'display','columnGap','alignItems','marginBottom'),
+            preview:{tag:preview.tagName.toLowerCase(),alt:preview.getAttribute('alt'),...box(preview),...css(preview,'objectFit','borderTopLeftRadius','borderTopWidth','borderTopStyle','borderTopColor','flexShrink')},
+            h3:{text:copy.querySelector('h3').textContent,...font(copy.querySelector('h3'))},
+            filename:{text:copy.querySelector('p.small').textContent,...font(copy.querySelector('p.small'))},
+            size:sizeChip?{text:sizeChip.textContent,...css(sizeChip,'marginTop','fontSize','borderTopColor','borderTopLeftRadius','paddingLeft')}:null},
+        scopeNote:note?{text:note.textContent,...edges(note),...css(note,'borderLeftWidth','borderLeftStyle','borderLeftColor','fontSize','color')}:null,
+        manual:{...fieldOf(manualField),help:{text:help.textContent,...font(help)},described:help.id===manual.getAttribute('aria-describedby')},
+        modelPanel:{...edges(panel),...css(panel,'marginTop'),h3:{text:panel.querySelector('h3').textContent,...font(panel.querySelector('h3'))},
+            fields:[...panel.querySelectorAll('.field')].map(fieldOf),
+            twoCol:options.querySelector('.two-col')?{columns:getComputedStyle(options.querySelector('.two-col')).gridTemplateColumns.split(' ').length,
+                ...css(options.querySelector('.two-col'),'display','columnGap')}:null,
+            generate:buttonOf(generate)},
+        suggestions:{labelled:suggestions.getAttribute('aria-labelledby')===suggestions.querySelector('h3').id,...css(suggestions,'marginTop'),
+            title:{text:suggestions.querySelector('h3').textContent,...font(suggestions.querySelector('h3'))},
+            count:{text:count.textContent,...font(count)},
+            copy:list.querySelector(':scope > p.small')?.textContent??null,
+            rows,
+            status:{text:status.textContent,role:status.getAttribute('role'),...css(status,'marginTop','fontSize','color')}},
+        actions:{...css(actions,'position','bottom','marginLeft','marginRight','marginBottom','paddingTop','paddingLeft','backgroundColor',
+            'borderTopWidth','borderTopStyle','borderTopColor','zIndex','display','justifyContent','columnGap'),
+            buttons:[...actions.querySelectorAll('button')].map(buttonOf)},
+        noPageScroll:document.documentElement.scrollWidth<=innerWidth+1,
+    };
+}
+
+/** The first image card's Info and Tag group (M1), with its card's filename. */
+export function cardActionsInventory(card) {
+    const css=(node,...keys)=>{const style=getComputedStyle(node);return Object.fromEntries(keys.map(key=>[key,style[key]]));};
+    const info=card.querySelector('.media-info'),actions=info.querySelector(':scope > .image-card-actions');
+    return {
+        info:css(info,'display','justifyContent','alignItems','columnGap'),
+        actions:{...css(actions,'display','alignItems','columnGap','flexShrink','minWidth'),lastInInfo:info.lastElementChild===actions},
+        buttons:[...actions.querySelectorAll(':scope > button')].map(node=>{const rect=node.getBoundingClientRect();return {
+            className:node.className,label:node.getAttribute('aria-label'),icon:node.querySelector('svg')?.innerHTML.replace(/\s+/g,' ').trim()??null,
+            width:Math.round(rect.width),height:Math.round(rect.height),
+            ...css(node,'minWidth','minHeight','paddingLeft','paddingTop','borderTopLeftRadius','borderTopColor','backgroundColor','color')};}),
+    };
+}
+
+// D84 (plan D61): the legacy engines' values, labels and sizes, GGUF JoyCaption in the concept's Ollama slot.
+const MODEL_LABELS=Object.freeze({'wd14-swinv2':['wd14-swinv2','WD14 SwinV2 · ~400 MB'],'wd14-vit':['wd14-vit','WD14 ViT · ~400 MB'],
+    ollama:['gguf','GGUF JoyCaption (Recommended) · ~4.5 GB'],huggingface:['hf','HuggingFace JoyCaption · ~16 GB']});
+const CONCEPT_LABELS=Object.freeze({'WD14 SwinV2 · fast classifier':'WD14 SwinV2 · ~400 MB','WD14 ViT · visual tags':'WD14 ViT · ~400 MB',
+    'Ollama · local language model':'GGUF JoyCaption (Recommended) · ~4.5 GB','HuggingFace · language model':'HuggingFace JoyCaption · ~16 GB'});
+
+/**
+ * The concept inventory with the recorded deviations applied:
+ * - no prototype note (Global Constraints: `Prototype · suggestions are simulated…` is not shipped);
+ * - D83 (ruling T4): while suggestions exist, "Remove anything you do not want to add." leads them, and every
+ *   suggestion is a chip with a `Remove <name>` × instead of a checked checkbox row: the concept's own `.chip` corner
+ *   (`chip`: a concept `.chip` inventory, the heading's size chip) around the row's own text (12px) in the checked
+ *   row's colours (accent border, `--soft-blue`);
+ * - D84 (plan D61): the legacy engine labels and values, and the server's default instructions (`instructions`);
+ * - D91 (plan D68): a Suggestions count field after the model's own fields, on the concept's number field tokens
+ *   (`numberField`: the concept's General confidence field) with the configured count (`limit`);
+ * - D29: the prototype's "Sample suggestions from …" status reads "Suggestions from …".
+ * The opened state is compared unmapped: both show the concept's on-screen "Generate suggestions with the selected
+ * model before approving." (its renderTagModelOptions, a1062–1076, replaces the markup's sentence at once).
+ */
+export function mapConceptTagging(reference,{instructions,limit=10,chip,numberField}) {
+    const [select,...own]=reference.modelPanel.fields;
+    const suggestionsField={...numberField,label:'Suggestions',control:{...numberField.control,id:'image-suggestion-limit',
+        value:String(limit),min:'1',max:'100',step:'1'}};
+    const mappedOwn=own.map(field=>field.control.id==='image-tag-instructions'?{...field,control:{...field.control,value:instructions}}:field);
+    const rows=reference.suggestions.rows.map(row=>({...row,
+        control:row.control?.kind==='checkbox'&&row.control.checked?{kind:'remove',label:`Remove ${row.name}`}:row.control,
+        borderTopLeftRadius:chip.borderTopLeftRadius}));
+    const copy=rows.length&&reference.suggestions.copy===null?'Remove anything you do not want to add.':reference.suggestions.copy;
+    const status=reference.suggestions.status.text.replace(/^Sample suggestions from (.+)\. Approve tags to save\.$/,
+        (_,label)=>`Suggestions from ${CONCEPT_LABELS[label]??label}. Approve tags to save.`);
+    return {...reference,
+        demo:null,
+        order:reference.order.filter(name=>name!=='demo-note'),
+        modelPanel:{...reference.modelPanel,
+            fields:[{...select,control:{...select.control,value:MODEL_LABELS[select.control.value]?.[0]??select.control.value,
+                options:select.control.options.map(([value])=>MODEL_LABELS[value])}},...mappedOwn,suggestionsField]},
+        suggestions:{...reference.suggestions,copy,rows,status:{...reference.suggestions.status,text:status}},
+    };
+}
+
+/**
+ * Production's own chip geometry beyond the concept's `.chip` (D83): the chip's padding hugs its × button. Only the
+ * leaves D83 replaces are dropped from both sides; names, confidences, colours, radius and font stay compared.
+ */
+export function withoutChipGeometry(inventory) {
+    return {...inventory,suggestions:{...inventory.suggestions,rows:inventory.suggestions.rows.map(({paddingTop,paddingLeft,alignItems,...row})=>row)}};
+}

tests/browser/tagging-concept.spec.mjs

@@ -0,0 +1,307 @@
+// The Tag image dialog and the card's Info and Tag group against the approved concept (#249 PR C, plan Task C4, Mock
+// inventory M1 and M2). Production runs on the concept-shaped fixture with the concept image's own tags saved as its
+// image tags and the injected fake model answering the concept's own sample (alpine: mountain 98%, observatory 95%,
+// snow 88%, blue_hour 81%), so every remaining difference is a recorded deviation applied by mapConceptTagging
+// (D29, D83, D84, D91) or a test failure. No model is ever downloaded: every test checks no download request was sent.
+import { test, expect } from './fixtures.mjs';
+import { seedConcept, openConcept, openProduction, saveImageTags, mutate, SIZES, THEMES } from './concept-inventory.mjs';
+import { taggingInventory, cardActionsInventory, mapConceptTagging, diffTaggingInventory, withoutChipGeometry } from './tagging-canonical-comparison.mjs';
+
+const MODULE = '/prompt_manager/js/workspace/image-tagging.mjs';
+// The concept's first image (app.js images[0]) and its sample suggestions (generateTags, scene "alpine").
+const CONCEPT_FILE = 'alpine_00001.png';
+const CONCEPT_TAGS = ['cinematic', 'blue_hour', 'architecture'];
+const ALPINE = [['mountain', 0.98], ['observatory', 0.95], ['snow', 0.88], ['blue_hour', 0.81]];
+// The fake model's server default instructions (tests/browser/server.py TestModel.default_prompt), D84.
+const INSTRUCTIONS = 'Describe the image';
+const CONCEPT_MODEL = { gguf: 'ollama' };
+
+const tagDialog = page => page.getByRole('dialog', { name: 'Tag image', exact: true });
+const firstCard = (page, side) => page.locator(side === 'concept' ? '#image-grid > article.media-card' : '#workspace-main article[data-image-key]').first();
+
+test.beforeEach(async ({ page, request }) => {
+    const manifest = await seedConcept(request);
+    expect(manifest).toBeTruthy();
+    await expect(await request.post('/__test__/tagging/suggestions', { data: { scored: ALPINE } })).toBeOK();
+    page.downloads = [];
+    page.on('request', item => { if (/\/autotag\/download\//.test(item.url())) page.downloads.push(item.url()); });
+    page.errors = [];
+    page.on('pageerror', error => page.errors.push(error.message));
+});
+test.afterEach(async ({ page, request }) => {
+    await leave(page);
+    expect(page.downloads).toEqual([]);
+    expect(page.errors).toEqual([]);
+    await request.post('/__test__/tagging/suggestions', { data: { scored: null } });
+});
+
+async function conceptPath(request) {
+    const listing = await (await request.get('/prompt_manager/images/output?mode=workspace&limit=40')).json();
+    const image = listing.images.find(item => item.filename === CONCEPT_FILE);
+    expect(image).toBeTruthy();
+    return image.path;
+}
+
+/**
+ * Close an open Tag dialog before opening the next page, so each comparison starts from a settled page. Leaving with
+ * the dialog open is covered on its own (workspace-tagging-lifecycle.spec.mjs: the keepalive release on pagehide).
+ */
+async function leave(page) {
+    if (await tagDialog(page).count()) {
+        await page.keyboard.press('Escape');
+        await expect(tagDialog(page)).toHaveCount(0);
+    }
+}
+/** Open `side`'s Tag image dialog from its first card, settled (production: the capture preview and saved tags). */
+async function openTagDialog(page, side, theme) {
+    await leave(page);
+    if (side === 'concept') await openConcept(page, 'images', theme);
+    else await openProduction(page, 'images', theme);
+    await firstCard(page, side).getByRole('button', { name: `Tag or auto-tag ${CONCEPT_FILE}`, exact: true }).click();
+    await expect(tagDialog(page)).toBeVisible();
+    await settled(page, side);
+}
+async function settled(page, side) {
+    if (side === 'production') {
+        await expect(tagDialog(page).locator('img.tag-image-preview')).toHaveAttribute('src', /\/image-annotations\/capture\//);
+        await expect(tagDialog(page).locator('#manual-image-tags')).toHaveValue(CONCEPT_TAGS.join(', '));
+    }
+    await page.evaluate(async () => {
+        await document.fonts.ready;
+        await Promise.all([...document.querySelectorAll('#modal img')].map(image => image.decode().catch(() => {})));
+    });
+}
+async function generateWith(page, side, model) {
+    const select = tagDialog(page).getByRole('combobox', { name: 'Tagging model', exact: true });
+    if (model) await select.selectOption(side === 'concept' ? CONCEPT_MODEL[model] ?? model : model);
+    await tagDialog(page).getByRole('button', { name: 'Generate suggestions', exact: true }).click();
+    await expect(tagDialog(page).locator('#suggestion-count')).not.toHaveText('Not generated');
+}
+// The pointer rests outside the dialog, so no control is measured in its hover state.
+const inventory = async page => {
+    await page.mouse.move(0, 0);
+    return page.locator('dialog#modal').evaluate(taggingInventory);
+};
+
+/** The concept and production inventories of one dialog state, the concept mapped (D29, D83, D84, D91). */
+async function compareState(page, { theme, act = async () => {} }) {
+    await openTagDialog(page, 'concept', theme);
+    const conceptOpened = await inventory(page);
+    await act(page, 'concept');
+    const concept = await inventory(page);
+    await openTagDialog(page, 'production', theme);
+    await act(page, 'production');
+    const actual = await inventory(page);
+    const numberField = conceptOpened.modelPanel.fields.find(field => field.control.id === 'image-general-threshold');
+    const reference = mapConceptTagging(concept, { instructions: INSTRUCTIONS, chip: concept.heading.size, numberField });
+    return { reference: withoutChipGeometry(reference), actual: withoutChipGeometry(actual) };
+}
+
+for (const [width, height] of SIZES) for (const theme of THEMES) {
+    test(`concept and production Tag dialog match at ${width}x${height} ${theme}`, async ({ page, request }) => {
+        test.setTimeout(90_000);
+        await saveImageTags(request, await conceptPath(request), CONCEPT_TAGS);
+        await page.setViewportSize({ width, height });
+        // Opened, generated with WD14 (the concept's default), then generated with a language model.
+        for (const [state, act] of [
+            ['opened', async () => {}],
+            ['generated WD14', (page, side) => generateWith(page, side)],
+            ['generated GGUF', (page, side) => generateWith(page, side, 'gguf')],
+        ]) {
+            const { reference, actual } = await compareState(page, { theme, act });
+            expect(diffTaggingInventory(reference, actual), `${state} at ${width}x${height} ${theme}`).toEqual([]);
+        }
+        // Pinned concept geometry (M2), so a mapping mistake on both sides cannot pass unseen.
+        const { actual } = await compareState(page, { theme, act: (page, side) => generateWith(page, side) });
+        const small = width <= 700;
+        expect([actual.heading.preview.width, actual.heading.preview.height]).toEqual(small ? [86, 86] : [112, 100]);
+        expect(actual.heading.columnGap).toBe('16px');
+        expect([actual.modelPanel.paddingTop, actual.modelPanel.borderTopLeftRadius]).toEqual(['16px', '6px']);
+        expect(actual.actions.position).toBe('sticky');
+        expect(actual.suggestions.rows.map(row => [row.name, row.confidence, row.control.label])).toEqual(
+            ALPINE.map(([name, confidence]) => [name, `${Math.round(confidence * 100)}%`, `Remove ${name}`]));
+        const accent = await page.evaluate(() => {
+            const probe = document.createElement('i');
+            probe.style.cssText = 'border-top:1px solid var(--ca-accent);background:var(--soft-blue)';
+            document.body.append(probe);
+            const style = getComputedStyle(probe), tokens = [style.borderTopColor, style.backgroundColor];
+            probe.remove();
+            return tokens;
+        });
+        for (const row of actual.suggestions.rows) expect([row.borderTopColor, row.backgroundColor]).toEqual(accent);
+        expect(actual.dialog.noHorizontalOverflow).toBe(true);
+        expect(actual.noPageScroll).toBe(true);
+    });
+}
+
+test('an empty result matches the concept below its threshold', async ({ page, request }) => {
+    await saveImageTags(request, await conceptPath(request), CONCEPT_TAGS);
+    await expect(await request.post('/__test__/tagging/suggestions', { data: { scored: [] } })).toBeOK();
+    const { reference, actual } = await compareState(page, { theme: 'dark', act: async (page, side) => {
+        // The concept filters its sample by the General confidence; above every sample nothing is shown.
+        if (side === 'concept') await tagDialog(page).getByRole('spinbutton', { name: 'General confidence', exact: true }).fill('1');
+        await generateWith(page, side);
+        if (side === 'concept') await tagDialog(page).getByRole('spinbutton', { name: 'General confidence', exact: true }).fill('0.35');
+    } });
+    expect(actual.suggestions.copy).toBe('No suggestions met this threshold. Adjust it or enter tags manually.');
+    expect(diffTaggingInventory(reference, actual)).toEqual([]);
+});
+
+test('review-scoped dialog shows the scope note', async ({ page, request }) => {
+    const path = await conceptPath(request);
+    await saveImageTags(request, path, CONCEPT_TAGS);
+    await page.setViewportSize({ width: 1600, height: 1100 });
+    await openConcept(page, 'review', 'dark');
+    await page.locator('#main section.surface').getByRole('button', { name: 'Tag image', exact: true }).click();
+    await expect(tagDialog(page)).toBeVisible();
+    await settled(page, 'concept');
+    const opened = await inventory(page);
+    // Needs review ships with PR F; until then the review source opens the same dialog from the module.
+    await leave(page);
+    await openProduction(page, 'images', 'dark');
+    await page.evaluate(async ({ module, path }) => {
+        const { openImageTagging } = await import(module);
+        const { createScope } = await import('/prompt_manager/js/shared/toolkit/core.js');
+        const api = window.ApiClient.createApiClient();
+        const listing = await api.get('/prompt_manager/images/output?mode=workspace&limit=40');
+        openImageTagging({ api, scope: createScope(), image: listing.images.find(item => item.path === path), source: 'review' });
+    }, { module: MODULE, path });
+    await expect(tagDialog(page)).toBeVisible();
+    await settled(page, 'production');
+    const actual = await inventory(page);
+    expect(actual.scopeNote.text).toBe('Tags belong to this image. Resolve its prompt association separately in Needs review.');
+    expect(actual.order).toEqual(['tag-image-heading', 'tag-scope-note small', 'field', 'tag-model-panel', 'tag-suggestion-panel', 'dialog-actions']);
+    const numberField = opened.modelPanel.fields.find(field => field.control.id === 'image-general-threshold');
+    const reference = mapConceptTagging(opened, { instructions: INSTRUCTIONS, chip: opened.heading.size, numberField });
+    expect(diffTaggingInventory(reference, actual)).toEqual([]);
+});
+
+for (const [width, height] of SIZES) test(`card action group matches M1 at ${width}x${height}`, async ({ page }) => {
+    await page.setViewportSize({ width, height });
+    await openConcept(page, 'images', 'dark');
+    const reference = await firstCard(page, 'concept').evaluate(cardActionsInventory);
+    await openProduction(page, 'images', 'dark');
+    const actual = await firstCard(page, 'production').evaluate(cardActionsInventory);
+    expect(actual.buttons.map(button => button.label)).toEqual([`Inspect ${CONCEPT_FILE}`, `Tag or auto-tag ${CONCEPT_FILE}`]);
+    expect(diffTaggingInventory(reference, actual)).toEqual([]);
+});
+
+test('the comparison notices a changed panel border or primary label', async ({ page, request }) => {
+    await saveImageTags(request, await conceptPath(request), CONCEPT_TAGS);
+    const { reference, actual } = await compareState(page, { theme: 'dark' });
+    expect(diffTaggingInventory(reference, actual)).toEqual([]);
+    const undo = await mutate(page, { css: '.tag-model-panel { border: 0 !important; }' });
+    const borderless = withoutChipGeometry(await inventory(page));
+    expect(diffTaggingInventory(reference, borderless).map(item => item.path)).toEqual(
+        expect.arrayContaining(['modelPanel.borderTopWidth', 'modelPanel.borderTopStyle']));
+    await undo();
+    await tagDialog(page).locator('.dialog-actions .ca-primary').evaluate(node => { node.textContent = 'Save'; });
+    const relabelled = withoutChipGeometry(await inventory(page));
+    expect(diffTaggingInventory(reference, relabelled).map(item => item.path)).toEqual(['actions.buttons.1.text']);
+});
+
+// Paired captures for the PR (no screenshot baselines): every production state beside the concept's nearest one. The
+// states the concept lacks (unavailable model, saving, conflict, file changed) pair with the concept's opened dialog.
+test('paired Tag dialog captures', async ({ page, request }, testInfo) => {
+    test.setTimeout(120_000);
+    const path = await conceptPath(request);
+    await saveImageTags(request, path, CONCEPT_TAGS);
+    await page.setViewportSize({ width: 1600, height: 1100 });
+    const shot = async name => {
+        await page.evaluate(async () => { await document.fonts.ready; await Promise.all([...document.images].map(image => image.decode().catch(() => {}))); });
+        const file = testInfo.outputPath(`${name}.png`);
+        await page.screenshot({ path: file });
+        await testInfo.attach(name, { path: file, contentType: 'image/png' });
+    };
+    const both = async (name, act) => {
+        for (const side of ['concept', 'production']) {
+            await openTagDialog(page, side, 'dark');
+            await act(page, side);
+            await shot(`${side}-${name}`);
+        }
+    };
+    await both('opened', async () => {});
+    await both('generated-wd14', (page, side) => generateWith(page, side));
+    await both('generated-gguf', (page, side) => generateWith(page, side, 'gguf'));
+    await expect(await request.post('/__test__/tagging/suggestions', { data: { scored: [] } })).toBeOK();
+    await both('empty', async (page, side) => {
+        if (side === 'concept') await tagDialog(page).getByRole('spinbutton', { name: 'General confidence', exact: true }).fill('1');
+        await generateWith(page, side);
+    });
+    await expect(await request.post('/__test__/tagging/suggestions', { data: { scored: ALPINE } })).toBeOK();
+    const save = () => tagDialog(page).locator('.dialog-actions .ca-primary').click();
+    const status = () => tagDialog(page).locator('#tag-generation-status');
+
+    // Unavailable model.
+    await expect(await request.post('/__test__/tagging/models', { data: { unavailable: ['wd14-swinv2'] } })).toBeOK();
+    await openTagDialog(page, 'production', 'dark');
+    await generateWith(page, 'production').catch(() => {});
+    await expect(status()).toHaveText('WD14 SwinV2 is not downloaded. Download it from the legacy Auto-tag page.');
+    await shot('production-unavailable');
+    await expect(await request.post('/__test__/tagging/models', { data: { unavailable: [] } })).toBeOK();
+
+    // Saving: the write is held until the capture is taken.
+    let release;
+    const held = new Promise(resolve => { release = resolve; });
+    const writes = url => url.pathname === '/prompt_manager/image-annotations/tags';
+    await page.route(writes, async route => { await held; await route.continue(); });
+    await openTagDialog(page, 'production', 'dark');
+    const sent = page.waitForRequest(item => writes(new URL(item.url())));
+    await save();
+    await sent;
+    await expect(tagDialog(page).locator('.dialog-actions .ca-primary')).toHaveAttribute('aria-disabled', 'true');
+    await shot('production-saving');
+    release();
+    await expect(tagDialog(page)).toHaveCount(0);
+    await page.unroute(writes);
+
+    // Conflict: another writer saved this file's set after the dialog observed it.
+    await openTagDialog(page, 'production', 'dark');
+    await saveImageTags(request, path, ['changed_elsewhere']);
+    await save();
+    await expect(status()).not.toHaveText('');
+    await shot('production-conflict');
+    await page.keyboard.press('Escape');
+
+    // File changed: the server's own envelope for bytes replaced after the preview (B1 `file-changed`).
+    await page.route(writes, route => route.fulfill({ status: 409, contentType: 'application/json',
+        body: JSON.stringify({ success: false, code: 'file-changed', error: 'The file changed since it was previewed. Reopen the Tag dialog.' }) }));
+    await saveImageTags(request, path, CONCEPT_TAGS);
+    await openTagDialog(page, 'production', 'dark');
+    await save();
+    await expect(status()).not.toHaveText('');
+    await shot('production-file-changed');
+    await page.unroute(writes);
+    await page.keyboard.press('Escape');
+
+    // Unconfirmed save (a 504): D116's Retry leads the actions, before Cancel and the disabled primary.
+    await page.route(writes, route => route.fulfill({ status: 504, body: 'Gateway Timeout' }));
+    await openTagDialog(page, 'production', 'dark');
+    await save();
+    await expect(status()).toHaveText('The save could not be confirmed. Retry sends the same tags again.');
+    expect((await inventory(page)).actions.buttons.map(button => [button.text, button.primary, button.disabled])).toEqual(
+        [['Retry', false, false], ['Cancel', false, false], ['Save manual tags', true, true]]);
+    await shot('production-unknown-outcome');
+    await page.unroute(writes);
+    await page.keyboard.press('Escape');
+
+    // Review scope: the concept's Needs review dialog and production's review-sourced dialog.
+    await leave(page);
+    await openConcept(page, 'review', 'dark');
+    await page.locator('#main section.surface').getByRole('button', { name: 'Tag image', exact: true }).click();
+    await expect(tagDialog(page)).toBeVisible();
+    await shot('concept-review');
+    await leave(page);
+    await openProduction(page, 'images', 'dark');
+    await page.evaluate(async ({ module, path }) => {
+        const { openImageTagging } = await import(module);
+        const { createScope } = await import('/prompt_manager/js/shared/toolkit/core.js');
+        const api = window.ApiClient.createApiClient();
+        const listing = await api.get('/prompt_manager/images/output?mode=workspace&limit=40');
+        openImageTagging({ api, scope: createScope(), image: listing.images.find(item => item.path === path), source: 'review' });
+    }, { module: MODULE, path });
+    await expect(tagDialog(page).locator('.tag-scope-note')).toBeVisible();
+    await settled(page, 'production');
+    await shot('production-review');
+    await leave(page);
+});

tests/browser/vocabulary-canonical-comparison.mjs

@@ -64,12 +64,10 @@ export function pageInventoryOf(main) {
 // Recorded deviations only (plan "Mock inventory" + decisions D1-D15 + ledger rulings). Everything else must be equal.
 export function mapVocabularyConcept(reference) {
     return {...reference,
-        heading:{...reference.heading,actions:{...reference.heading.actions,
-            // D2: Auto-tag images is omitted until the image-annotation slice owns image tags.
-            buttons:reference.heading.actions.buttons.filter(item=>item.text!=='Auto-tag images')}},
         table:{...reference.table,rows:reference.table.rows.map(row=>({...row,
-            // D3: no image-tag owner yet, so the Images cell shows "—" with a visually hidden "Not available yet".
-            images:'—',imagesHidden:'Not available yet'}))},
+            // Data, not a deviation: the concept tags each sample image with its prompt's tags; the vocabulary fixture
+            // tags no image, so every Images count is 0 (#249 C2 retired #245 D3).
+            images:'0'}))},
         operations:reference.operations.map(row=>({...row,p:{...row.p,
             // Ledger ruling: "1 positive prompts" -> singular, consistent with the shared plural helper (#238/#243).
             text:row.p.text.replace(/^1 positive prompts$/,'1 positive prompt')}})),

tests/browser/vocabulary-visual.spec.mjs

@@ -1,4 +1,4 @@
-import {test,expect} from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import {test,expect} from './fixtures.mjs';
 import {captureCanonicalState} from './history-canonical-comparison.mjs';
 
 // Artifact-only captures of Tags & categories states that have no concept counterpart: the Manage categories

tests/browser/worker-server.spec.mjs

@@ -0,0 +1,35 @@
+// Each Playwright worker owns its server: `request`, `page` and manual
+// `browser.newContext()` all reach it through `baseURL`, and another server
+// process never sees its database.
+import { test, expect } from './fixtures.mjs';
+import { startBrowserServer } from '../../tools/vendor/start-browser-server.mjs';
+
+const search = text => `/prompt_manager/search?text=${encodeURIComponent(text)}`;
+
+test('every fixture reaches this worker\'s own server and another server keeps its own database', async ({ request, page, browser, baseURL, workerServer }, testInfo) => {
+    expect(baseURL).toBe(workerServer.url);
+    await expect(await request.post('/__test__/reset')).toBeOK();
+    const marker = `worker ${testInfo.workerIndex} isolation ${testInfo.project.name} ${testInfo.repeatEachIndex}`;
+    await expect(await request.post('/prompt_manager/save', { data: { text: marker } })).toBeOK();
+    expect((await (await request.get(search(marker))).json()).count).toBe(1);
+
+    const viaPage = await page.request.get(search(marker));
+    expect(viaPage.url().startsWith(`${baseURL}/`)).toBe(true);
+    expect((await viaPage.json()).count).toBe(1);
+    const context = await browser.newContext();
+    try {
+        const response = await (await context.newPage()).goto(search(marker));
+        expect(response.url().startsWith(`${baseURL}/`)).toBe(true);
+        expect((await response.json()).count).toBe(1);
+    } finally {
+        await context.close();
+    }
+
+    const other = await startBrowserServer();
+    try {
+        expect(other.url).not.toBe(baseURL);
+        expect((await (await fetch(`${other.url}${search(marker)}`)).json()).count).toBe(0);
+    } finally {
+        await other.stop();
+    }
+});

tests/browser/workflows.spec.mjs

@@ -1,5 +1,5 @@
 import { startCoverage, saveCoverage } from '../../tools/vendor/browser-coverage.mjs';
-import { test, expect } from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import { test, expect } from './fixtures.mjs';
 
 test.beforeEach(async ({ request, page, browserName }) => {
     await startCoverage(page, browserName);
@@ -64,11 +64,16 @@ test('actual Tags page reloads the selected filter and opens its image viewer',
     await page.goto('/prompt_manager/admin#/tags');
     await page.locator('[data-tag-name="cats"]').click();
     await expect(page.locator('#tagPromptsGrid')).toContainText('Cat portrait');
+    const dogs = page.waitForResponse(response => response.url().includes('/tags/dogs/prompts?'));
     await page.locator('[data-tag-name="dogs"]').click();
+    const [shown] = (await (await dogs).json()).prompts[0].images;
     await expect(page.locator('#tagPromptsGrid')).toContainText('Dog landscape');
     await expect(page.locator('#tagPromptsGrid')).not.toContainText('Cat portrait');
     await page.locator('#tagPromptsGrid img').first().click();
     await expect(page.locator('.viewer-canvas img')).toBeVisible();
+    // The viewer's title names the shown image through the page's title callback. ViewerJS renders it only once the
+    // image is viewed, so closing at "visible" alone left the callback unrun on slower hosts (CI coverage missed it).
+    await expect(page.locator('.viewer-title')).toHaveText(shown.filename);
     await page.locator('.viewer-container').getByRole('button', { name: 'Close', exact: true }).click();
     await expect(page.locator('.viewer-container')).toHaveCount(0);
 });

tests/browser/workspace-activity.spec.mjs

@@ -1,5 +1,5 @@
 import { startCoverage, saveCoverage } from '../../tools/vendor/browser-coverage.mjs';
-import { test, expect } from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import { test, expect } from './fixtures.mjs';
 import { seedConcept, openConcept, openProduction } from './concept-inventory.mjs';
 
 // Activity (#249 Activity-1): the concept's Activity page (app.js renderActivity 732–738), its top-bar button

tests/browser/workspace-history.spec.mjs

@@ -1,5 +1,5 @@
 import {startCoverage,saveCoverage} from '../../tools/vendor/browser-coverage.mjs';
-import {test,expect} from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import {test,expect} from './fixtures.mjs';
 import {readFile,writeFile} from 'node:fs/promises';
 import {createHash} from 'node:crypto';
 
@@ -317,6 +317,19 @@ test('marker labels disclose exact observed dequeue and outcome including times
     }
 });
 
+// A timeline scroll re-renders the markers on the next frame. On a slow runner that frame lands between the press and
+// the release of a marker tap (Playwright's own scroll-into-view scrolls the timeline first); the re-render must leave
+// the pressed marker's nodes alone, or Chromium touch drops the click. The scroll event is dispatched so the marker
+// stays under the pointer, and the two frames are the re-render itself.
+test('a timeline re-render between press and release keeps the marker tap',async({page})=>{
+    const f=page.fixture;await open(page,f.detail_id);
+    const target=marker(page,f.runs.shared,'z');await expect(target).toHaveAttribute('aria-disabled','false');await target.scrollIntoViewIfNeeded();
+    const box=await target.boundingBox();await page.mouse.move(box.x+box.width/2,box.y+box.height/2);await page.mouse.down();
+    await recorded(page).locator('.history-timeline').evaluate(timeline=>{timeline.dispatchEvent(new Event('scroll'));return new Promise(resolve=>requestAnimationFrame(()=>requestAnimationFrame(resolve)));});
+    await page.mouse.up();
+    await expect(dialog(page).locator('[data-receipt]')).toHaveCount(5);
+});
+
 test('same-epoch refresh retains recorded context mode and exact receipt after fresh candidate reads',async({page})=>{
     const f=page.fixture;await open(page,f.detail_id);await marker(page,f.runs.shared,'z').click();await expect(dialog(page).locator('[data-receipt]')).toHaveCount(5);
     await dialog(page).getByRole('button',{name:'Effective',exact:true}).click();
@@ -1006,7 +1019,7 @@ test('shared gallery uses concept cards, image opens the viewer, empty copy',asy
     await expect(star).toHaveAccessibleName(`Favorite ${filename}`);await expect(star).toHaveAttribute('aria-pressed',String(image.favorite));
     await expect(star).toHaveText(image.favorite?'★':'☆');
     await expect(first.locator('.media-info h3')).toHaveCount(1);
-    await expect(first.locator('.image-card-actions > button.icon-btn')).toHaveAccessibleName(`Inspect ${filename}`);
+    await expect(first.locator('.image-card-actions > button.icon-btn:first-child')).toHaveAccessibleName(`Inspect ${filename}`);
     // C3 activation: the image opens the viewer and Escape returns focus to it.
     const openButton=first.locator('.image-open');await openButton.click();
     const viewer=page.locator('dialog#viewer-dialog');await expect(viewer).toBeVisible();

tests/browser/workspace-images.spec.mjs

@@ -1,5 +1,5 @@
 import { startCoverage, saveCoverage } from '../../tools/vendor/browser-coverage.mjs';
-import { test, expect } from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import { test, expect } from './fixtures.mjs';
 import { seedConcept, openConcept, openProduction } from './concept-inventory.mjs';
 
 const workspace = '/prompt_manager/workspace#/images';
@@ -545,7 +545,7 @@ async function staleSourceRecovery(page, request, roots) {
     await expect(await request.post('/prompt_manager/settings', { data: { gallery_root_paths: roots } })).toBeOK();
     const selected = initial.images.at(-1), key = `${selected.root_index}:${selected.path}`;
     const stale = page.waitForResponse(response => response.url().includes('/images/inspect/') && response.status() === 409);
-    await page.locator(`article[data-image-key="${key}"] .image-card-actions button`).click();expect((await stale).status()).toBe(409);
+    await page.locator(`article[data-image-key="${key}"] .image-card-actions button:first-child`).click();expect((await stale).status()).toBe(409);
     await expect(details(page)).toContainText('The source changed. Refresh the image collection before inspecting this file.');
     await page.locator(`article[data-image-key="${key}"] .image-open`).click();
     const viewer = page.locator('dialog#viewer-dialog');await expect(viewer).toContainText('The source changed.');
@@ -1028,8 +1028,7 @@ async function cardInventory(page, concept) {
         const walk = new Function(`return (${walkSource})()`)();
         return nodes.map(card => {
             const copy = card.cloneNode(true);
-            // D9: no tag button until image annotations. The D8 checkbox and the D7 star are compared.
-            if (concept) for (const node of copy.querySelectorAll('[aria-label^="Tag or auto-tag "]')) node.remove();
+            // Info then Tag (#249 C2 retired D9). The D8 checkbox and the D7 star are compared.
             const open = card.querySelector('.image-open'), img = open.querySelector('img');
             const star = card.querySelector('.image-overlays > button.star-button');
             const box = card.querySelector('.image-overlays > label > input');
@@ -1039,7 +1038,7 @@ async function cardInventory(page, concept) {
                 size: [img.getAttribute('width'), img.getAttribute('height')],
                 title: card.querySelector('.media-info h3').textContent, subtitle: card.querySelector('.media-info p').textContent,
                 ratio: card.style.getPropertyValue('--media-ratio').replace(/\s/g, ''),
-                inspect: card.querySelector('.image-card-actions > button').getAttribute('aria-label'),
+                actions: [...card.querySelectorAll('.image-card-actions > button')].map(button => button.getAttribute('aria-label')),
                 star: star && { text: star.textContent, pressed: star.getAttribute('aria-pressed'), label: star.getAttribute('aria-label') },
                 select: box && { type: box.type, checked: box.checked, label: box.getAttribute('aria-label') },
                 paths: walk(copy, `article.${[...card.classList].filter(value => value !== 'selected').sort().join('.')}`, []) };
@@ -1209,7 +1208,7 @@ async function zoomedGallery(page, request, { cssZoom }) {
     await page.keyboard.press('Tab');
     await expect(page.locator(`article[aria-posinset="${columns + 1}"] .image-overlays .star-button`)).toBeFocused();
     await page.keyboard.press('Tab');
-    await expect(page.locator(`article[aria-posinset="${columns + 1}"] .image-card-actions button`)).toBeFocused();
+    await expect(page.locator(`article[aria-posinset="${columns + 1}"] .image-card-actions button:first-child`)).toBeFocused();
     await page.evaluate(() => {
         const card = document.querySelector('article[aria-posinset="1"]');
         const grid = document.querySelector('.image-grid');
@@ -1512,6 +1511,25 @@ test('Refine opened before discovery settles lists the discovered sources and fo
 });
 
 // #249 final review m1: as in the old field row, a Source offers only its own folders.
+test('a search reload keeps the Filters label node, so a click landing during the reload is not lost', async ({ page, request }) => {
+    // WebKit drops a click whose mousedown hit a text node that was replaced before mouseup; the reload reports the
+    // (unchanged) active count, so the label must keep its node.
+    await expect(await request.post('/__test__/reset', { data: { extra_images: 80 } })).toBeOK();
+    await page.goto(workspace);
+    const button = filtersButton(page);
+    await button.click();await expect(refineDialog(page).getByLabel('Folder', { exact: true }).locator('option[value="bulk"]')).toHaveCount(1);await page.keyboard.press('Escape');
+    await refine(page, { Source: '0', Folder: 'bulk' });
+    await expect(button).toHaveText('Filters · 2');
+    const label = await button.evaluateHandle(node => node.firstChild);
+    const reloaded = page.waitForResponse(response => {
+        const url = new URL(response.url());return imagePages(url) && url.searchParams.get('q') === 'png';
+    });
+    await search(page).fill('png');
+    await reloaded;
+    await expect(button).toHaveText('Filters · 2');
+    expect(await button.evaluate((node, text) => node.firstChild === text && text.isConnected, label)).toBe(true);
+});
+
 test('Refine offers only the chosen source folders', async ({ page, request }) => {
     await expect(await request.post('/__test__/add-nested-image')).toBeOK();
     await page.goto(workspace);

tests/browser/workspace-library-editor.spec.mjs

@@ -1,6 +1,6 @@
 import {writeFile} from 'node:fs/promises';
 import {startCoverage,saveCoverage} from '../../tools/vendor/browser-coverage.mjs';
-import {test,expect} from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import {test,expect} from './fixtures.mjs';
 import {captureCanonicalState} from './history-canonical-comparison.mjs';
 import {inventoryOf,mapConcept,diffInventory} from './library-editor-canonical-comparison.mjs';
 

tests/browser/workspace-library.spec.mjs

@@ -1,5 +1,5 @@
 import {startCoverage,saveCoverage} from '../../tools/vendor/browser-coverage.mjs';
-import {test,expect} from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import {test,expect} from './fixtures.mjs';
 
 const route='/prompt_manager/workspace#/library';
 const cards=page=>page.locator('article[data-prompt-id]');
@@ -25,6 +25,8 @@ async function loadShelfPages(page,more,pages) {
         await more.evaluate(node=>node.click());await response;await expect(more).toHaveAttribute('aria-disabled','false');
     }
 }
+/** Resolves in the rendering frame after every frame callback the page had already queued (callbacks run in request order). */
+const nextFrame=page=>page.evaluate(()=>new Promise(resolve=>requestAnimationFrame(()=>resolve())));
 async function openHistory(page,text='Cat portrait') {
     await cards(page).filter({has:page.getByText(text,{exact:true})}).getByRole('button',{name:'Explore history',exact:true}).click();
     await expect(history(page)).toBeVisible();
@@ -103,7 +105,8 @@ test('library infinite loading keeps unique parents and keyboard-accessible hist
         const appended=await (await nextPage).json();expect(appended.items.length).toBeGreaterThan(0);hasMore=appended.has_more;
         if(hasMore)await expect(more).toHaveAttribute('aria-disabled','false');
     }
-    await expect(more).toBeHidden();expect(await cards(page).count()).toBeLessThan(80);
+    // The last page's render queued the window update for the next frame; count after that frame, not before it.
+    await expect(more).toBeHidden();await nextFrame(page);expect(await cards(page).count()).toBeLessThan(80);
     const ids=await cards(page).evaluateAll(nodes=>nodes.map(node=>node.dataset.promptId));expect(new Set(ids).size).toBe(ids.length);
     await expect(page.getByRole('button',{name:/next page|previous page/i})).toHaveCount(0);
     const firstId=await cards(page).first().getAttribute('data-prompt-id');
@@ -253,6 +256,19 @@ test('returning from history restores library context and mobile content stays i
     await expect(history(page)).toHaveCount(0);await expect(page.locator('dialog[open]')).toHaveCount(0);
 });
 
+test('leaving the shelf before its first page answers and returning reads page one without a page error',async({page})=>{
+    // Hold the first shelf page so the route changes while the shelf is still loading (a slow CI runner's natural order).
+    let reached;const firstRead=new Promise(resolve=>{reached=resolve;});
+    await page.route(url=>url.pathname==='/prompt_manager/library'&&!url.searchParams.has('cursor'),route=>{reached();});
+    await page.goto(route);await firstRead;await expect(page.locator('.library-status')).toHaveText('Loading prompt library…');
+    await page.evaluate(()=>{window.location.hash='#/tags';});
+    await expect(page.locator('#workspace-main > .heading h1')).toHaveText('Tags & categories');
+    await page.unrouteAll({behavior:'ignoreErrors'});
+    await page.evaluate(()=>{window.location.hash='#/library';});
+    await expect(page.locator('#workspace-main > .heading h1')).toHaveText('Prompt library');
+    await expect(cards(page)).toHaveCount(3);await expect(page.locator('.section-caption')).toContainText('3 positive prompts');
+    expect(page.libraryErrors).toEqual([]);
+});
 test('empty shelf visits restore saved scroll without inventing a parent anchor',async({page})=>{
     await page.setViewportSize({width:1600,height:300});await page.goto(route);
     await page.getByRole('searchbox',{name:'Search workspace',exact:true}).fill('No matching saved positive');
@@ -376,14 +392,18 @@ async function observeLibrary(page, hold=false) {
             holdFrames(){holdingFrames=true;},
             releaseFrames(){holdingFrames=false;const callbacks=[...heldFrames.values()];heldFrames.clear();for(const callback of callbacks)callback();},
             equal(){for(const record of records){const entries=[...record.latest.values()].filter(entry=>record.targets.has(entry.target));if(entries.length)record.callback(entries,record.native);}},
+            // Resolves once two consecutive rendering frames find no Library frame queued. Each check runs inside a native
+            // frame callback, so a frame the page queues later in that update (a scroll event, a ResizeObserver delivery) is
+            // still pending there. A timer is not a frame: under CPU load a 60 ms wait could end before the scroll's frame.
+            // A frame that never drains (a rAF loop, or settling while frames are held) fails after `limit` frames with the stats.
+            settled(limit=600){return new Promise((resolve,reject)=>{let quiet=0,frames=0;const check=()=>{frames++;quiet=pending.size?0:quiet+1;
+                if(quiet>=2)resolve();else if(frames>=limit)reject(new Error(`Library never settled within ${limit} frames: ${JSON.stringify(window.libraryObservation.stats())}`));else raf(check);};raf(check);});},
             stats(){return {requested,pending:pending.size,heldFrames:heldFrames.size,measurements:records.reduce((sum,record)=>sum+record.entries.size,0),mutations,unmounted,targets:records.reduce((sum,record)=>sum+record.targets.size,0)};},
         };
     },{hold});
 }
 async function settleLibrary(page) {
-    await expect.poll(async()=>page.evaluate(()=>window.libraryObservation.stats().pending)).toBe(0);
-    await page.evaluate(()=>new Promise(resolve=>setTimeout(resolve,60)));
-    await expect.poll(async()=>page.evaluate(()=>window.libraryObservation.stats().pending)).toBe(0);
+    await page.evaluate(()=>window.libraryObservation.settled());
 }
 async function firstVisibleParent(page) {
     return cards(page).evaluateAll(nodes=>{
@@ -683,6 +703,17 @@ test('title and Explore history are buttons that open history',async({page,reque
     await page.keyboard.press('ArrowLeft');await expect(cards(page).nth(1).locator('.card-title > button')).toBeFocused();
 });
 
+test('settling fails with the observation stats when a Library frame stays queued',async({page,request})=>{
+    await request.post('/__test__/reset',{data:{native_library:true,extra_prompts:200}});
+    await observeLibrary(page);await page.addInitScript(()=>{window.IntersectionObserver=undefined;});
+    await page.goto(route);await expect(cards(page).first()).toBeVisible();await settleLibrary(page);
+    await page.evaluate(()=>{window.libraryObservation.holdFrames();dispatchEvent(new Event('scroll'));});
+    await expect.poll(()=>page.evaluate(()=>window.libraryObservation.stats().heldFrames)).toBeGreaterThan(0);
+    const failure=await page.evaluate(()=>window.libraryObservation.settled(30).then(()=>'settled',error=>error.message));
+    expect(failure).toMatch(/^Library never settled within 30 frames: \{.*"heldFrames":[1-9]/);
+    await page.evaluate(()=>window.libraryObservation.releaseFrames());await settleLibrary(page);
+});
+
 test('virtualized rows keep measured heights with the clamp',async({page,request,isMobile})=>{
     await request.post('/__test__/reset',{data:{native_library:true,extra_prompts:200}});
     await observeLibrary(page);await page.addInitScript(()=>{window.IntersectionObserver=undefined;});

tests/browser/workspace-saved-views.spec.mjs

@@ -1,5 +1,5 @@
 import { startCoverage, saveCoverage } from '../../tools/vendor/browser-coverage.mjs';
-import { test, expect } from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import { test, expect } from './fixtures.mjs';
 
 // Saved views (#249 D16, maintainer ruling 2026-10-21): the toolbar's Save view (concept app.js a359, a1330–1346),
 // the sidebar SAVED VIEWS rows (a331–335) and `#/saved/<id>` (a1158–1162), on Images and the Prompt library.

tests/browser/workspace-tagging-lifecycle.spec.mjs

@@ -0,0 +1,71 @@
+// The Tag image dialog's capture lifetime when the page or the dialog goes away mid-flight (#249 PR C final review I4,
+// I5), run on all four projects: leaving the page with the dialog open releases the capture (a `keepalive` DELETE sent
+// on `pagehide`, which also closes the dialog), and closing the dialog while its save is in flight keeps the capture
+// until the save settles, then still reports the save. The fixture is tests/browser/workspace_tagging_fixture.py.
+import { startCoverage, saveCoverage } from '../../tools/vendor/browser-coverage.mjs';
+import { test, expect } from './fixtures.mjs';
+
+const images = '/prompt_manager/workspace#/images';
+const FIRST = 'tagging/dunes_00001.png';
+const TAGS_URL = '**/prompt_manager/image-annotations/tags';
+const dialog = page => page.getByRole('dialog', { name: 'Tag image', exact: true });
+const manual = page => dialog(page).getByRole('textbox', { name: 'Image tags · edit manually', exact: true });
+const tagButton = page => page.locator(`article[data-image-key="0:${FIRST}"]`).getByRole('button', { name: 'Tag or auto-tag dunes_00001.png', exact: true });
+const openCaptures = async request => (await (await request.get('/__test__/tagging/captures')).json()).captures;
+
+test.beforeEach(async ({ page, request, browserName }) => {
+    await startCoverage(page, browserName);
+    await expect(await request.post('/__test__/tagging/reset', { data: {} })).toBeOK();
+    page.problems = [];
+    page.on('pageerror', error => page.problems.push(`pageerror: ${error.message}`));
+    page.on('console', message => { if (message.type() === 'error') page.problems.push(`console: ${message.text()}`); });
+});
+test.afterEach(async ({ page, browserName }) => {
+    await saveCoverage(page, browserName);
+    expect(page.problems).toEqual([]);
+});
+
+async function openTag(page) {
+    await page.goto(images);
+    const observed = page.waitForResponse(item => item.url().endsWith('/image-annotations/observe'));
+    await tagButton(page).click();
+    await expect(dialog(page)).toBeVisible();
+    expect((await observed).status()).toBe(200);
+    await expect(manual(page)).toBeEditable();
+}
+
+test('leaving the page with the dialog open releases its capture without a console error (I4)', async ({ page, request }) => {
+    await openTag(page);
+    expect(await openCaptures(request)).toBe(1);
+    // Another document: the old one's pagehide closes the dialog and sends the release with keepalive.
+    await page.goto('/prompt_manager/workspace?left=1#/tags');
+    await expect(page.locator('#workspace-main h1')).toHaveText('Tags & categories');
+    await expect.poll(() => openCaptures(request)).toBe(0);
+    await expect(dialog(page)).toHaveCount(0);
+});
+
+test('closing the dialog during a save releases the capture after the save and still reports it (I5)', async ({ page, request }) => {
+    await openTag(page);
+    await manual(page).fill('saved_after_close');
+    const order = [];
+    page.on('response', item => { if (item.url().endsWith('/image-annotations/tags')) order.push(`PUT ${item.status()}`); });
+    page.on('request', item => { if (item.url().includes('/image-annotations/capture/') && item.method() === 'DELETE') order.push('DELETE'); });
+    let release;
+    const held = new Promise(resolve => { release = resolve; });
+    const sent = new Promise(resolve => {
+        page.route(TAGS_URL, async route => { resolve(); await held; await route.continue(); });
+    });
+    await dialog(page).getByRole('button', { name: 'Save manual tags', exact: true }).click();
+    await sent;
+    await page.keyboard.press('Escape');
+    await expect(dialog(page)).toHaveCount(0);
+    expect(await openCaptures(request)).toBe(1);
+    expect(order).toEqual([]);
+    release();
+    await expect(page.locator('#workspace-toast')).toHaveText('1 tags saved for dunes_00001.png.');
+    await expect.poll(() => order).toEqual(['PUT 200', 'DELETE']);
+    await expect.poll(() => openCaptures(request)).toBe(0);
+    // The page followed the late save: the selected file's details show the new image tag.
+    await expect(page.getByRole('complementary', { name: 'Selected image details', exact: true }).getByText('saved_after_close', { exact: true })).toBeVisible();
+    await page.unroute(TAGS_URL);
+});

tests/browser/workspace-tagging.spec.mjs

@@ -0,0 +1,749 @@
+// The Tag image dialog, the card Tag button and the Tags page hooks (#249 PR C, plan Task C2, Mock inventory M1, M2, M5)
+// against the real routes and the injected fake model (tests/browser/workspace_tagging_fixture.py). No model is ever
+// downloaded or loaded: the fake answers, and every test checks that no download request was sent.
+import { startCoverage, saveCoverage } from '../../tools/vendor/browser-coverage.mjs';
+import { test, expect } from './fixtures.mjs';
+import { saveImageTags } from './concept-inventory.mjs';
+import { taggingInventory } from './tagging-canonical-comparison.mjs';
+
+const images = '/prompt_manager/workspace#/images';
+const MODULE = '/prompt_manager/js/workspace/image-tagging.mjs';
+const MODEL_LABELS = ['WD14 SwinV2 · ~400 MB', 'WD14 ViT · ~400 MB', 'GGUF JoyCaption (Recommended) · ~4.5 GB', 'HuggingFace JoyCaption · ~16 GB'];
+// The concept's on-screen opening sentence (its renderTagModelOptions replaces the markup's at once, D83).
+const MODEL_CHANGED = 'Generate suggestions with the selected model before approving.';
+const CHIP_HINT = 'Remove anything you do not want to add.';
+const ALL_REMOVED = 'All suggestions removed. Generate again or enter tags manually.';
+const card = (page, path) => page.locator(`article[data-image-key="0:${path}"]`);
+const tagButton = (page, path) => card(page, path).getByRole('button', { name: `Tag or auto-tag ${path.split('/').pop()}`, exact: true });
+const dialog = page => page.getByRole('dialog', { name: 'Tag image', exact: true });
+const manual = page => dialog(page).getByRole('textbox', { name: 'Image tags · edit manually', exact: true });
+const model = page => dialog(page).getByRole('combobox', { name: 'Tagging model', exact: true });
+const generate = page => dialog(page).getByRole('button', { name: 'Generate suggestions', exact: true });
+const chips = page => dialog(page).locator('#tag-suggestions .chip.tag-suggestion');
+const status = page => dialog(page).locator('#tag-generation-status');
+const toast = page => page.locator('#workspace-toast');
+const nav = (page, label) => page.locator('aside.sidebar nav').getByRole('link', { name: new RegExp(`^${label}(?:\\s*[\\d,]+)?$`) });
+const FIRST = 'tagging/dunes_00001.png', SECOND = 'tagging/dunes_00002.png', COPY = 'copies/dunes_00001.png';
+const VIDEO = 'tagging/clip.webm', UNLINKED = 'loose/unlinked.png';
+
+let manifest;
+test.beforeEach(async ({ page, request, browserName }) => {
+    await startCoverage(page, browserName);
+    const response = await request.post('/__test__/tagging/reset', { data: {} });
+    await expect(response).toBeOK();
+    manifest = await response.json();
+    page.taggingErrors = [];
+    page.on('pageerror', error => page.taggingErrors.push(error.message));
+    page.downloads = [];
+    page.on('request', item => { if (/\/autotag\/download\//.test(item.url())) page.downloads.push(item.url()); });
+});
+test.afterEach(async ({ page, browserName }) => {
+    await saveCoverage(page, browserName);
+    expect(page.taggingErrors).toEqual([]);
+    expect(page.downloads).toEqual([]);
+});
+
+/** Images, then the Tag dialog of `path`, settled on its observation (the manual field holds the saved tags). */
+async function openTag(page, path = FIRST, { goto = true } = {}) {
+    if (goto) await page.goto(images);
+    const observed = page.waitForResponse(item => item.url().endsWith('/image-annotations/observe'));
+    await tagButton(page, path).click();
+    await expect(dialog(page)).toBeVisible();
+    expect((await observed).status()).toBe(200);
+    await expect(dialog(page).locator('img.tag-image-preview, video.tag-image-preview')).toHaveAttribute('src', /\/prompt_manager\/image-annotations\/capture\//);
+}
+async function generated(page) {
+    const answer = page.waitForResponse(item => item.url().endsWith('/image-annotations/suggest'));
+    await generate(page).click();
+    return answer;
+}
+/** Read the saved tags of `path` through the real observe route, then release that capture. */
+async function savedTags(request, path) {
+    const target = { root_index: 0, root_scope: await scopeOf(request), path };
+    const observed = await (await request.post('/prompt_manager/image-annotations/observe', { data: { target } })).json();
+    expect(observed.success).toBe(true);
+    await request.delete(`/prompt_manager/image-annotations/capture/${observed.capture.token}`);
+    return observed.tags;
+}
+async function scopeOf(request) {
+    const listing = await (await request.get('/prompt_manager/images/output?mode=workspace&limit=40')).json();
+    return listing.images[0].root_scope;
+}
+
+test('tag dialog control inventory and order match M2', async ({ page }) => {
+    await openTag(page);
+    await expect(page.locator('dialog#modal')).toHaveClass(/(^| )tagging-modal( |$)/);
+    await expect(dialog(page).locator('#modal-title')).toHaveText('Tag image');
+    const body = dialog(page).locator('#modal-body');
+    expect(await body.evaluate(node => [...node.children].map(child => child.className))).toEqual(
+        ['tag-image-heading', 'field', 'tag-model-panel', 'tag-suggestion-panel', 'dialog-actions']);
+    const heading = body.locator('.tag-image-heading');
+    await expect(heading.locator('h3')).toHaveText(manifest.title);
+    await expect(heading.locator('p.small')).toHaveText('dunes_00001.png');
+    await expect(heading.locator('span.chip')).toHaveText('96 × 64');
+    await expect(heading.locator('img.tag-image-preview')).toHaveAttribute('alt', manifest.title);
+    await expect(body.locator('.tag-scope-note')).toHaveCount(0);
+    await expect(manual(page)).toHaveAttribute('id', 'manual-image-tags');
+    await expect(manual(page)).toHaveValue('');
+    await expect(manual(page)).toHaveAccessibleDescription('Comma-separated. Generated suggestions are added only after you approve them.');
+    const panel = body.locator('.tag-model-panel');
+    await expect(panel.locator('h3')).toHaveText('Auto-tag this image');
+    await expect(model(page).locator('option')).toHaveText(MODEL_LABELS);
+    expect(await model(page).locator('option').evaluateAll(nodes => nodes.map(node => node.value))).toEqual(['wd14-swinv2', 'wd14-vit', 'gguf', 'hf']);
+    await expect(model(page)).toHaveValue('wd14-swinv2');
+    const options = panel.locator('#image-tag-model-options');
+    const general = options.getByRole('spinbutton', { name: 'General confidence', exact: true });
+    const character = options.getByRole('spinbutton', { name: 'Character confidence', exact: true });
+    const limit = options.getByRole('spinbutton', { name: 'Suggestions', exact: true });
+    await expect(options.locator('.two-col .field')).toHaveCount(2);
+    await expect(general).toHaveValue('0.35');
+    await expect(character).toHaveValue('0.85');
+    await expect(limit).toHaveValue('10');
+    for (const [input, attributes] of [[general, ['0', '1', '0.05']], [character, ['0', '1', '0.05']], [limit, ['1', '100', '1']]]) {
+        expect(await input.evaluate(node => [node.min, node.max, node.step])).toEqual(attributes);
+    }
+    await expect(generate(page)).toHaveClass(/ca-primary/);
+    await expect(generate(page).locator('svg.icon')).toHaveCount(1);
+    const suggestions = body.locator('section.tag-suggestion-panel');
+    await expect(suggestions).toHaveAttribute('aria-labelledby', 'tag-suggestions-title');
+    await expect(suggestions.locator('h3#tag-suggestions-title')).toHaveText('Review suggestions');
+    await expect(suggestions.locator('#suggestion-count')).toHaveText('Not generated');
+    await expect(suggestions.locator('#tag-suggestions > p.small')).toHaveText(MODEL_CHANGED);
+    await expect(status(page)).toHaveAttribute('role', 'status');
+    await expect(status(page)).toHaveText('');
+    await expect(body.locator('.dialog-actions button')).toHaveText(['Cancel', 'Save manual tags']);
+    await expect(body.locator('.dialog-actions button').last()).toHaveClass(/ca-primary/);
+    await model(page).selectOption('gguf');
+    await expect(options.getByRole('textbox', { name: 'Tagging instructions', exact: true })).toHaveValue('Describe the image');
+    await expect(options.getByRole('spinbutton', { name: 'Suggestions', exact: true })).toHaveValue('10');
+    await expect(options.locator('.two-col')).toHaveCount(0);
+});
+
+test('generate shows real fixture percentages for WD14 and Suggested for GGUF, at most ten chips', async ({ page }) => {
+    await openTag(page);
+    const answer = await generated(page);
+    expect(answer.request().postDataJSON()).toMatchObject({ model_type: 'wd14-swinv2', options: { general_threshold: 0.35, character_threshold: 0.85, suggestion_limit: 10 } });
+    await expect(chips(page)).toHaveCount(10);
+    // D83: ruling T4's instruction leads the chips, only while chips exist.
+    await expect(dialog(page).locator('#tag-suggestions > *')).toHaveCount(2);
+    await expect(dialog(page).locator('#tag-suggestions > p.small')).toHaveText(CHIP_HINT);
+    await expect(chips(page).first().locator(':scope > span').first()).toHaveText('sandstone');
+    await expect(chips(page).locator('.tag-confidence').first()).toHaveText('98%');
+    await expect(chips(page).nth(1).locator('.tag-confidence')).toHaveText('97%');
+    await expect(chips(page).first().getByRole('button', { name: 'Remove sandstone', exact: true })).toBeVisible();
+    await expect(dialog(page).locator('#suggestion-count')).toHaveText('10 suggestions');
+    await expect(status(page)).toHaveText('Suggestions from WD14 SwinV2 · ~400 MB. Approve tags to save.');
+    await expect(dialog(page).locator('.dialog-actions .ca-primary')).toHaveText('Approve selected tags');
+    for (const id of ['gguf', 'hf']) {
+        await model(page).selectOption(id);
+        await expect(chips(page)).toHaveCount(0);
+        await generated(page);
+        await expect(chips(page)).toHaveCount(10);
+        await expect(chips(page).locator('.tag-confidence')).toHaveText(Array(10).fill('Suggested'));
+        await expect(status(page)).toHaveText(`Suggestions from ${MODEL_LABELS[id === 'gguf' ? 2 : 3]}. Approve tags to save.`);
+    }
+    await dialog(page).getByRole('spinbutton', { name: 'Suggestions', exact: true }).fill('3');
+    const fewer = await generated(page);
+    expect(fewer.request().postDataJSON().options).toEqual({ instructions: 'Describe the image', suggestion_limit: 3 });
+    await expect(chips(page)).toHaveCount(3);
+    await expect(dialog(page).locator('#suggestion-count')).toHaveText('3 suggestions');
+});
+
+test('removing chips and editing manually then approving persists across reload', async ({ page, request }) => {
+    await openTag(page);
+    await manual(page).fill('Minimal, sky_blue');
+    await dialog(page).getByRole('spinbutton', { name: 'Suggestions', exact: true }).fill('3');
+    await generated(page);
+    await expect(chips(page)).toHaveCount(3);
+    const remove = dialog(page).getByRole('button', { name: 'Remove desert', exact: true });
+    await remove.click();
+    await expect(chips(page)).toHaveCount(2);
+    // Focus moves to the next chip's ×.
+    await expect(dialog(page).getByRole('button', { name: 'Remove arch', exact: true })).toBeFocused();
+    const saved = page.waitForResponse(item => item.url().endsWith('/image-annotations/tags') && item.request().method() === 'PUT');
+    await dialog(page).getByRole('button', { name: 'Approve selected tags', exact: true }).click();
+    expect((await saved).request().postDataJSON().tags).toEqual(['Minimal', 'sky_blue', 'sandstone', 'arch']);
+    await expect(dialog(page)).toHaveCount(0);
+    await expect(toast(page)).toHaveText('4 tags saved for dunes_00001.png.');
+    await expect(tagButton(page, FIRST)).toBeFocused();
+    expect(await savedTags(request, FIRST)).toEqual(['Minimal', 'sky_blue', 'sandstone', 'arch']);
+    await page.reload();
+    await openTag(page, FIRST, { goto: false });
+    await expect(manual(page)).toHaveValue('Minimal, sky_blue, sandstone, arch');
+    // Removing every chip shows the all-removed copy; the manual set still saves.
+    await dialog(page).getByRole('spinbutton', { name: 'Suggestions', exact: true }).fill('2');
+    await generated(page);
+    for (const name of ['sandstone', 'desert']) await dialog(page).getByRole('button', { name: `Remove ${name}`, exact: true }).click();
+    await expect(dialog(page).locator('#tag-suggestions > p.small')).toHaveText(ALL_REMOVED);
+    await expect(generate(page)).toBeFocused();
+    await dialog(page).getByRole('button', { name: 'Approve selected tags', exact: true }).click();
+    await expect(toast(page)).toHaveText('4 tags saved for dunes_00001.png.');
+});
+
+test('cancel escape and close write nothing', async ({ page, request }) => {
+    const writes = [];
+    page.on('request', item => { if (item.url().endsWith('/image-annotations/tags')) writes.push(item.method()); });
+    await page.goto(images);
+    for (const dismiss of [
+        () => dialog(page).getByRole('button', { name: 'Cancel', exact: true }).click(),
+        () => page.keyboard.press('Escape'),
+        () => dialog(page).getByRole('button', { name: 'Close dialog', exact: true }).click(),
+    ]) {
+        const released = page.waitForResponse(item => item.url().includes('/image-annotations/capture/') && item.request().method() === 'DELETE');
+        await openTag(page, FIRST, { goto: false });
+        await manual(page).fill('never_saved');
+        await dismiss();
+        await expect(dialog(page)).toHaveCount(0);
+        expect((await released).ok()).toBe(true);
+        await expect(tagButton(page, FIRST)).toBeFocused();
+        await expect(page.locator('dialog#modal.tagging-modal')).toHaveCount(0);
+    }
+    expect(writes).toEqual([]);
+    expect(await savedTags(request, FIRST)).toEqual([]);
+});
+
+test('model change discards suggestions and relabels primary', async ({ page }) => {
+    await openTag(page);
+    await manual(page).fill('kept');
+    await generated(page);
+    await expect(dialog(page).locator('.dialog-actions .ca-primary')).toHaveText('Approve selected tags');
+    await dialog(page).getByRole('spinbutton', { name: 'General confidence', exact: true }).fill('0.5');
+    await model(page).selectOption('wd14-vit');
+    await expect(chips(page)).toHaveCount(0);
+    await expect(dialog(page).locator('#tag-suggestions > p.small')).toHaveText(MODEL_CHANGED);
+    await expect(dialog(page).locator('#suggestion-count')).toHaveText('Not generated');
+    await expect(status(page)).toHaveText('');
+    await expect(dialog(page).locator('.dialog-actions .ca-primary')).toHaveText('Save manual tags');
+    await expect(manual(page)).toHaveValue('kept');
+    // Like the concept, the new model's fields start from the defaults.
+    await expect(dialog(page).getByRole('spinbutton', { name: 'General confidence', exact: true })).toHaveValue('0.35');
+});
+
+test('invalid threshold reports validity without request', async ({ page }) => {
+    const sent = [];
+    page.on('request', item => { if (item.url().endsWith('/image-annotations/suggest')) sent.push(item.url()); });
+    await openTag(page);
+    const general = dialog(page).getByRole('spinbutton', { name: 'General confidence', exact: true });
+    await general.fill('1.5');
+    await generate(page).click();
+    await expect(general).toBeFocused();
+    expect(await general.evaluate(node => node.validity.rangeOverflow)).toBe(true);
+    await general.fill('');
+    await generate(page).click();
+    await expect.poll(() => general.evaluate(node => node.validationMessage)).toBe('Enter a confidence from 0 to 1.');
+    await general.fill('0.35');
+    expect(await general.evaluate(node => node.validity.valid)).toBe(true);
+    await model(page).selectOption('gguf');
+    const instructions = dialog(page).getByRole('textbox', { name: 'Tagging instructions', exact: true });
+    await instructions.fill('   ');
+    await generate(page).click();
+    await expect.poll(() => instructions.evaluate(node => node.validationMessage)).toBe('Enter tagging instructions of at most 4000 characters.');
+    await expect(status(page)).toHaveText('');
+    expect(sent).toEqual([]);
+});
+
+test('unavailable model shows inline status and never downloads', async ({ page, request }) => {
+    await expect(await request.post('/__test__/tagging/models', { data: { unavailable: ['wd14-swinv2'] } })).toBeOK();
+    await openTag(page);
+    expect((await generated(page)).status()).toBe(409);
+    await expect(status(page)).toHaveText('WD14 SwinV2 is not downloaded. Download it from the legacy Auto-tag page.');
+    await expect(chips(page)).toHaveCount(0);
+    await expect(dialog(page).locator('#suggestion-count')).toHaveText('Not generated');
+    // D116: Retry leads the actions while the controller allows it, and generates again.
+    await expect(dialog(page).locator('.dialog-actions button')).toHaveText(['Retry', 'Cancel', 'Save manual tags']);
+    await expect(await request.post('/__test__/tagging/models', { data: { unavailable: [] } })).toBeOK();
+    const again = page.waitForResponse(item => item.url().endsWith('/image-annotations/suggest'));
+    await dialog(page).getByRole('button', { name: 'Retry', exact: true }).click();
+    expect((await again).status()).toBe(200);
+    await expect(chips(page)).toHaveCount(10);
+    await expect(dialog(page).locator('.dialog-actions button')).toHaveText(['Cancel', 'Approve selected tags']);
+    await expect(dialog(page).locator('.dialog-actions .ca-primary')).toBeFocused();
+});
+
+test('video tag is manual only', async ({ page, request }) => {
+    await openTag(page, VIDEO);
+    const preview = dialog(page).locator('video.tag-image-preview');
+    await expect(preview).toHaveAttribute('preload', 'metadata');
+    expect(await preview.evaluate(node => node.muted)).toBe(true);
+    await expect(dialog(page).locator('.tag-image-heading span.chip')).toHaveCount(0);
+    await expect(generate(page)).toHaveAttribute('aria-disabled', 'true');
+    await expect(status(page)).toHaveText('Auto-tagging is available for supported still images only.');
+    let suggested = false;
+    page.on('request', item => { if (item.url().endsWith('/image-annotations/suggest')) suggested = true; });
+    // Pressing the aria-disabled Generate sends nothing.
+    await generate(page).focus();
+    await page.keyboard.press('Enter');
+    await manual(page).fill('clip_tag');
+    await dialog(page).getByRole('button', { name: 'Save manual tags', exact: true }).click();
+    await expect(toast(page)).toHaveText('1 tags saved for clip.webm.');
+    expect(suggested).toBe(false);
+    expect(await savedTags(request, VIDEO)).toEqual(['clip_tag']);
+});
+
+test('sticky footer reachable at 390x844 while content scrolls', async ({ page }) => {
+    await page.setViewportSize({ width: 390, height: 844 });
+    await openTag(page);
+    await model(page).selectOption('gguf');
+    await generated(page);
+    await expect(chips(page)).toHaveCount(10);
+    const shell = page.locator('dialog#modal');
+    const actions = dialog(page).locator('.dialog-actions');
+    expect(await actions.evaluate(node => getComputedStyle(node).position)).toBe('sticky');
+    expect(await shell.evaluate(node => node.scrollHeight > node.clientHeight)).toBe(true);
+    for (const top of [0, 10_000]) {
+        await shell.evaluate((node, value) => { node.scrollTop = value; }, top);
+        const box = await actions.boundingBox(), frame = await shell.boundingBox();
+        expect(box.y + box.height).toBeLessThanOrEqual(frame.y + frame.height + 1);
+        expect(box.y).toBeGreaterThanOrEqual(frame.y);
+        await expect(dialog(page).getByRole('button', { name: 'Approve selected tags', exact: true })).toBeInViewport();
+    }
+    expect(await page.evaluate(() => document.documentElement.scrollWidth <= innerWidth + 1)).toBe(true);
+});
+
+test('focus starts on close and returns to the Tag trigger', async ({ page }) => {
+    await openTag(page, SECOND);
+    await expect(dialog(page).getByRole('button', { name: 'Close dialog', exact: true })).toBeFocused();
+    await page.keyboard.press('Escape');
+    await expect(dialog(page)).toHaveCount(0);
+    await expect(tagButton(page, SECOND)).toBeFocused();
+});
+
+test('every Images and Library gallery card has Info then Tag with mock labels', async ({ page }) => {
+    const expectCards = async locator => {
+        const labels = await locator.evaluateAll(nodes => nodes.map(node => {
+            const buttons = [...node.querySelectorAll('.media-info > .image-card-actions > button')];
+            return { file: node.querySelector('.image-open').getAttribute('aria-label').replace(/^Open /, ''),
+                buttons: buttons.map(button => [button.className, button.getAttribute('aria-label'), button.querySelector('svg.icon') !== null]) };
+        }));
+        expect(labels.length).toBeGreaterThan(0);
+        for (const { file, buttons } of labels) {
+            expect(buttons).toEqual([['icon-btn', `Inspect ${file}`, true], ['icon-btn', `Tag or auto-tag ${file}`, true]]);
+        }
+        return labels.map(({ file }) => file);
+    };
+    await page.goto(images);
+    await expect(page.locator('article[data-image-key]')).toHaveCount(manifest.files.length);
+    expect(await expectCards(page.locator('article[data-image-key]'))).toEqual(manifest.files.map(file => file.filename));
+    await page.goto(`/prompt_manager/workspace#/library/${manifest.positive_id}`);
+    const gallery = page.locator('.library-history .history-gallery article[data-image-key]');
+    await expect(gallery.first()).toBeVisible();
+    await expectCards(gallery);
+    const observed = page.waitForResponse(item => item.url().endsWith('/image-annotations/observe'));
+    await gallery.first().getByRole('button', { name: /^Tag or auto-tag / }).click();
+    await expect(dialog(page)).toBeVisible();
+    expect((await observed).status()).toBe(200);
+    await dialog(page).getByRole('button', { name: 'Cancel', exact: true }).click();
+    await expect(gallery.first().getByRole('button', { name: /^Tag or auto-tag / })).toBeFocused();
+});
+
+test('Tag opens the dialog without selecting the card', async ({ page }) => {
+    const panel = page.getByRole('complementary', { name: 'Selected image details', exact: true });
+    await page.goto(images);
+    await expect(card(page, FIRST)).toHaveClass(/\bselected\b/);
+    const before = await panel.innerText();
+    await openTag(page, SECOND, { goto: false });
+    await page.keyboard.press('Escape');
+    await expect(card(page, SECOND)).not.toHaveClass(/\bselected\b/);
+    await expect(card(page, FIRST)).toHaveClass(/\bselected\b/);
+    expect(await panel.innerText()).toBe(before);
+});
+
+test('tagging one of two cards sharing a prompt leaves the other card, the prompt and the pair unchanged', async ({ page, request }) => {
+    const before = await (await request.get('/__test__/tagging/snapshot')).json();
+    expect(await savedTags(request, SECOND)).toEqual([]);
+    expect(await savedTags(request, COPY)).toEqual([]);
+    const pairBefore = await (await request.get(`/prompt_manager/library/${manifest.positive_id}`)).json();
+    await openTag(page, FIRST);
+    await manual(page).fill('only_this_file');
+    await dialog(page).getByRole('button', { name: 'Save manual tags', exact: true }).click();
+    await expect(toast(page)).toHaveText('1 tags saved for dunes_00001.png.');
+    expect(await savedTags(request, FIRST)).toEqual(['only_this_file']);
+    expect(await savedTags(request, SECOND)).toEqual([]);
+    expect(await savedTags(request, COPY)).toEqual([]);
+    expect(await (await request.get('/__test__/tagging/snapshot')).json()).toEqual(before);
+    const pairAfter = await (await request.get(`/prompt_manager/library/${manifest.positive_id}`)).json();
+    expect(pairAfter.success).toBe(true);
+    expect({ ...pairAfter, library_epoch: null }).toEqual({ ...pairBefore, library_epoch: null });
+});
+
+test('Tags page Auto-tag images opens the dialog for the selected image and is disabled without one', async ({ page }) => {
+    await page.goto('/prompt_manager/workspace#/tags');
+    const autoTag = page.locator('#workspace-main .heading-actions').getByRole('button', { name: 'Auto-tag images', exact: true });
+    await expect(page.locator('#workspace-main .heading-actions button')).toHaveText(['+ Create tag', 'Auto-tag images']);
+    await expect(autoTag).toHaveClass(/ca-primary/);
+    await expect(autoTag).toHaveAttribute('aria-disabled', 'true');
+    // An aria-disabled control still takes a press; it only explains.
+    await autoTag.focus();
+    await page.keyboard.press('Enter');
+    await expect(toast(page)).toHaveText('Select an image in Images first.');
+    await expect(dialog(page)).toHaveCount(0);
+    await nav(page, 'Images').click();
+    await card(page, SECOND).getByRole('button', { name: 'Inspect dunes_00002.png', exact: true }).click();
+    await expect(card(page, SECOND)).toHaveClass(/selected/);
+    await nav(page, 'Tags & categories').click();
+    await expect(autoTag).not.toHaveAttribute('aria-disabled', 'true');
+    const observed = page.waitForResponse(item => item.url().endsWith('/image-annotations/observe'));
+    await autoTag.click();
+    expect((await observed).request().postDataJSON().target.path).toBe(SECOND);
+    await expect(dialog(page).locator('.tag-image-heading p.small')).toHaveText('dunes_00002.png');
+    await page.keyboard.press('Escape');
+    await expect(autoTag).toBeFocused();
+});
+
+test('Images column shows image_count and updates after save', async ({ page }) => {
+    await page.goto(images);
+    await card(page, FIRST).getByRole('button', { name: 'Inspect dunes_00001.png', exact: true }).click();
+    await nav(page, 'Tags & categories').click();
+    const row = name => page.locator('#workspace-main table tbody tr').filter({ has: page.locator('td:first-child').getByText(name, { exact: true }) });
+    await expect(row('minimal').locator('td:nth-child(3)')).toHaveText('0');
+    await expect(row('editorial').locator('td:nth-child(3)')).toHaveText('0');
+    await page.locator('#workspace-main .heading-actions').getByRole('button', { name: 'Auto-tag images', exact: true }).click();
+    await expect(manual(page)).toBeVisible();
+    await expect(dialog(page).locator('img.tag-image-preview')).toHaveAttribute('src', /capture/);
+    await manual(page).fill('minimal, editorial');
+    const reread = page.waitForResponse(item => item.url().includes('/prompt_manager/vocabulary/tags'));
+    await dialog(page).getByRole('button', { name: 'Save manual tags', exact: true }).click();
+    await reread;
+    await expect(row('minimal').locator('td:nth-child(3)')).toHaveText('1');
+    await expect(row('editorial').locator('td:nth-child(3)')).toHaveText('1');
+    await expect(row('warm_light').locator('td:nth-child(3)')).toHaveText('0');
+});
+
+test('virtualized cards release listeners on unmount', async ({ page, request }) => {
+    const extra = 140;
+    await expect(await request.post('/__test__/tagging/reset', { data: { extra_images: extra } })).toBeOK();
+    const count = () => page.evaluate(async url => (await import(url)).imageTagsListenerCount(), MODULE);
+    await page.goto(images);
+    await expect(card(page, FIRST)).toBeVisible();
+    const mounted = await count();
+    // Leaving and returning to Images several times keeps one subscription per mounted page.
+    for (let round = 0; round < 3; round += 1) {
+        await nav(page, 'Tags & categories').click();
+        await expect(page.locator('#workspace-main h1')).toHaveText('Tags & categories');
+        await nav(page, 'Images').click();
+        await expect(card(page, FIRST)).toBeVisible();
+    }
+    expect(await count()).toBe(mounted);
+    // Cards scrolled out are unmounted; the delegated Tag still works on a card rendered again.
+    const continuation = page.locator('#gallery-continuation');
+    await expect(async () => {
+        await continuation.scrollIntoViewIfNeeded();
+        await expect(continuation).toContainText('You’re all caught up', { timeout: 1000 });
+    }).toPass();
+    await continuation.scrollIntoViewIfNeeded();
+    await expect(card(page, FIRST)).toHaveCount(0);
+    expect(await count()).toBe(mounted);
+    await page.evaluate(() => window.scrollTo(0, 0));
+    await expect(card(page, FIRST)).toBeVisible();
+    await openTag(page, FIRST, { goto: false });
+    await page.keyboard.press('Escape');
+    await expect(tagButton(page, FIRST)).toBeFocused();
+    expect(await count()).toBe(mounted);
+});
+
+// Task C3 (D88): the side panel and the viewer details draw the file's own current image tags first, then D21's linked
+// tags not already shown (case-insensitive). Panel image tags are Tags filter buttons; the viewer's chips are plain.
+const panel = page => page.getByRole('complementary', { name: 'Selected image details', exact: true });
+const tagRow = page => panel(page).locator('.detail-inner > .tags').last();
+const inspected = (page, path) => page.waitForResponse(item => new URL(item.url()).pathname === `/prompt_manager/images/inspect/${path}`);
+const infoButton = (page, path) => card(page, path).getByRole('button', { name: `Inspect ${path.split('/').pop()}`, exact: true });
+
+test('inspector shows saved image tags after Save without reselecting', async ({ page }) => {
+    await page.goto(images);
+    await expect(card(page, FIRST)).toHaveClass(/\bselected\b/);
+    // D21/D39: the linked recipe's own tag, plain.
+    await expect(tagRow(page).locator('.chip')).toHaveText(manifest.pair_tags);
+    await expect(tagRow(page).locator('button.chip')).toHaveCount(0);
+    await openTag(page, FIRST, { goto: false });
+    await manual(page).fill('dunes_tag, Editorial');
+    const reread = inspected(page, FIRST);
+    await dialog(page).getByRole('button', { name: 'Save manual tags', exact: true }).click();
+    expect((await (await reread).json()).image_tags).toEqual({ state: 'current', tags: ['dunes_tag', 'Editorial'] });
+    // The image tags lead as filter buttons; the recipe's "editorial" is already shown as "Editorial".
+    await expect(tagRow(page).locator('.chip')).toHaveText(['dunes_tag', 'Editorial']);
+    await expect(tagRow(page).locator('button.chip')).toHaveText(['dunes_tag', 'Editorial']);
+    await expect(card(page, FIRST)).toHaveClass(/\bselected\b/);
+    await expect(panel(page).locator('.panel-heading h2')).toHaveText(manifest.title);
+});
+
+test('clicking an image tag chip filters Images to the tagged unlinked file', async ({ page, request }) => {
+    expect(await saveImageTags(request, UNLINKED, ['loose_tag'])).toEqual(['loose_tag']);
+    await page.goto(images);
+    const read = inspected(page, UNLINKED);
+    await infoButton(page, UNLINKED).click();
+    await read;
+    await expect(panel(page).locator('.panel-heading h2')).toHaveText('unlinked.png');
+    await expect(tagRow(page).locator('.chip')).toHaveText(['loose_tag']);
+    const listing = page.waitForRequest(item => item.url().includes('/images/output?') && new URL(item.url()).searchParams.get('tags') === 'loose_tag');
+    await tagRow(page).getByRole('button', { name: 'loose_tag', exact: true }).click();
+    await listing;
+    await expect(page.locator('.gallery-body article[data-image-key]')).toHaveCount(1);
+    await expect(card(page, UNLINKED)).toBeVisible();
+    await expect(page.getByRole('button', { name: 'Filters · 1', exact: true })).toBeVisible();
+});
+
+test('viewer details show the same chips', async ({ page, request }) => {
+    await saveImageTags(request, FIRST, ['dunes_tag', 'Editorial', 'sunlit']);
+    await page.goto(images);
+    await expect(tagRow(page).locator('button.chip')).toHaveText(['dunes_tag', 'Editorial', 'sunlit']);
+    await card(page, FIRST).locator('.image-open').click();
+    const details = page.locator('#viewer-dialog[open] #viewer-details');
+    await expect(details.locator('.panel-heading h2')).toHaveText(manifest.title);
+    const row = details.locator('.tags').last();
+    await expect(row.locator('.chip')).toHaveText(['dunes_tag', 'Editorial', 'sunlit']);
+    await expect(row.locator('span.chip')).toHaveCount(3);
+    await expect(row.locator('button')).toHaveCount(0);
+    // A file without image tags: the viewer shows only D21's linked tags.
+    await page.getByRole('button', { name: 'Next media', exact: true }).click();
+    await expect(page.locator('#viewer-dialog[open] #viewer-title')).toHaveText(`${manifest.title} · dunes_00002.png`);
+    await expect(details.locator('.tags').last().locator('.chip')).toHaveText(manifest.pair_tags);
+});
+
+test('search finds an image by its image tag', async ({ page, request }) => {
+    await saveImageTags(request, UNLINKED, ['Loose_Tag']);
+    await page.goto(images);
+    await expect(page.locator('.gallery-body article[data-image-key]')).toHaveCount(manifest.files.length);
+    const listing = page.waitForRequest(item => item.url().includes('/images/output?') && new URL(item.url()).searchParams.get('q') === 'loose_t');
+    await page.getByRole('searchbox', { name: 'Search workspace', exact: true }).fill('loose_t');
+    await listing;
+    await expect(page.locator('.gallery-body article[data-image-key]')).toHaveCount(1);
+    await expect(card(page, UNLINKED)).toBeVisible();
+});
+
+// Task C4 rulings: a save re-reads the selected file's details without a loading flash, and under a Tags filter or a
+// search the loaded listing is re-read in place (scroll and selection kept while the file still matches).
+const cards = page => page.locator('.gallery-body article[data-image-key]');
+const listingOf = (page, check) => page.waitForResponse(item => item.url().includes('/images/output?') && check(new URL(item.url()).searchParams));
+
+test('the post-save re-read keeps the panel content until the new read lands', async ({ page }) => {
+    await page.goto(images);
+    await expect(card(page, FIRST)).toHaveClass(/\bselected\b/);
+    await expect(tagRow(page).locator('.chip')).toHaveText(manifest.pair_tags);
+    await page.evaluate(() => {
+        const node = document.querySelector('aside[aria-label="Selected image details"]');
+        window.__loadingSeen = false;
+        window.__loadingWatch = new MutationObserver(() => { if (node.textContent.includes('Loading file details')) window.__loadingSeen = true; });
+        window.__loadingWatch.observe(node, { childList: true, subtree: true, characterData: true });
+    });
+    let release;
+    const held = new Promise(resolve => { release = resolve; });
+    const matcher = url => url.pathname === `/prompt_manager/images/inspect/${FIRST}`;
+    await page.route(matcher, async route => { await held; await route.continue(); });
+    await openTag(page, FIRST, { goto: false });
+    await manual(page).fill('held_tag');
+    const reread = page.waitForRequest(item => new URL(item.url()).pathname === `/prompt_manager/images/inspect/${FIRST}`);
+    await dialog(page).getByRole('button', { name: 'Save manual tags', exact: true }).click();
+    await reread;
+    await expect(toast(page)).toHaveText('1 tags saved for dunes_00001.png.');
+    // While the re-read is held, the panel still shows the file's details.
+    await expect(panel(page).locator('.panel-heading h2')).toHaveText(manifest.title);
+    await expect(tagRow(page).locator('.chip')).toHaveText(manifest.pair_tags);
+    expect(await page.evaluate(() => window.__loadingSeen)).toBe(false);
+    const landed = inspected(page, FIRST);
+    release();
+    await landed;
+    await expect(tagRow(page).locator('.chip')).toHaveText(['held_tag', ...manifest.pair_tags]);
+    expect(await page.evaluate(() => { window.__loadingWatch.disconnect(); return window.__loadingSeen; })).toBe(false);
+    await page.unroute(matcher);
+});
+
+test('a save under a Tags filter re-reads the listing: the selection stays while it matches and follows when it leaves', async ({ page, request }) => {
+    await saveImageTags(request, UNLINKED, ['shared']);
+    await saveImageTags(request, SECOND, ['shared']);
+    await page.goto(images);
+    const read = inspected(page, SECOND);
+    await infoButton(page, SECOND).click();
+    await read;
+    const filtered = listingOf(page, query => query.get('tags') === 'shared');
+    await tagRow(page).getByRole('button', { name: 'shared', exact: true }).click();
+    await filtered;
+    await expect(cards(page)).toHaveCount(2);
+    // The newest match (SECOND) is selected once the filter applies.
+    await expect(card(page, SECOND)).toHaveClass(/\bselected\b/);
+    // Both matches checked: the selection bar keeps only what the refreshed result still lists (I7).
+    for (const path of [SECOND, UNLINKED]) await card(page, path).getByRole('checkbox', { name: `Select ${path.split('/').pop()}`, exact: true }).check();
+    const bulkbar = page.locator('#bulkbar .bulkbar');
+    await expect(bulkbar.locator('strong')).toHaveText('2 selected');
+    // Tagging the other match keeps both cards and SECOND selected; the listing is re-read.
+    await openTag(page, UNLINKED, { goto: false });
+    await manual(page).fill('shared, extra');
+    let reread = listingOf(page, query => query.get('tags') === 'shared' && !query.has('cursor'));
+    await dialog(page).getByRole('button', { name: 'Save manual tags', exact: true }).click();
+    await reread;
+    await expect(cards(page)).toHaveCount(2);
+    await expect(card(page, SECOND)).toHaveClass(/\bselected\b/);
+    await expect(tagButton(page, UNLINKED)).toBeFocused();
+    // SECOND loses the filter tag: it leaves the result and the panel follows the remaining match.
+    await openTag(page, SECOND, { goto: false });
+    await manual(page).fill('');
+    reread = listingOf(page, query => query.get('tags') === 'shared' && !query.has('cursor'));
+    const next = inspected(page, UNLINKED);
+    await dialog(page).getByRole('button', { name: 'Save manual tags', exact: true }).click();
+    await reread;
+    await next;
+    await expect(cards(page)).toHaveCount(1);
+    await expect(card(page, SECOND)).toHaveCount(0);
+    await expect(card(page, UNLINKED)).toHaveClass(/\bselected\b/);
+    await expect(panel(page).locator('.panel-heading h2')).toHaveText('unlinked.png');
+    await expect(page.getByText('1 of 1 image', { exact: false }).first()).toBeVisible();
+    await expect(bulkbar.locator('strong')).toHaveText('1 selected');
+    await expect(card(page, UNLINKED).getByRole('checkbox')).toBeChecked();
+    // Export metadata acts on exactly the one file the bar counts.
+    const exported = [];
+    page.on('request', item => { if (item.url().includes('/prompt_manager/images/inspect/')) exported.push(new URL(item.url()).pathname); });
+    const download = page.waitForEvent('download');
+    await bulkbar.getByRole('button', { name: 'Export metadata', exact: true }).click();
+    await download;
+    expect(exported).toEqual([`/prompt_manager/images/inspect/${UNLINKED}`]);
+});
+
+test('a save under a search keeps the scroll position and the loaded range', async ({ page, request }) => {
+    await expect(await request.post('/__test__/tagging/reset', { data: { extra_images: 140 } })).toBeOK();
+    await page.goto(images);
+    const searched = listingOf(page, query => query.get('q') === 'bulk');
+    await page.getByRole('searchbox', { name: 'Search workspace', exact: true }).fill('bulk');
+    await searched;
+    const continuation = page.locator('#gallery-continuation');
+    // Load at least two pages (the search lists 40 a page), then rest a third of the way down.
+    const loadedCount = async () => {
+        const text = await continuation.textContent();
+        return /caught up/.test(text) ? 140 : Number(/^(\d+) of 140/.exec(text)?.[1] ?? 0);
+    };
+    await expect.poll(async () => { await continuation.scrollIntoViewIfNeeded(); return loadedCount(); }).toBeGreaterThanOrEqual(80);
+    await page.evaluate(() => window.scrollTo(0, Math.round(document.documentElement.scrollHeight / 3)));
+    await expect(continuation).not.toBeInViewport();
+    const loaded = await continuation.textContent();
+    await expect.poll(() => continuation.textContent()).toBe(loaded);
+    const y = await page.evaluate(() => window.scrollY);
+    expect(y).toBeGreaterThan(0);
+    const visible = await cards(page).evaluateAll(nodes => nodes.find(node => { const box = node.getBoundingClientRect(); return box.top > 80 && box.bottom < innerHeight; })?.dataset.imageKey);
+    expect(visible).toBeTruthy();
+    const path = visible.slice(2);
+    await openTag(page, path, { goto: false });
+    await manual(page).fill('kept_in_place');
+    const reread = listingOf(page, query => query.get('q') === 'bulk' && !query.has('cursor'));
+    await dialog(page).getByRole('button', { name: 'Save manual tags', exact: true }).click();
+    await reread;
+    await expect(tagButton(page, path)).toBeFocused();
+    await expect(continuation).toHaveText(loaded);
+    expect(Math.abs(await page.evaluate(() => window.scrollY) - y)).toBeLessThanOrEqual(1);
+});
+
+// Final review round (I1, I2, I9, M5, M6).
+test('the option fields show the server defaults that Generate sends (I1)', async ({ page, request }) => {
+    await expect(await request.post('/__test__/tagging/defaults', { data: { general: 0.5, character: 0.9, limit: 7 } })).toBeOK();
+    await openTag(page);
+    const field = name => dialog(page).getByRole('spinbutton', { name, exact: true });
+    await expect(field('General confidence')).toHaveValue('0.5');
+    await expect(field('Character confidence')).toHaveValue('0.9');
+    await expect(field('Suggestions')).toHaveValue('7');
+    const answer = await generated(page);
+    expect(answer.request().postDataJSON().options).toEqual({ general_threshold: 0.5, character_threshold: 0.9, suggestion_limit: 7 });
+    await expect(chips(page)).toHaveCount(7);
+    // A model change re-renders the new model's fields from the server defaults, and keeps the count (D91).
+    await field('Suggestions').fill('4');
+    await model(page).selectOption('wd14-vit');
+    await expect(field('General confidence')).toHaveValue('0.5');
+    await expect(field('Suggestions')).toHaveValue('4');
+});
+
+test('a tag conflict shows the other writer\'s tags before Save, and a second Save keeps them (I2)', async ({ page, request }) => {
+    await openTag(page);
+    await manual(page).fill('mine');
+    // Another writer (a second tab, the legacy page) saves while the dialog is open.
+    expect(await saveImageTags(request, FIRST, ['changed_elsewhere'])).toEqual(['changed_elsewhere']);
+    const conflict = page.waitForResponse(item => item.url().endsWith('/image-annotations/tags') && item.request().method() === 'PUT');
+    await dialog(page).getByRole('button', { name: 'Save manual tags', exact: true }).click();
+    expect((await conflict).status()).toBe(409);
+    await expect(status(page)).toContainText('Saved now: changed_elsewhere.');
+    await expect(manual(page)).toHaveValue('changed_elsewhere, mine');
+    await expect(dialog(page)).toBeVisible();
+    const second = page.waitForResponse(item => item.url().endsWith('/image-annotations/tags') && item.request().method() === 'PUT');
+    await dialog(page).getByRole('button', { name: 'Save manual tags', exact: true }).click();
+    expect((await second).request().postDataJSON().tags).toEqual(['changed_elsewhere', 'mine']);
+    await expect(toast(page)).toHaveText('2 tags saved for dunes_00001.png.');
+    expect(await savedTags(request, FIRST)).toEqual(['changed_elsewhere', 'mine']);
+});
+
+test('D116: Retry leads the actions after an unconfirmed save and resends the same tags', async ({ page, request }) => {
+    await openTag(page);
+    await manual(page).fill('retried');
+    let lost = true;
+    await page.route('**/prompt_manager/image-annotations/tags', route => (lost ? (lost = false, route.fulfill({ status: 504, body: 'Gateway Timeout' })) : route.continue()));
+    await dialog(page).getByRole('button', { name: 'Save manual tags', exact: true }).click();
+    await expect(status(page)).toHaveText('The save could not be confirmed. Retry sends the same tags again.');
+    const inventory = await page.locator('dialog#modal').evaluate(taggingInventory);
+    expect(inventory.actions.buttons.map(button => [button.text, button.primary, button.disabled])).toEqual(
+        [['Retry', false, false], ['Cancel', false, false], ['Save manual tags', true, true]]);
+    expect(inventory.actions.buttons[0].minHeight).toBe(inventory.actions.buttons[1].minHeight);
+    const resent = page.waitForResponse(item => item.url().endsWith('/image-annotations/tags') && item.request().method() === 'PUT');
+    await dialog(page).getByRole('button', { name: 'Retry', exact: true }).click();
+    expect((await resent).request().postDataJSON().tags).toEqual(['retried']);
+    await expect(toast(page)).toHaveText('1 tags saved for dunes_00001.png.');
+    expect(await savedTags(request, FIRST)).toEqual(['retried']);
+    await page.unroute('**/prompt_manager/image-annotations/tags');
+});
+
+test('D116: Retry after an expired capture observes again and keeps the typed draft (M5)', async ({ page, request }) => {
+    await openTag(page);
+    await manual(page).fill('kept_draft');
+    // The store forgets every capture (as its idle sweep would), so Save answers capture-expired.
+    await expect(await request.post('/__test__/tagging/suggestions', { data: { scored: null } })).toBeOK();
+    await dialog(page).getByRole('button', { name: 'Save manual tags', exact: true }).click();
+    await expect(dialog(page).locator('.dialog-actions .ca-primary')).toHaveAttribute('aria-disabled', 'true');
+    const observed = page.waitForResponse(item => item.url().endsWith('/image-annotations/observe'));
+    await dialog(page).getByRole('button', { name: 'Retry', exact: true }).click();
+    expect((await observed).status()).toBe(200);
+    await expect(dialog(page).locator('.dialog-actions button')).toHaveText(['Cancel', 'Save manual tags']);
+    await expect(manual(page)).toHaveValue('kept_draft');
+    await dialog(page).getByRole('button', { name: 'Save manual tags', exact: true }).click();
+    await expect(toast(page)).toHaveText('1 tags saved for dunes_00001.png.');
+    expect(await savedTags(request, FIRST)).toEqual(['kept_draft']);
+});
+
+test('a tag save from Tags reloads a filtered Images visit instead of restoring it (M6)', async ({ page, request }) => {
+    await saveImageTags(request, SECOND, ['shared']);
+    await saveImageTags(request, UNLINKED, ['shared']);
+    await page.goto(images);
+    await infoButton(page, SECOND).click();
+    const filtered = listingOf(page, query => query.get('tags') === 'shared');
+    await tagRow(page).getByRole('button', { name: 'shared', exact: true }).click();
+    await filtered;
+    await expect(cards(page)).toHaveCount(2);
+    await expect(card(page, SECOND)).toHaveClass(/\bselected\b/);
+    // Tags › Auto-tag images tags the selected SECOND without the filter tag.
+    await nav(page, 'Tags & categories').click();
+    await page.locator('#workspace-main .heading-actions').getByRole('button', { name: 'Auto-tag images', exact: true }).click();
+    await expect(manual(page)).toHaveValue('shared');
+    await manual(page).fill('other');
+    await dialog(page).getByRole('button', { name: 'Save manual tags', exact: true }).click();
+    await expect(toast(page)).toHaveText('1 tags saved for dunes_00002.png.');
+    const reloaded = listingOf(page, query => query.get('tags') === 'shared' && !query.has('cursor'));
+    await nav(page, 'Images').click();
+    await reloaded;
+    await expect(cards(page)).toHaveCount(1);
+    await expect(card(page, UNLINKED)).toBeVisible();
+    await expect(page.getByRole('button', { name: 'Filters · 1', exact: true })).toBeVisible();
+});
+
+test('a failed listing re-read after a save says so and keeps the cards (M2)', async ({ page, request }) => {
+    await saveImageTags(request, SECOND, ['shared']);
+    await saveImageTags(request, UNLINKED, ['shared']);
+    await page.goto(images);
+    await infoButton(page, SECOND).click();
+    const filtered = listingOf(page, query => query.get('tags') === 'shared');
+    await tagRow(page).getByRole('button', { name: 'shared', exact: true }).click();
+    await filtered;
+    await expect(cards(page)).toHaveCount(2);
+    await openTag(page, SECOND, { goto: false });
+    await manual(page).fill('');
+    const failing = url => url.pathname === '/prompt_manager/images/output' && url.searchParams.get('tags') === 'shared';
+    await page.route(failing, route => route.fulfill({ status: 500, contentType: 'application/json', body: '{"success":false,"error":"down"}' }));
+    const reread = page.waitForResponse(item => failing(new URL(item.url())));
+    await dialog(page).getByRole('button', { name: 'Save manual tags', exact: true }).click();
+    expect((await reread).status()).toBe(500);
+    await expect(toast(page)).toHaveText('Images could not be refreshed after the tag save. Refresh image collection to retry.');
+    await expect(cards(page)).toHaveCount(2);
+    await expect(page.locator('#gallery-continuation [aria-disabled="true"]')).toHaveCount(0);
+    await page.unroute(failing);
+});

tests/browser/workspace-visual.spec.mjs

@@ -3,7 +3,7 @@ import { readFile } from 'node:fs/promises';
 import { createHash } from 'node:crypto';
 import { captureCanonicalState, historyGeometry, compareCanonicalHistory, nativeDialogState, exerciseCurrentPairSend, assertRecordedHistory } from './history-canonical-comparison.mjs';
 import { startCoverage, saveCoverage } from '../../tools/vendor/browser-coverage.mjs';
-import { test, expect } from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import { test, expect } from './fixtures.mjs';
 
 test.use({ locale: 'en-US', timezoneId: 'UTC', reducedMotion: 'reduce' });
 const workspace = '/prompt_manager/workspace';
@@ -107,7 +107,7 @@ for (const [width, height] of [[1600, 1100], [390, 844]]) for (const theme of ['
         await expect(roles.getByRole('button', { name: 'Send prompt pair', exact: true })).toHaveCount(0);
         await expect(page.locator('#workspace-main > .heading button.send')).toBeEnabled();
         await expect(page.locator('.library-history .history-gallery article[data-image-key]')).toHaveCount(2);
-        await page.locator('.library-history .history-gallery .image-card-actions button').first().click();
+        await page.locator('.library-history .history-gallery .image-card-actions button:first-child').first().click();
         await expect(page.getByRole('complementary', { name: 'Selected image details', exact: true })).toContainText('canonical-alt.png');
         await page.locator('.library-history .history-gallery .media-card.selected .image-open').click();
         await expect(page.locator('dialog#viewer-dialog')).toBeVisible();await page.keyboard.press('Escape');
@@ -296,7 +296,7 @@ async function containedImageContent(page) {
         const failures = [];
         for (const card of nodes) {
             const outer = card.getBoundingClientRect();
-            for (const selector of ['.media-info', '.media-info h3', '.media-info p', '.image-card-actions button']) {
+            for (const selector of ['.media-info', '.media-info h3', '.media-info p', '.image-card-actions']) {
                 const node = card.querySelector(selector), box = node.getBoundingClientRect();
                 if (box.width <= 0 || box.height <= 0 || box.left < outer.left - 1 || box.right > outer.right + 1 ||
                     box.top < outer.top - 1 || box.bottom > outer.bottom + 1 ||
@@ -306,7 +306,7 @@ async function containedImageContent(page) {
                         content: { left: box.left, right: box.right, top: box.top, bottom: box.bottom } });
                 }
             }
-            const action = card.querySelector('.image-card-actions button').getBoundingClientRect();
+            const action = card.querySelector('.image-card-actions').getBoundingClientRect();
             const range = document.createRange(); range.selectNodeContents(card.querySelector('.media-info p'));
             for (const text of range.getClientRects()) {
                 if (Math.min(text.right, action.right) - Math.max(text.left, action.left) > .5 &&
@@ -351,7 +351,7 @@ for (const [width, height] of [[390, 844], [1600, 1100]]) for (const layout of [
             const filename = (await longCard.locator('.image-open').getAttribute('aria-label')).replace(/^Open /, '');
             expect(filename).toMatch(/layered-context-.*\.png/);
             // D19: Info selects into the panel; the image itself opens the viewer.
-            await longCard.locator('.image-card-actions button').focus(); await page.keyboard.press('Enter');
+            await longCard.locator('.image-card-actions button:first-child').focus(); await page.keyboard.press('Enter');
             await expect(page.locator('.detail-panel .panel-heading h2')).toHaveText(filename);
             // D19: the card image, not the panel, opens the viewer.
             const fullPreview = longCard.locator('.image-open');
@@ -420,7 +420,7 @@ test('supplemental extreme zoom grows real footers preserves the visible anchor
     await page.evaluate(() => { document.documentElement.style.zoom = '2'; });
     await expect.poll(() => cards.nth(1).evaluate(node => node.offsetLeft)).toBe(0);
     const selected = cards.nth(16), key = await selected.getAttribute('data-image-key');
-    await selected.locator('.image-card-actions button').click(); await selected.locator('.image-open').focus();
+    await selected.locator('.image-card-actions button:first-child').click(); await selected.locator('.image-open').focus();
     const filename = (await selected.locator('.image-open').getAttribute('aria-label')).replace(/^Open /, '');
     await selected.evaluate(node => scrollBy(0, node.getBoundingClientRect().top - 120));
     await page.evaluate(() => {
@@ -497,7 +497,7 @@ test('supplemental wrapped real footer reservation survives append without movin
     await page.setViewportSize({ width: 390, height: 844 }); await page.goto(`${workspace}#/images`);
     const cards = page.locator('article[data-image-key]'), grid = page.locator('.image-grid');
     await expect(cards).toHaveCount(40);
-    await cards.first().locator('.image-card-actions button').click();
+    await cards.first().locator('.image-card-actions button:first-child').click();
     const selectedKey = await cards.first().getAttribute('data-image-key');
     await page.evaluate(() => { document.documentElement.style.zoom = '4'; });
     await expect.poll(() => grid.evaluate(node => Number.parseFloat(node.style.getPropertyValue('--image-footer-height')))).toBeGreaterThan(54);
@@ -609,7 +609,7 @@ test('Images visual layouts retain real delayed media long filenames and loaded
         await page.setViewportSize({width,height});
         for (const theme of ['dark','light']) {
             await page.getByRole('combobox',{name:'Theme',exact:true}).selectOption(theme);
-            await page.locator('.media-card .image-card-actions button').first().click();
+            await page.locator('.media-card .image-card-actions button:first-child').first().click();
             await expect(page.locator('.detail-panel .panel-heading h2')).toHaveText('study-01.png');
             for (const layout of ['Grid','List','Masonry']) {
                 await page.getByRole('button',{name:layout,exact:true}).click();await page.evaluate(()=>scrollTo(0,0));
@@ -621,7 +621,7 @@ test('Images visual layouts retain real delayed media long filenames and loaded
             const longCard=page.locator('article[data-image-key]').filter({has:page.locator('[aria-label^="Open a-very-long-geometric-composition-"]')});
             const longName=(await longCard.locator('.image-open').getAttribute('aria-label')).replace(/^Open /,'');
             expect(longName).toMatch(/layered-context-.*\.png/);
-            await longCard.locator('.image-card-actions button').click();
+            await longCard.locator('.image-card-actions button:first-child').click();
             await expect(page.locator('.detail-panel .panel-heading h2')).toHaveText(longName);
             await orderedMedia(page);
             await attachEvidence(page,testInfo,`after-images-loaded-scroll-${width}-${theme}`,false);

tests/browser/workspace-vocabulary.spec.mjs

@@ -1,6 +1,6 @@
 import {startCoverage,saveCoverage} from '../../tools/vendor/browser-coverage.mjs';
 import {writeFile} from 'node:fs/promises';
-import {test,expect} from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import {test,expect} from './fixtures.mjs';
 import {captureCanonicalState} from './history-canonical-comparison.mjs';
 import {pageInventoryOf,inventoryOf,diffInventory,mapVocabularyConcept,mapVocabularyDialog} from './vocabulary-canonical-comparison.mjs';
 
@@ -50,9 +50,12 @@ test('tags route renders the concept composition without a legacy handoff',async
     await expect(main(page).getByRole('heading',{level:1})).toHaveText('Tags & categories');
     await expect(main(page).locator(':scope > .heading p.muted')).toHaveText('Organize once. Reuse everywhere.');
     const actions=main(page).locator(':scope > .heading > .heading-actions button');
-    await expect(actions).toHaveText(['+ Create tag']);
+    // The concept's heading actions (#249 C2 retired #245 D2): Auto-tag images is primary; with no image selected in
+    // Images it is aria-disabled (D115).
+    await expect(actions).toHaveText(['+ Create tag','Auto-tag images']);
     await expect(actions.first()).not.toHaveClass(/ca-primary/);
-    await expect(main(page).getByRole('button',{name:'Auto-tag images'})).toHaveCount(0);
+    await expect(actions.last()).toHaveClass(/ca-primary/);
+    await expect(actions.last()).toHaveAttribute('aria-disabled','true');
     await expect(surfaces(page)).toHaveCount(2);
     await expect(surfaces(page).locator('h2')).toHaveText(['Tag vocabulary','Categories']);
     await expect(surfaces(page).locator(':scope > p:visible')).toHaveText(['Category/tag autocomplete uses this same vocabulary across nodes and dialogs.',
@@ -65,9 +68,9 @@ test('tags route renders the concept composition without a legacy handoff',async
     expect(manifest.tags).toEqual(expected.map(([name])=>name));
     await expect(rows(page).locator('td:nth-child(1)')).toHaveText(expected.map(([name])=>name));
     await expect(rows(page).locator('td:nth-child(2)')).toHaveText(expected.map(([,count])=>String(count)));
-    const images=await rows(page).locator('td:nth-child(3)').evaluateAll(cells=>cells.map(cell=>[cell.querySelector('[aria-hidden="true"]')?.textContent,cell.querySelector('.ca-visually-hidden')?.textContent]));
-    expect(images).toEqual(expected.map(()=>['—','Not available yet']));
-    await expect(rows(page).first().locator('td:nth-child(3)')).toHaveAccessibleName('Not available yet');
+    // Images counts the files carrying each tag (#249 C2 retired #245 D3); this fixture tags no image.
+    await expect(rows(page).locator('td:nth-child(3)')).toHaveText(expected.map(()=>'0'));
+    await expect(rows(page).locator('td:nth-child(3) .ca-visually-hidden')).toHaveCount(0);
     const labels=await rows(page).evaluateAll(nodes=>nodes.map(row=>[...row.querySelectorAll('td:nth-child(4) .row-actions button')].map(button=>button.textContent)));
     expect(labels).toEqual(expected.map(()=>['Browse','Rename / merge']));
     const operations=surfaces(page).nth(1).locator('.operation');

tests/browser/workspace.spec.mjs

@@ -1,5 +1,5 @@
 import { startCoverage, saveCoverage } from '../../tools/vendor/browser-coverage.mjs';
-import { test, expect } from '../../tools/vendor/node_modules/@playwright/test/index.mjs';
+import { test, expect } from './fixtures.mjs';
 
 const workspace = '/prompt_manager/workspace';
 const statistic = (page, label) => page.locator('#workspace-main dt')

tests/browser/workspace_concept_fixture.py

@@ -158,6 +158,9 @@ MAX_EXTRA_IMAGES = 120
 # Unlinked, seedless images (D11 "W × H · PNG") newer than the concept's, for D3 checks.
 MAX_SEEDLESS_IMAGES = 2
 SEEDLESS_SIZE = (1536, 1152)
+# The same 4:3 shape with a short "96 × 72 · PNG" subtitle that stays on one line at
+# 390px beside the card's Info and Tag (the D3 footer check).
+SHORT_SEEDLESS_SIZE = (96, 72)
 SESSION = "5c1f9a7e-3f0b-4a55-9d0e-2b8c4f6a7d10"
 _TABLES = (
     "execution_run_outputs",
@@ -170,6 +173,9 @@ _TABLES = (
     "execution_context_texts",
     "saved_views",
     "image_favorites",
+    "media_image_tags",
+    "media_annotation_bytes",
+    "media_annotation_files",
     "image_prompt_pairs",
     "prompt_pair_tags",
     "prompt_pairs",
@@ -323,9 +329,9 @@ def _record_run(
     return start.run_id
 
 
-def _write_seedless_png(path: Path) -> None:
+def _write_seedless_png(path: Path, size: tuple[int, int] = SEEDLESS_SIZE) -> None:
     """A plain PNG with no metadata: no prompt, no seed, no link."""
-    with Image.new("RGB", SEEDLESS_SIZE, "#2b3442") as image:
+    with Image.new("RGB", size, "#2b3442") as image:
         image.save(path)
 
 
@@ -342,13 +348,17 @@ def seed_workspace_concept(
     *,
     extra_images: int = 0,
     seedless_images: int = 0,
+    short_seedless: bool = False,
 ) -> dict:
     """Replace this server's disposable rows and files with the concept's data.
 
     ``extra_images`` continues the concept's rounds (same prompts, dimensions and
     seed rule) after its 24 images, for bounded-rendering and Load more checks.
-    ``seedless_images`` adds unlinked PNGs without metadata, newest of all.
+    ``seedless_images`` adds unlinked PNGs without metadata, newest of all;
+    ``short_seedless`` makes them ``SHORT_SEEDLESS_SIZE`` instead.
     """
+    if not isinstance(short_seedless, bool):
+        raise TypeError("short_seedless must be a boolean")
     _bounded("extra_images", extra_images, MAX_EXTRA_IMAGES)
     _bounded("seedless_images", seedless_images, MAX_SEEDLESS_IMAGES)
     _clear(api, root)
@@ -456,7 +466,9 @@ def seed_workspace_concept(
     for number in range(seedless_images):
         name = f"seedless_{number + 1:05}.png"
         path = folder / name
-        _write_seedless_png(path)
+        _write_seedless_png(
+            path, SHORT_SEEDLESS_SIZE if short_seedless else SEEDLESS_SIZE
+        )
         stamp = now - 10 - number * 10
         os.utime(path, (stamp, stamp))
         seedless.append({"filename": name, "path": f"concepts/{name}"})

tests/browser/workspace_library_editor_fixture.py

@@ -28,6 +28,9 @@ EDITOR_METADATA = {
 _TABLES = (
     "saved_views",
     "image_favorites",
+    "media_image_tags",
+    "media_annotation_bytes",
+    "media_annotation_files",
     "image_prompt_pairs",
     "prompt_pair_tags",
     "prompt_pairs",

tests/browser/workspace_tagging_fixture.py

@@ -0,0 +1,313 @@
+"""Owned, deterministic files for the Tag image dialog workflows (#249 PR C).
+
+Two files share one positive prompt and one saved pair, an equal-byte copy of
+the first sits at another path, and a video and an unlinked file follow. The
+injected fake model (``tests/browser/server.py`` ``TestModel``) answers
+Generate; no model file exists and nothing is ever downloaded.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+from pathlib import Path
+import shutil
+import time
+from typing import TYPE_CHECKING
+
+from aiohttp import web
+from PIL import Image
+from PIL.PngImagePlugin import PngInfo
+
+if TYPE_CHECKING:
+    from py.api import PromptManagerAPI
+
+POSITIVE = (
+    "Monolithic sandstone pavilion in a vast desert, sweeping dunes, warm "
+    "terracotta, late afternoon sun, architectural photography, quiet geometry"
+)
+NEGATIVE = "plastic, harsh shadows, noise, watermark"
+TITLE = "Quiet geometry"
+PROMPT_TAGS = ["minimal", "warm_light"]
+PAIR_TAGS = ["editorial"]
+# Newest first, so the Images order (Recently added) is the manifest's.
+FILES = (
+    ("tagging/dunes_00001.png", "linked"),
+    ("tagging/dunes_00002.png", "linked"),
+    ("copies/dunes_00001.png", "copy"),
+    ("tagging/clip.webm", "video"),
+    ("loose/unlinked.png", "unlinked"),
+)
+MODELS = ("wd14-swinv2", "wd14-vit", "gguf", "hf")
+# Unlinked copies enough to virtualize the Images grid (it renders all up to 120).
+MAX_EXTRA_IMAGES = 160
+_TABLES = (
+    "saved_views",
+    "image_favorites",
+    "media_image_tags",
+    "media_annotation_bytes",
+    "media_annotation_files",
+    "image_prompt_pairs",
+    "prompt_pair_tags",
+    "prompt_pairs",
+    "generated_images",
+    "prompt_tags",
+    "tags",
+    "prompt_library_state",
+    "prompts",
+)
+# Every prompt-side owner row a Tag save could touch, for before/after
+# comparisons: tagging one image never changes a prompt, a pair or a link.
+_SNAPSHOT = {
+    "prompts": "SELECT id, text, category, tags, updated_at FROM prompts ORDER BY id",
+    "prompt_tags": "SELECT prompt_id, tag_id FROM prompt_tags ORDER BY 1, 2",
+    "prompt_pairs": "SELECT id, positive_prompt_id, negative_prompt_id, category,"
+    " archived, updated_at FROM prompt_pairs ORDER BY id",
+    "prompt_pair_tags": "SELECT pair_id, tag_id FROM prompt_pair_tags ORDER BY 1, 2",
+    "generated_images": "SELECT id, prompt_id, image_path FROM generated_images"
+    " ORDER BY id",
+    "image_prompt_pairs": "SELECT id, image_id, pair_id FROM image_prompt_pairs"
+    " ORDER BY id",
+}
+
+
+def _png(path: Path, colour: tuple[int, int, int], size=(96, 64)) -> None:
+    graph = {
+        "1": {"class_type": "CLIPTextEncode", "inputs": {"text": POSITIVE}},
+        "2": {"class_type": "CLIPTextEncode", "inputs": {"text": NEGATIVE}},
+        "3": {
+            "class_type": "KSampler",
+            "inputs": {"positive": ["1", 0], "negative": ["2", 0], "seed": 7},
+        },
+        "4": {"class_type": "SaveImage", "inputs": {"images": ["3", 0]}},
+    }
+    info = PngInfo()
+    info.add_text("prompt", json.dumps(graph))
+    path.parent.mkdir(parents=True, exist_ok=True)
+    with Image.new("RGB", size, colour) as image:
+        image.save(path, pnginfo=info)
+
+
+def seed_tagging(api: PromptManagerAPI, root: Path, extra_images: int = 0) -> dict:
+    """Replace this server's disposable rows and files with the tagging set."""
+    db = api.db
+    with db.model.get_connection() as conn:
+        for table in _TABLES:
+            conn.execute(f"DELETE FROM {table}")
+    for name in ("output", "input"):
+        shutil.rmtree(root / name)
+        (root / name).mkdir()
+    output = root / "output"
+    _png(output / FILES[0][0], (196, 140, 92))
+    _png(output / FILES[1][0], (150, 110, 80))
+    (output / FILES[2][0]).parent.mkdir(parents=True, exist_ok=True)
+    shutil.copyfile(output / FILES[0][0], output / FILES[2][0])
+    shutil.copyfile(
+        Path(__file__).resolve().parents[1] / "fixtures" / "tiny.webm",
+        output / FILES[3][0],
+    )
+    _png(output / FILES[4][0], (60, 70, 90), size=(64, 64))
+    now = time.time()
+    for index, (relative, _) in enumerate(FILES):
+        stamp = now - 60 - index * 10
+        os.utime(output / relative, (stamp, stamp))
+    for number in range(extra_images):
+        extra = output / "bulk" / f"bulk_{number:04}.png"
+        extra.parent.mkdir(exist_ok=True)
+        shutil.copyfile(output / FILES[4][0], extra)
+        stamp = now - 600 - number
+        os.utime(extra, (stamp, stamp))
+
+    receipt = db.create_library_recipe(
+        POSITIVE,
+        NEGATIVE,
+        expected_epoch=db.get_execution_epoch(),
+        parent_metadata={
+            "title": TITLE,
+            "category": "Architecture",
+            "tags": list(PROMPT_TAGS),
+        },
+    )
+    positive_id = receipt["positive"]["id"]
+    pair_id = receipt["pair"]["id"]
+    with db.model.get_connection() as conn:
+        tag_ids = []
+        for name in PAIR_TAGS:
+            conn.execute("INSERT OR IGNORE INTO tags (name) VALUES (?)", (name,))
+            tag_ids.append(
+                conn.execute("SELECT id FROM tags WHERE name=?", (name,)).fetchone()[0]
+            )
+        conn.executemany(
+            "INSERT INTO prompt_pair_tags (pair_id, tag_id) VALUES (?, ?)",
+            [(pair_id, tag_id) for tag_id in tag_ids],
+        )
+    for relative, kind in FILES:
+        if kind in {"linked", "video"}:
+            image_id = db.link_image_to_prompt(positive_id, str(output / relative))
+            if kind == "linked":
+                db.link_image_to_pair(pair_id, image_id, {"source": "fixture"})
+    api._generate_gallery_thumbnails_sync(api._get_all_output_dirs(), (150, 150))
+    api.invalidate_gallery_cache()
+    return {
+        "positive_id": positive_id,
+        "pair_id": pair_id,
+        "title": TITLE,
+        "prompt_tags": list(PROMPT_TAGS),
+        "pair_tags": list(PAIR_TAGS),
+        "files": [
+            {"path": relative, "filename": Path(relative).name, "kind": kind}
+            for relative, kind in FILES
+        ],
+    }
+
+
+def tagging_snapshot(api: PromptManagerAPI) -> dict:
+    """Plain rows of every prompt, pair and link table."""
+    with api.db.model.get_connection() as conn:
+        return {
+            name: [list(row) for row in conn.execute(query)]
+            for name, query in _SNAPSHOT.items()
+        }
+
+
+def register_tagging_commands(
+    routes: web.RouteTableDef, api: PromptManagerAPI, root: Path, model
+) -> None:
+    """Register fixed seed, snapshot and fake-model commands; no paths or SQL input."""
+
+    @routes.post("/__test__/tagging/reset")
+    async def reset(request: web.Request) -> web.Response:
+        data = await request.json() if request.can_read_body else {}
+        if request.query or not isinstance(data, dict) or set(data) - {"extra_images"}:
+            raise web.HTTPBadRequest()
+        extra = data.get("extra_images", 0)
+        if isinstance(extra, bool) or not isinstance(extra, int):
+            raise web.HTTPBadRequest()
+        if not 0 <= extra <= MAX_EXTRA_IMAGES:
+            raise web.HTTPBadRequest()
+        model.unavailable = set()
+        model.suggestions = None
+        model.loaded = None
+        _set_defaults(model, DEFAULTS)
+        # A test that ends with the dialog open leaves its capture behind; drop
+        # every capture so the next test starts below the store's limits.
+        await api._run_in_executor(api.close_annotation_captures)
+        manifest = await api._run_in_executor(seed_tagging, api, root, extra)
+        return web.json_response({"success": True, **manifest})
+
+    @routes.get("/__test__/tagging/snapshot")
+    async def snapshot(request: web.Request) -> web.Response:
+        if request.query:
+            raise web.HTTPBadRequest()
+        return web.json_response(await api._run_in_executor(tagging_snapshot, api))
+
+    @routes.post("/__test__/tagging/models")
+    async def models(request: web.Request) -> web.Response:
+        """Mark fake models as not downloaded (``{"unavailable": [...]}``)."""
+        data = await request.json() if request.can_read_body else None
+        if (
+            request.query
+            or not isinstance(data, dict)
+            or set(data) != {"unavailable"}
+            or not isinstance(data["unavailable"], list)
+            or any(name not in MODELS for name in data["unavailable"])
+        ):
+            raise web.HTTPBadRequest()
+        model.unavailable = set(data["unavailable"])
+        return web.json_response({"success": True})
+
+    @routes.post("/__test__/tagging/suggestions")
+    async def suggestions(request: web.Request) -> web.Response:
+        """Answer Generate with fixed scored names (the concept's own sample).
+
+        ``{"scored": [[name, confidence], ...]}`` (at most 100, confidence in
+        [0, 1]) or ``{"scored": null}`` for the default list. Every capture is
+        released too, so a concept comparison starts below the store's limits
+        whatever the previous test left open.
+        """
+        data = await request.json() if request.can_read_body else None
+        scored = data.get("scored") if isinstance(data, dict) else None
+        if (
+            request.query
+            or not isinstance(data, dict)
+            or set(data) != {"scored"}
+            or not (scored is None or _valid_scored(scored))
+        ):
+            raise web.HTTPBadRequest()
+        model.suggestions = (
+            None if scored is None else [(name, float(value)) for name, value in scored]
+        )
+        model.unavailable = set()
+        await api._run_in_executor(api.close_annotation_captures)
+        return web.json_response({"success": True})
+
+    @routes.post("/__test__/tagging/defaults")
+    async def defaults(request: web.Request) -> web.Response:
+        """Serve other tagging defaults from ``GET /autotag/models``.
+
+        ``{"general": g, "character": c, "limit": n}``: the WD14 thresholds
+        (in [0, 1]) and the persisted suggestion count (1..100), so a test can
+        tell the server's values from the dialog's built-in ones. Reset
+        restores the standard 0.35 / 0.85 / 10.
+        """
+        data = await request.json() if request.can_read_body else None
+        if request.query or not _valid_defaults(data):
+            raise web.HTTPBadRequest()
+        _set_defaults(model, (data["general"], data["character"], data["limit"]))
+        return web.json_response({"success": True})
+
+    @routes.get("/__test__/tagging/captures")
+    async def captures(request: web.Request) -> web.Response:
+        """How many captures the store holds (a released capture is gone)."""
+        if request.query:
+            raise web.HTTPBadRequest()
+        store = api._captures
+        count = store.stats()["captures"] if store is not None else 0
+        return web.json_response({"captures": count})
+
+
+# The standard tagging defaults: WD14 general and character thresholds, and
+# the suggestion count (``AutotagConfig``).
+DEFAULTS = (0.35, 0.85, 10)
+
+
+def _set_defaults(model, values) -> None:
+    from py.config import AutotagConfig
+
+    general, character, limit = values
+    model.wd14_general_threshold = general
+    model.wd14_character_threshold = character
+    AutotagConfig.SUGGESTION_LIMIT = limit
+
+
+def _valid_defaults(data: object) -> bool:
+    unit = (
+        lambda value: isinstance(value, (int, float))
+        and not isinstance(value, bool)
+        and 0 <= value <= 1
+    )
+    return (
+        isinstance(data, dict)
+        and set(data) == {"general", "character", "limit"}
+        and unit(data["general"])
+        and unit(data["character"])
+        and type(data["limit"]) is int
+        and 1 <= data["limit"] <= 100
+    )
+
+
+def _valid_scored(scored: object) -> bool:
+    return (
+        isinstance(scored, list)
+        and len(scored) <= 100
+        and all(
+            isinstance(row, list)
+            and len(row) == 2
+            and isinstance(row[0], str)
+            and 0 < len(row[0]) <= 50
+            and isinstance(row[1], (int, float))
+            and not isinstance(row[1], bool)
+            and 0 <= row[1] <= 1
+            for row in scored
+        )
+    )

tests/browser/workspace_visual_fixture.py

@@ -114,6 +114,9 @@ def seed_workspace_visual(api: PromptManagerAPI, root: Path) -> None:
         for table in (
             "saved_views",
             "image_favorites",
+            "media_image_tags",
+            "media_annotation_bytes",
+            "media_annotation_files",
             "image_prompt_pairs",
             "prompt_pair_tags",
             "prompt_pairs",

tests/browser/workspace_vocabulary_fixture.py

@@ -81,6 +81,9 @@ MAX_EXTRA_CATEGORIES = 200
 _TABLES = (
     "saved_views",
     "image_favorites",
+    "media_image_tags",
+    "media_annotation_bytes",
+    "media_annotation_files",
     "image_prompt_pairs",
     "prompt_pair_tags",
     "prompt_pairs",

tests/fixtures/README.md

@@ -13,6 +13,75 @@ ffmpeg -f lavfi -i testsrc=size=32x32:rate=4:duration=1 \
 
 Browser tests use the committed file; FFmpeg is not required to run them.
 
+# Local Inspect header fixtures
+
+`local-inspect/tiny.jpg` (48x32) and `local-inspect/tiny.webp` (40x24) are original
+synthetic FFmpeg color test patterns, under 4 KB, dedicated to the public domain
+under CC0-1.0. `tests/js/png-local-reader.test.js` reads their dimensions with the
+browser header reader and `tests/test_local_inspect_parity.py` checks the same two
+files against Pillow.
+
+Reproduce with FFmpeg:
+
+```sh
+ffmpeg -f lavfi -i testsrc=size=48x32:rate=1 -frames:v 1 -q:v 10 tiny.jpg
+ffmpeg -f lavfi -i testsrc=size=40x24:rate=1 -frames:v 1 -c:v libwebp -quality 40 tiny.webp
+```
+
+# Local Inspect video fixtures
+
+`local-inspect/core-faststart.mp4`, `core-moov-end.mp4`, `core.mov`, `core.m4v`,
+`core.mkv`, `core-tags.webm`, `legacy-comment.mp4` and `legacy-comment.webm` are
+0.2-second 16x16 black clips (each under 4 KB) carrying an original synthetic
+ComfyUI graph in their container tags, dedicated to the public domain under
+CC0-1.0. The `core-*` files carry one tag per key as core SaveVideo/SaveWEBM write
+them (MP4/MOV/M4V through `-movflags use_metadata_tags`; Matroska upper-cases the
+names); `core-faststart.mp4` also carries `parameters` and
+`comfyassets_prompt_pairs`. The `legacy-comment.*` files carry the old
+VideoHelperSuite `comment` JSON envelope (an ilst `©cmt` atom in MP4, a `COMMENT`
+tag in WebM). `video-tags-v1.json` records `ffprobe -show_entries format_tags -of
+json` for each file (ffmpeg 9.0.2); `tests/test_utils_video_metadata.py`
+`VideoFixtureFfprobeTests` re-checks it when ffprobe is installed, and
+`tests/js/workspace-local-video.test.js` plus `tests/test_local_inspect_parity.py`
+hold the browser video reader to the same maps.
+
+Reproduce with FFmpeg, from `tests/fixtures/local-inspect`:
+
+```sh
+PROMPT='{"3":{"class_type":"KSampler","inputs":{"seed":7,"steps":20,"cfg":8.0,"sampler_name":"euler","scheduler":"normal","denoise":1.0,"model":["4",0],"positive":["6",0],"negative":["7",0],"latent_image":["5",0]}},"4":{"class_type":"CheckpointLoaderSimple","inputs":{"ckpt_name":"synthetic-v1.safetensors"}},"5":{"class_type":"EmptyLatentImage","inputs":{"width":512,"height":512,"batch_size":1}},"6":{"class_type":"CLIPTextEncode","inputs":{"text":"a red fox in snow, detailed fur","clip":["4",1]}},"7":{"class_type":"CLIPTextEncode","inputs":{"text":"blurry, lowres","clip":["4",1]}},"8":{"class_type":"VAEDecode","inputs":{"samples":["3",0],"vae":["4",2]}},"9":{"class_type":"SaveVideo","inputs":{"filename_prefix":"video/synthetic","video":["8",0]}}}'
+WORKFLOW='{"last_node_id":9,"nodes":[{"id":3,"type":"KSampler","widgets_values":[7,"fixed",20,8.0,"euler","normal",1.0]},{"id":6,"type":"CLIPTextEncode","widgets_values":["a red fox in snow, detailed fur"]},{"id":7,"type":"CLIPTextEncode","widgets_values":["blurry, lowres"]}],"links":[],"version":0.4}'
+PARAMETERS=$'a red fox in snow, detailed fur\nNegative prompt: blurry, lowres\nSteps: 20, Sampler: Euler, CFG scale: 8.0, Seed: 7, Size: 512x512'
+PAIRS='{"version":1,"nodes":{"6":{"records":[{"positive_text":"a red fox in snow, detailed fur","negative_text":"blurry, lowres","negative_effective_text":"blurry, lowres","prepend_text":"","append_text":""}]}}}'
+COMMENT=$(PROMPT="$PROMPT" WORKFLOW="$WORKFLOW" python3 -c 'import json,os; print(json.dumps({"prompt": os.environ["PROMPT"], "workflow": json.loads(os.environ["WORKFLOW"])}))')
+SRC=(-v quiet -y -f lavfi -i "color=c=black:s=16x16:d=0.2")
+ffmpeg "${SRC[@]}" -movflags use_metadata_tags+faststart -metadata prompt="$PROMPT" -metadata workflow="$WORKFLOW" -metadata parameters="$PARAMETERS" -metadata comfyassets_prompt_pairs="$PAIRS" core-faststart.mp4
+ffmpeg "${SRC[@]}" -movflags use_metadata_tags -metadata prompt="$PROMPT" -metadata workflow="$WORKFLOW" core-moov-end.mp4
+ffmpeg "${SRC[@]}" -movflags use_metadata_tags -metadata prompt="$PROMPT" -metadata workflow="$WORKFLOW" core.mov
+ffmpeg "${SRC[@]}" -movflags use_metadata_tags -metadata prompt="$PROMPT" -metadata workflow="$WORKFLOW" core.m4v
+ffmpeg "${SRC[@]}" -metadata prompt="$PROMPT" -metadata workflow="$WORKFLOW" core.mkv
+ffmpeg "${SRC[@]}" -c:v libvpx -metadata prompt="$PROMPT" -metadata workflow="$WORKFLOW" core-tags.webm
+ffmpeg "${SRC[@]}" -metadata comment="$COMMENT" legacy-comment.mp4
+ffmpeg "${SRC[@]}" -c:v libvpx -metadata comment="$COMMENT" legacy-comment.webm
+for f in core-faststart.mp4 core-moov-end.mp4 core.mov core.m4v core.mkv core-tags.webm legacy-comment.mp4 legacy-comment.webm; do
+  ffprobe -v quiet -show_entries format_tags -of json "$f"   # -> video-tags-v1.json "fixtures"[f] = .format.tags
+done
+```
+
+# Local Inspect projection cases
+
+`local-inspect/projection-cases.json` holds synthetic record sets (prompt, workflow,
+parameters and pair-record texts, optional `record_formats`, a video `carrier_text`
+and `limited` records) that exercise every branch of the server's inspection
+projection: Python's Decimal spellings and exact-construction limits, the 4300-digit
+integer limit, duplicate keys, NaN/Infinity, the depth-64 and node ceilings, prompt
+node key filtering and the carrier rule. `records: null` stands for an unreadable
+file. `expect` pins both graph statuses; `needs` names an interpreter feature a case
+depends on. `tests/js/workspace-local-projection.test.js` runs them through
+`web/js/workspace/local-projection.mjs`, and `tests/test_local_inspect_parity.py`
+asserts that its answer equals `utils.inspection_response._metadata` byte for byte.
+All texts are original synthetic content, dedicated to the public domain under
+CC0-1.0.
+
 # Inspection parity fixtures
 
 `inspection_parity/` holds small synthetic PNGs (`f01b.png` ... `f24.png`, each

tests/fixtures/local-inspect/projection-cases.json

@@ -0,0 +1,849 @@
+{
+ "version": 1,
+ "about": "Synthetic record sets for utils/inspection_response.py _metadata parity with web/js/workspace/local-projection.mjs projectMetadata. records null = an unreadable file. record_formats default to original-text for each record. expect pins both graph statuses; needs names an interpreter feature the Python side must have.",
+ "cases": [
+  {
+   "name": "unreadable_read",
+   "records": null,
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "absent_records",
+   "records": {},
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "prompt_and_workflow_parsed",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"x\":1}}}",
+    "workflow": "{\"nodes\":[],\"version\":0.4}"
+   },
+   "expect": {
+    "prompt_graph": "parsed",
+    "workflow_graph": "parsed"
+   }
+  },
+  {
+   "name": "all_four_records",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"x\":1}}}",
+    "workflow": "{\"nodes\":[],\"version\":0.4}",
+    "parameters": "cat\nNegative prompt: dog\nSteps: 20, Sampler: Euler, Seed: 7",
+    "comfyassets_prompt_pairs": "{\"version\":1,\"nodes\":{}}"
+   },
+   "expect": {
+    "prompt_graph": "parsed",
+    "workflow_graph": "parsed"
+   }
+  },
+  {
+   "name": "parameters_only",
+   "records": {
+    "parameters": "cat\nSteps: 20"
+   },
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "number_spellings",
+   "records": {
+    "workflow": "{\"nodes\":[],\"a\":[1e3,1e0,-0.0,6.50,0.0100,1.0e2,0e5,0.0000001,0.000001,1E+3,1e-0,0.5e1,0.05e1,0.0e0,1.5e+300,123e-8,2.5e-3,1.0,-1.50,0.0,100.0,1234567e-6,0.5E-6,0.5E-5,1e22,0.000000,-0.000000e-3,7.0e-6,7.0e-7,12345678901234567890.123,9007199254740992.0]}"
+   },
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "parsed"
+   }
+  },
+  {
+   "name": "integer_spellings",
+   "records": {
+    "workflow": "{\"a\":[9007199254740991,-9007199254740991,9007199254740992,-9007199254740992,10000000000000000,123456789012345678901234567890,0,-0,42,-7,100]}"
+   },
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "parsed"
+   }
+  },
+  {
+   "name": "nonfinite_literals_are_null",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"n\":NaN,\"i\":Infinity,\"m\":-Infinity},\"_meta\":{\"n\":NaN}}}",
+    "workflow": "{\"a\":[NaN,Infinity,-Infinity],\"b\":NaN}"
+   },
+   "expect": {
+    "prompt_graph": "parsed",
+    "workflow_graph": "parsed"
+   }
+  },
+  {
+   "name": "exponent_edges_valid",
+   "records": {
+    "workflow": "{\"a\":[1e999999999999999999,1e-1999999999999999997,0.0e1000000000000000000,1.5e999999999999999999,0.1e-1999999999999999996]}"
+   },
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "parsed"
+   }
+  },
+  {
+   "name": "exponent_edge_invalid_1e1000000000000000000",
+   "records": {
+    "workflow": "{\"a\":1e1000000000000000000}"
+   },
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "exponent_edge_invalid_15e999999999999999999",
+   "records": {
+    "workflow": "{\"a\":15e999999999999999999}"
+   },
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "exponent_edge_invalid_1e_1999999999999999998",
+   "records": {
+    "workflow": "{\"a\":1e-1999999999999999998}"
+   },
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "exponent_edge_invalid_1_5e_1999999999999999997",
+   "records": {
+    "workflow": "{\"a\":1.5e-1999999999999999997}"
+   },
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "exponent_edge_invalid_0e1000000000000000000",
+   "records": {
+    "workflow": "{\"a\":0e1000000000000000000}"
+   },
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "exponent_edge_invalid_1e9223372036854775808",
+   "records": {
+    "workflow": "{\"a\":1e9223372036854775808}"
+   },
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "exponent_edge_invalid_0_0e1000000000000000001",
+   "records": {
+    "workflow": "{\"a\":0.0e1000000000000000001}"
+   },
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "exponent_edge_invalid_1e99999999999999999999999999",
+   "records": {
+    "workflow": "{\"a\":1e99999999999999999999999999}"
+   },
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "int_4300_digits_kept",
+   "records": {
+    "workflow": "{\"a\":7777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777}"
+   },
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "parsed"
+   }
+  },
+  {
+   "name": "int_4301_digits_malformed",
+   "records": {
+    "workflow": "{\"a\":77777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777}"
+   },
+   "needs": "int-max-str-digits",
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "int_negative_4301_digits_malformed",
+   "records": {
+    "workflow": "{\"a\":-77777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777777}"
+   },
+   "needs": "int-max-str-digits",
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "string_escapes",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"t\":\"é😀\\u0000\\u001f/\\\\\\\"\\b\\f\\n\\r\\t\",\"raw\":\"é😀 — 
\"}}}",
+    "workflow": "{\"s\":\"\\u00E9\\uD83D\\uDE00\\uD800\\uDC00x\\u0000\\u001f\\/\\\\\\\"\\b\\f\\n\\r\\t\\uD800\\u0041\\uDC00\"}"
+   },
+   "expect": {
+    "prompt_graph": "parsed",
+    "workflow_graph": "parsed"
+   }
+  },
+  {
+   "name": "lone_surrogate_in_raw_text",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"t\":\"𐏿\\udc00\"}}}",
+    "parameters": "\ud800 lone\nSteps: 1"
+   },
+   "expect": {
+    "prompt_graph": "parsed",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "keys_with_escapes_proto_and_order",
+   "records": {
+    "prompt": "{\"10\":{\"class_type\":\"B\",\"inputs\":{}},\"3\":{\"class_type\":\"A\",\"inputs\":{}},\"__proto__\":{\"class_type\":\"P\",\"inputs\":{\"constructor\":1}},\"constructor\":{\"class_type\":\"C\",\"inputs\":{}},\"\\u0000\":5,\"\":[]}",
+    "workflow": "{\"10\":1,\"3\":2,\"b\":3,\"a\":4,\"__proto__\":{\"x\":1},\"\":{}}"
+   },
+   "expect": {
+    "prompt_graph": "parsed",
+    "workflow_graph": "parsed"
+   }
+  },
+  {
+   "name": "duplicate_key_top_level",
+   "records": {
+    "workflow": "{\"a\":1,\"a\":2}"
+   },
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "duplicate_key_nested",
+   "records": {
+    "workflow": "{\"a\":{\"b\":[{\"c\":1,\"c\":1}]}}"
+   },
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "duplicate_key_in_dropped_prompt_meta",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"a\":1},\"_meta\":{\"t\":1,\"t\":2}}}"
+   },
+   "expect": {
+    "prompt_graph": "malformed",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "duplicate_key_in_prompt_inputs",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"a\":1,\"a\":1}}}"
+   },
+   "expect": {
+    "prompt_graph": "malformed",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "duplicate_key_after_deep_nest",
+   "records": {
+    "workflow": "{\"x\":[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]],\"a\":1,\"a\":2}"
+   },
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "workflow_depth_64_kept",
+   "records": {
+    "workflow": "{\"a\":[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]}"
+   },
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "parsed"
+   }
+  },
+  {
+   "name": "workflow_depth_65_limited",
+   "records": {
+    "workflow": "{\"a\":[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]}"
+   },
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "limited"
+   }
+  },
+  {
+   "name": "prompt_inputs_depth_64_kept",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"v\":[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]}}}"
+   },
+   "expect": {
+    "prompt_graph": "parsed",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "prompt_inputs_depth_65_limited",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"v\":[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]}}}"
+   },
+   "expect": {
+    "prompt_graph": "limited",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "prompt_dropped_meta_depth_65_limited",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{},\"_meta\":{\"v\":[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]}}}"
+   },
+   "expect": {
+    "prompt_graph": "limited",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "prompt_non_object_nodes_kept",
+   "records": {
+    "prompt": "{\"1\":5,\"2\":[1,2.50,\"x\",null,true],\"3\":\"s\",\"4\":null,\"5\":{\"inputs\":[1],\"class_type\":null,\"other\":{\"deep\":[NaN]}}}"
+   },
+   "expect": {
+    "prompt_graph": "parsed",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "prompt_root_array_malformed",
+   "records": {
+    "prompt": "[]"
+   },
+   "expect": {
+    "prompt_graph": "malformed",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "prompt_root_string_malformed",
+   "records": {
+    "prompt": "\"{}\""
+   },
+   "expect": {
+    "prompt_graph": "malformed",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "prompt_root_number_malformed",
+   "records": {
+    "prompt": "5"
+   },
+   "expect": {
+    "prompt_graph": "malformed",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "prompt_root_null_malformed",
+   "records": {
+    "prompt": "null"
+   },
+   "expect": {
+    "prompt_graph": "malformed",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "workflow_root_array_malformed",
+   "records": {
+    "workflow": "[1,2]"
+   },
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "workflow_root_nan_malformed",
+   "records": {
+    "workflow": "NaN"
+   },
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "empty_text_malformed",
+   "records": {
+    "prompt": "",
+    "workflow": "   "
+   },
+   "expect": {
+    "prompt_graph": "malformed",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "bom_malformed",
+   "records": {
+    "workflow": "{}"
+   },
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "trailing_data_malformed",
+   "records": {
+    "workflow": "{\"a\":1} x",
+    "prompt": "{}\u0000"
+   },
+   "expect": {
+    "prompt_graph": "malformed",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "python_whitespace_only",
+   "records": {
+    "workflow": " \t\r\n{\n\t\"a\" :\r 1 }\n",
+    "prompt": "\u000b{}"
+   },
+   "expect": {
+    "prompt_graph": "malformed",
+    "workflow_graph": "parsed"
+   }
+  },
+  {
+   "name": "syntax_errors",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{},}}",
+    "workflow": "[1,]"
+   },
+   "expect": {
+    "prompt_graph": "malformed",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "bad_escape_and_control_char",
+   "records": {
+    "prompt": "{\"1\":\"\\x\"}",
+    "workflow": "{\"a\":\"\u0001\"}"
+   },
+   "expect": {
+    "prompt_graph": "malformed",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "number_grammar",
+   "records": {
+    "prompt": "{\"1\":01}",
+    "workflow": "{\"a\":1.,\"b\":.5}"
+   },
+   "expect": {
+    "prompt_graph": "malformed",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "nodes_not_list_counts_zero",
+   "records": {
+    "workflow": "{\"nodes\":{\"a\":1},\"definitions\":{\"subgraphs\":[{\"nodes\":5},7,{\"nodes\":[1,2]}]}}"
+   },
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "parsed"
+   }
+  },
+  {
+   "name": "limited_workflow_record",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"x\":1}}}",
+    "workflow": "{\"nodes\":[],\"version\":0.4}"
+   },
+   "limited": [
+    "workflow"
+   ],
+   "expect": {
+    "prompt_graph": "parsed",
+    "workflow_graph": "limited"
+   }
+  },
+  {
+   "name": "limited_workflow_without_text",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"x\":1}}}"
+   },
+   "limited": [
+    "workflow"
+   ],
+   "expect": {
+    "prompt_graph": "parsed",
+    "workflow_graph": "limited"
+   }
+  },
+  {
+   "name": "invalid_workflow_format",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"x\":1}}}"
+   },
+   "record_formats": {
+    "prompt": "original-text",
+    "workflow": "invalid"
+   },
+   "expect": {
+    "prompt_graph": "parsed",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "reserialized_without_carrier_is_limited",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"x\":1}}}",
+    "workflow": "{\"nodes\":[],\"version\":0.4}"
+   },
+   "record_formats": {
+    "prompt": "reserialized-json",
+    "workflow": "reserialized-json"
+   },
+   "expect": {
+    "prompt_graph": "limited",
+    "workflow_graph": "limited"
+   }
+  },
+  {
+   "name": "unknown_format_dropped_but_projected",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"x\":1}}}"
+   },
+   "record_formats": {
+    "prompt": "weird"
+   },
+   "expect": {
+    "prompt_graph": "parsed",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "non_text_record_dropped",
+   "records": {
+    "prompt": 5,
+    "workflow": "{\"nodes\":[],\"version\":0.4}"
+   },
+   "record_formats": {
+    "prompt": "original-text",
+    "workflow": "original-text"
+   },
+   "expect": {
+    "prompt_graph": "malformed",
+    "workflow_graph": "parsed"
+   }
+  },
+  {
+   "name": "format_without_text_is_malformed",
+   "records": {},
+   "record_formats": {
+    "prompt": "original-text"
+   },
+   "expect": {
+    "prompt_graph": "malformed",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "unknown_record_keys_ignored",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"x\":1}}}",
+    "Software": "ComfyUI",
+    "WORKFLOW": "{\"nodes\":[],\"version\":0.4}"
+   },
+   "record_formats": {
+    "prompt": "original-text",
+    "Software": "original-text",
+    "WORKFLOW": "original-text"
+   },
+   "expect": {
+    "prompt_graph": "parsed",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "carrier_prompt_string_workflow_object",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"x\":1}}}",
+    "workflow": "{\"nodes\": [], \"v\": 1}"
+   },
+   "record_formats": {
+    "prompt": "original-text",
+    "workflow": "reserialized-json"
+   },
+   "carrier_text": "{\"prompt\":\"{\\\"1\\\":{\\\"class_type\\\":\\\"A\\\",\\\"inputs\\\":{\\\"x\\\":1}}}\",\"workflow\":{\"nodes\":[],\"v\":1}}",
+   "expect": {
+    "prompt_graph": "parsed",
+    "workflow_graph": "parsed"
+   }
+  },
+  {
+   "name": "carrier_prompt_object_projected",
+   "records": {
+    "prompt": "{\"1\": {\"class_type\": \"A\", \"inputs\": {\"x\": 8.0}, \"_meta\": {\"t\": \"x\"}}}",
+    "workflow": "{\"nodes\":[],\"version\":0.4}"
+   },
+   "record_formats": {
+    "prompt": "reserialized-json",
+    "workflow": "original-text"
+   },
+   "carrier_text": "{\"prompt\":{\"1\":{\"class_type\":\"A\",\"inputs\":{\"x\":8.0},\"_meta\":{\"t\":\"x\"}}},\"workflow\":{\"nodes\":[],\"version\":0.4},\"title\":\"n\",\"extra\":[1e3]}",
+   "expect": {
+    "prompt_graph": "parsed",
+    "workflow_graph": "parsed"
+   }
+  },
+  {
+   "name": "carrier_root_array_fails_all",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"x\":1}}}",
+    "workflow": "{\"nodes\":[],\"version\":0.4}"
+   },
+   "carrier_text": "[1]",
+   "expect": {
+    "prompt_graph": "malformed",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "carrier_duplicate_key_fails_all",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"x\":1}}}"
+   },
+   "carrier_text": "{\"prompt\":\"{}\",\"prompt\":\"{}\"}",
+   "expect": {
+    "prompt_graph": "malformed",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "carrier_other_member_too_deep_limited",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"x\":1}}}"
+   },
+   "carrier_text": "{\"title\":[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]],\"prompt\":\"{\\\"1\\\":{\\\"class_type\\\":\\\"A\\\",\\\"inputs\\\":{\\\"x\\\":1}}}\"}",
+   "expect": {
+    "prompt_graph": "limited",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "carrier_other_member_depth_64_kept",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"x\":1}}}"
+   },
+   "carrier_text": "{\"title\":[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]],\"prompt\":\"{\\\"1\\\":{\\\"class_type\\\":\\\"A\\\",\\\"inputs\\\":{\\\"x\\\":1}}}\"}",
+   "expect": {
+    "prompt_graph": "parsed",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "carrier_prompt_object_depth_shift",
+   "records": {
+    "prompt": "x"
+   },
+   "record_formats": {
+    "prompt": "reserialized-json"
+   },
+   "carrier_text": "{\"prompt\":{\"1\":{\"class_type\":\"A\",\"inputs\":{\"v\":[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]}}}}",
+   "expect": {
+    "prompt_graph": "parsed",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "carrier_prompt_object_depth_shift_limited",
+   "records": {
+    "prompt": "x"
+   },
+   "record_formats": {
+    "prompt": "reserialized-json"
+   },
+   "carrier_text": "{\"prompt\":{\"1\":{\"class_type\":\"A\",\"inputs\":{\"v\":[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]}}}}",
+   "expect": {
+    "prompt_graph": "limited",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "carrier_prompt_string_invalid_json",
+   "records": {
+    "prompt": "{",
+    "workflow": "{\"nodes\":[],\"version\":0.4}"
+   },
+   "carrier_text": "{\"prompt\":\"{\",\"workflow\":\"{\\\"nodes\\\":[],\\\"version\\\":0.4}\"}",
+   "expect": {
+    "prompt_graph": "malformed",
+    "workflow_graph": "parsed"
+   }
+  },
+  {
+   "name": "carrier_prompt_string_too_deep",
+   "records": {
+    "prompt": "x"
+   },
+   "carrier_text": "{\"prompt\":\"{\\\"1\\\":{\\\"class_type\\\":\\\"A\\\",\\\"inputs\\\":{\\\"v\\\":[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]}}}\"}",
+   "expect": {
+    "prompt_graph": "limited",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "carrier_workflow_number_malformed",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"x\":1}}}",
+    "workflow": "5"
+   },
+   "record_formats": {
+    "prompt": "original-text",
+    "workflow": "original-text"
+   },
+   "carrier_text": "{\"prompt\":\"{\\\"1\\\":{\\\"class_type\\\":\\\"A\\\",\\\"inputs\\\":{\\\"x\\\":1}}}\",\"workflow\":5}",
+   "expect": {
+    "prompt_graph": "parsed",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "carrier_prompt_nonfinite_malformed",
+   "records": {
+    "prompt": "NaN"
+   },
+   "carrier_text": "{\"prompt\":NaN}",
+   "expect": {
+    "prompt_graph": "malformed",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "carrier_prompt_bigint_malformed",
+   "records": {
+    "prompt": "123456789012345678901"
+   },
+   "carrier_text": "{\"prompt\":123456789012345678901}",
+   "expect": {
+    "prompt_graph": "malformed",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "carrier_prompt_decimal_malformed",
+   "records": {
+    "prompt": "1.5"
+   },
+   "carrier_text": "{\"prompt\":1.5}",
+   "expect": {
+    "prompt_graph": "malformed",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "carrier_missing_member_malformed",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"x\":1}}}"
+   },
+   "carrier_text": "{\"workflow\":{}}",
+   "expect": {
+    "prompt_graph": "malformed",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "carrier_without_graph_records",
+   "records": {
+    "parameters": "cat"
+   },
+   "carrier_text": "{\"parameters\":\"cat\"}",
+   "expect": {
+    "prompt_graph": "absent",
+    "workflow_graph": "absent"
+   }
+  },
+  {
+   "name": "carrier_with_limited_workflow",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"x\":1}}}"
+   },
+   "limited": [
+    "workflow"
+   ],
+   "carrier_text": "{\"prompt\":\"{\\\"1\\\":{\\\"class_type\\\":\\\"A\\\",\\\"inputs\\\":{\\\"x\\\":1}}}\"}",
+   "expect": {
+    "prompt_graph": "parsed",
+    "workflow_graph": "limited"
+   }
+  },
+  {
+   "name": "carrier_invalid_workflow_format",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"x\":1}}}"
+   },
+   "record_formats": {
+    "prompt": "original-text",
+    "workflow": "invalid"
+   },
+   "carrier_text": "{\"prompt\":\"{\\\"1\\\":{\\\"class_type\\\":\\\"A\\\",\\\"inputs\\\":{\\\"x\\\":1}}}\",\"workflow\":null}",
+   "expect": {
+    "prompt_graph": "parsed",
+    "workflow_graph": "malformed"
+   }
+  },
+  {
+   "name": "carrier_non_string_text_malformed",
+   "records": {
+    "prompt": "{\"1\":{\"class_type\":\"A\",\"inputs\":{\"x\":1}}}"
+   },
+   "carrier_text": "",
+   "expect": {
+    "prompt_graph": "malformed",
+    "workflow_graph": "absent"
+   }
+  }
+ ]
+}

tests/fixtures/local-inspect/video-tags-v1.json

@@ -0,0 +1,60 @@
+{
+  "generator": "ffmpeg -show_entries format_tags -of json over tests/fixtures/local-inspect, see tests/fixtures/README.md",
+  "fixtures": {
+    "core-faststart.mp4": {
+      "major_brand": "isom",
+      "minor_version": "512",
+      "compatible_brands": "isomiso2avc1mp41",
+      "prompt": "{\"3\":{\"class_type\":\"KSampler\",\"inputs\":{\"seed\":7,\"steps\":20,\"cfg\":8.0,\"sampler_name\":\"euler\",\"scheduler\":\"normal\",\"denoise\":1.0,\"model\":[\"4\",0],\"positive\":[\"6\",0],\"negative\":[\"7\",0],\"latent_image\":[\"5\",0]}},\"4\":{\"class_type\":\"CheckpointLoaderSimple\",\"inputs\":{\"ckpt_name\":\"synthetic-v1.safetensors\"}},\"5\":{\"class_type\":\"EmptyLatentImage\",\"inputs\":{\"width\":512,\"height\":512,\"batch_size\":1}},\"6\":{\"class_type\":\"CLIPTextEncode\",\"inputs\":{\"text\":\"a red fox in snow, detailed fur\",\"clip\":[\"4\",1]}},\"7\":{\"class_type\":\"CLIPTextEncode\",\"inputs\":{\"text\":\"blurry, lowres\",\"clip\":[\"4\",1]}},\"8\":{\"class_type\":\"VAEDecode\",\"inputs\":{\"samples\":[\"3\",0],\"vae\":[\"4\",2]}},\"9\":{\"class_type\":\"SaveVideo\",\"inputs\":{\"filename_prefix\":\"video/synthetic\",\"video\":[\"8\",0]}}}",
+      "workflow": "{\"last_node_id\":9,\"nodes\":[{\"id\":3,\"type\":\"KSampler\",\"widgets_values\":[7,\"fixed\",20,8.0,\"euler\",\"normal\",1.0]},{\"id\":6,\"type\":\"CLIPTextEncode\",\"widgets_values\":[\"a red fox in snow, detailed fur\"]},{\"id\":7,\"type\":\"CLIPTextEncode\",\"widgets_values\":[\"blurry, lowres\"]}],\"links\":[],\"version\":0.4}",
+      "parameters": "a red fox in snow, detailed fur\nNegative prompt: blurry, lowres\nSteps: 20, Sampler: Euler, CFG scale: 8.0, Seed: 7, Size: 512x512",
+      "comfyassets_prompt_pairs": "{\"version\":1,\"nodes\":{\"6\":{\"records\":[{\"positive_text\":\"a red fox in snow, detailed fur\",\"negative_text\":\"blurry, lowres\",\"negative_effective_text\":\"blurry, lowres\",\"prepend_text\":\"\",\"append_text\":\"\"}]}}}",
+      "encoder": "Lavf63.1.102"
+    },
+    "core-moov-end.mp4": {
+      "major_brand": "isom",
+      "minor_version": "512",
+      "compatible_brands": "isomiso2avc1mp41",
+      "prompt": "{\"3\":{\"class_type\":\"KSampler\",\"inputs\":{\"seed\":7,\"steps\":20,\"cfg\":8.0,\"sampler_name\":\"euler\",\"scheduler\":\"normal\",\"denoise\":1.0,\"model\":[\"4\",0],\"positive\":[\"6\",0],\"negative\":[\"7\",0],\"latent_image\":[\"5\",0]}},\"4\":{\"class_type\":\"CheckpointLoaderSimple\",\"inputs\":{\"ckpt_name\":\"synthetic-v1.safetensors\"}},\"5\":{\"class_type\":\"EmptyLatentImage\",\"inputs\":{\"width\":512,\"height\":512,\"batch_size\":1}},\"6\":{\"class_type\":\"CLIPTextEncode\",\"inputs\":{\"text\":\"a red fox in snow, detailed fur\",\"clip\":[\"4\",1]}},\"7\":{\"class_type\":\"CLIPTextEncode\",\"inputs\":{\"text\":\"blurry, lowres\",\"clip\":[\"4\",1]}},\"8\":{\"class_type\":\"VAEDecode\",\"inputs\":{\"samples\":[\"3\",0],\"vae\":[\"4\",2]}},\"9\":{\"class_type\":\"SaveVideo\",\"inputs\":{\"filename_prefix\":\"video/synthetic\",\"video\":[\"8\",0]}}}",
+      "workflow": "{\"last_node_id\":9,\"nodes\":[{\"id\":3,\"type\":\"KSampler\",\"widgets_values\":[7,\"fixed\",20,8.0,\"euler\",\"normal\",1.0]},{\"id\":6,\"type\":\"CLIPTextEncode\",\"widgets_values\":[\"a red fox in snow, detailed fur\"]},{\"id\":7,\"type\":\"CLIPTextEncode\",\"widgets_values\":[\"blurry, lowres\"]}],\"links\":[],\"version\":0.4}",
+      "encoder": "Lavf63.1.102"
+    },
+    "core.mov": {
+      "major_brand": "qt  ",
+      "minor_version": "512",
+      "compatible_brands": "qt  ",
+      "prompt": "{\"3\":{\"class_type\":\"KSampler\",\"inputs\":{\"seed\":7,\"steps\":20,\"cfg\":8.0,\"sampler_name\":\"euler\",\"scheduler\":\"normal\",\"denoise\":1.0,\"model\":[\"4\",0],\"positive\":[\"6\",0],\"negative\":[\"7\",0],\"latent_image\":[\"5\",0]}},\"4\":{\"class_type\":\"CheckpointLoaderSimple\",\"inputs\":{\"ckpt_name\":\"synthetic-v1.safetensors\"}},\"5\":{\"class_type\":\"EmptyLatentImage\",\"inputs\":{\"width\":512,\"height\":512,\"batch_size\":1}},\"6\":{\"class_type\":\"CLIPTextEncode\",\"inputs\":{\"text\":\"a red fox in snow, detailed fur\",\"clip\":[\"4\",1]}},\"7\":{\"class_type\":\"CLIPTextEncode\",\"inputs\":{\"text\":\"blurry, lowres\",\"clip\":[\"4\",1]}},\"8\":{\"class_type\":\"VAEDecode\",\"inputs\":{\"samples\":[\"3\",0],\"vae\":[\"4\",2]}},\"9\":{\"class_type\":\"SaveVideo\",\"inputs\":{\"filename_prefix\":\"video/synthetic\",\"video\":[\"8\",0]}}}",
+      "workflow": "{\"last_node_id\":9,\"nodes\":[{\"id\":3,\"type\":\"KSampler\",\"widgets_values\":[7,\"fixed\",20,8.0,\"euler\",\"normal\",1.0]},{\"id\":6,\"type\":\"CLIPTextEncode\",\"widgets_values\":[\"a red fox in snow, detailed fur\"]},{\"id\":7,\"type\":\"CLIPTextEncode\",\"widgets_values\":[\"blurry, lowres\"]}],\"links\":[],\"version\":0.4}",
+      "encoder": "Lavf63.1.102"
+    },
+    "core.m4v": {
+      "major_brand": "M4V ",
+      "minor_version": "512",
+      "compatible_brands": "M4V isomiso2avc1",
+      "prompt": "{\"3\":{\"class_type\":\"KSampler\",\"inputs\":{\"seed\":7,\"steps\":20,\"cfg\":8.0,\"sampler_name\":\"euler\",\"scheduler\":\"normal\",\"denoise\":1.0,\"model\":[\"4\",0],\"positive\":[\"6\",0],\"negative\":[\"7\",0],\"latent_image\":[\"5\",0]}},\"4\":{\"class_type\":\"CheckpointLoaderSimple\",\"inputs\":{\"ckpt_name\":\"synthetic-v1.safetensors\"}},\"5\":{\"class_type\":\"EmptyLatentImage\",\"inputs\":{\"width\":512,\"height\":512,\"batch_size\":1}},\"6\":{\"class_type\":\"CLIPTextEncode\",\"inputs\":{\"text\":\"a red fox in snow, detailed fur\",\"clip\":[\"4\",1]}},\"7\":{\"class_type\":\"CLIPTextEncode\",\"inputs\":{\"text\":\"blurry, lowres\",\"clip\":[\"4\",1]}},\"8\":{\"class_type\":\"VAEDecode\",\"inputs\":{\"samples\":[\"3\",0],\"vae\":[\"4\",2]}},\"9\":{\"class_type\":\"SaveVideo\",\"inputs\":{\"filename_prefix\":\"video/synthetic\",\"video\":[\"8\",0]}}}",
+      "workflow": "{\"last_node_id\":9,\"nodes\":[{\"id\":3,\"type\":\"KSampler\",\"widgets_values\":[7,\"fixed\",20,8.0,\"euler\",\"normal\",1.0]},{\"id\":6,\"type\":\"CLIPTextEncode\",\"widgets_values\":[\"a red fox in snow, detailed fur\"]},{\"id\":7,\"type\":\"CLIPTextEncode\",\"widgets_values\":[\"blurry, lowres\"]}],\"links\":[],\"version\":0.4}",
+      "encoder": "Lavf63.1.102"
+    },
+    "core.mkv": {
+      "WORKFLOW": "{\"last_node_id\":9,\"nodes\":[{\"id\":3,\"type\":\"KSampler\",\"widgets_values\":[7,\"fixed\",20,8.0,\"euler\",\"normal\",1.0]},{\"id\":6,\"type\":\"CLIPTextEncode\",\"widgets_values\":[\"a red fox in snow, detailed fur\"]},{\"id\":7,\"type\":\"CLIPTextEncode\",\"widgets_values\":[\"blurry, lowres\"]}],\"links\":[],\"version\":0.4}",
+      "PROMPT": "{\"3\":{\"class_type\":\"KSampler\",\"inputs\":{\"seed\":7,\"steps\":20,\"cfg\":8.0,\"sampler_name\":\"euler\",\"scheduler\":\"normal\",\"denoise\":1.0,\"model\":[\"4\",0],\"positive\":[\"6\",0],\"negative\":[\"7\",0],\"latent_image\":[\"5\",0]}},\"4\":{\"class_type\":\"CheckpointLoaderSimple\",\"inputs\":{\"ckpt_name\":\"synthetic-v1.safetensors\"}},\"5\":{\"class_type\":\"EmptyLatentImage\",\"inputs\":{\"width\":512,\"height\":512,\"batch_size\":1}},\"6\":{\"class_type\":\"CLIPTextEncode\",\"inputs\":{\"text\":\"a red fox in snow, detailed fur\",\"clip\":[\"4\",1]}},\"7\":{\"class_type\":\"CLIPTextEncode\",\"inputs\":{\"text\":\"blurry, lowres\",\"clip\":[\"4\",1]}},\"8\":{\"class_type\":\"VAEDecode\",\"inputs\":{\"samples\":[\"3\",0],\"vae\":[\"4\",2]}},\"9\":{\"class_type\":\"SaveVideo\",\"inputs\":{\"filename_prefix\":\"video/synthetic\",\"video\":[\"8\",0]}}}",
+      "ENCODER": "Lavf63.1.102"
+    },
+    "core-tags.webm": {
+      "WORKFLOW": "{\"last_node_id\":9,\"nodes\":[{\"id\":3,\"type\":\"KSampler\",\"widgets_values\":[7,\"fixed\",20,8.0,\"euler\",\"normal\",1.0]},{\"id\":6,\"type\":\"CLIPTextEncode\",\"widgets_values\":[\"a red fox in snow, detailed fur\"]},{\"id\":7,\"type\":\"CLIPTextEncode\",\"widgets_values\":[\"blurry, lowres\"]}],\"links\":[],\"version\":0.4}",
+      "PROMPT": "{\"3\":{\"class_type\":\"KSampler\",\"inputs\":{\"seed\":7,\"steps\":20,\"cfg\":8.0,\"sampler_name\":\"euler\",\"scheduler\":\"normal\",\"denoise\":1.0,\"model\":[\"4\",0],\"positive\":[\"6\",0],\"negative\":[\"7\",0],\"latent_image\":[\"5\",0]}},\"4\":{\"class_type\":\"CheckpointLoaderSimple\",\"inputs\":{\"ckpt_name\":\"synthetic-v1.safetensors\"}},\"5\":{\"class_type\":\"EmptyLatentImage\",\"inputs\":{\"width\":512,\"height\":512,\"batch_size\":1}},\"6\":{\"class_type\":\"CLIPTextEncode\",\"inputs\":{\"text\":\"a red fox in snow, detailed fur\",\"clip\":[\"4\",1]}},\"7\":{\"class_type\":\"CLIPTextEncode\",\"inputs\":{\"text\":\"blurry, lowres\",\"clip\":[\"4\",1]}},\"8\":{\"class_type\":\"VAEDecode\",\"inputs\":{\"samples\":[\"3\",0],\"vae\":[\"4\",2]}},\"9\":{\"class_type\":\"SaveVideo\",\"inputs\":{\"filename_prefix\":\"video/synthetic\",\"video\":[\"8\",0]}}}",
+      "ENCODER": "Lavf63.1.102"
+    },
+    "legacy-comment.mp4": {
+      "major_brand": "isom",
+      "minor_version": "512",
+      "compatible_brands": "isomiso2avc1mp41",
+      "encoder": "Lavf63.1.102",
+      "comment": "{\"prompt\": \"{\\\"3\\\":{\\\"class_type\\\":\\\"KSampler\\\",\\\"inputs\\\":{\\\"seed\\\":7,\\\"steps\\\":20,\\\"cfg\\\":8.0,\\\"sampler_name\\\":\\\"euler\\\",\\\"scheduler\\\":\\\"normal\\\",\\\"denoise\\\":1.0,\\\"model\\\":[\\\"4\\\",0],\\\"positive\\\":[\\\"6\\\",0],\\\"negative\\\":[\\\"7\\\",0],\\\"latent_image\\\":[\\\"5\\\",0]}},\\\"4\\\":{\\\"class_type\\\":\\\"CheckpointLoaderSimple\\\",\\\"inputs\\\":{\\\"ckpt_name\\\":\\\"synthetic-v1.safetensors\\\"}},\\\"5\\\":{\\\"class_type\\\":\\\"EmptyLatentImage\\\",\\\"inputs\\\":{\\\"width\\\":512,\\\"height\\\":512,\\\"batch_size\\\":1}},\\\"6\\\":{\\\"class_type\\\":\\\"CLIPTextEncode\\\",\\\"inputs\\\":{\\\"text\\\":\\\"a red fox in snow, detailed fur\\\",\\\"clip\\\":[\\\"4\\\",1]}},\\\"7\\\":{\\\"class_type\\\":\\\"CLIPTextEncode\\\",\\\"inputs\\\":{\\\"text\\\":\\\"blurry, lowres\\\",\\\"clip\\\":[\\\"4\\\",1]}},\\\"8\\\":{\\\"class_type\\\":\\\"VAEDecode\\\",\\\"inputs\\\":{\\\"samples\\\":[\\\"3\\\",0],\\\"vae\\\":[\\\"4\\\",2]}},\\\"9\\\":{\\\"class_type\\\":\\\"SaveVideo\\\",\\\"inputs\\\":{\\\"filename_prefix\\\":\\\"video/synthetic\\\",\\\"video\\\":[\\\"8\\\",0]}}}\", \"workflow\": {\"last_node_id\": 9, \"nodes\": [{\"id\": 3, \"type\": \"KSampler\", \"widgets_values\": [7, \"fixed\", 20, 8.0, \"euler\", \"normal\", 1.0]}, {\"id\": 6, \"type\": \"CLIPTextEncode\", \"widgets_values\": [\"a red fox in snow, detailed fur\"]}, {\"id\": 7, \"type\": \"CLIPTextEncode\", \"widgets_values\": [\"blurry, lowres\"]}], \"links\": [], \"version\": 0.4}}"
+    },
+    "legacy-comment.webm": {
+      "COMMENT": "{\"prompt\": \"{\\\"3\\\":{\\\"class_type\\\":\\\"KSampler\\\",\\\"inputs\\\":{\\\"seed\\\":7,\\\"steps\\\":20,\\\"cfg\\\":8.0,\\\"sampler_name\\\":\\\"euler\\\",\\\"scheduler\\\":\\\"normal\\\",\\\"denoise\\\":1.0,\\\"model\\\":[\\\"4\\\",0],\\\"positive\\\":[\\\"6\\\",0],\\\"negative\\\":[\\\"7\\\",0],\\\"latent_image\\\":[\\\"5\\\",0]}},\\\"4\\\":{\\\"class_type\\\":\\\"CheckpointLoaderSimple\\\",\\\"inputs\\\":{\\\"ckpt_name\\\":\\\"synthetic-v1.safetensors\\\"}},\\\"5\\\":{\\\"class_type\\\":\\\"EmptyLatentImage\\\",\\\"inputs\\\":{\\\"width\\\":512,\\\"height\\\":512,\\\"batch_size\\\":1}},\\\"6\\\":{\\\"class_type\\\":\\\"CLIPTextEncode\\\",\\\"inputs\\\":{\\\"text\\\":\\\"a red fox in snow, detailed fur\\\",\\\"clip\\\":[\\\"4\\\",1]}},\\\"7\\\":{\\\"class_type\\\":\\\"CLIPTextEncode\\\",\\\"inputs\\\":{\\\"text\\\":\\\"blurry, lowres\\\",\\\"clip\\\":[\\\"4\\\",1]}},\\\"8\\\":{\\\"class_type\\\":\\\"VAEDecode\\\",\\\"inputs\\\":{\\\"samples\\\":[\\\"3\\\",0],\\\"vae\\\":[\\\"4\\\",2]}},\\\"9\\\":{\\\"class_type\\\":\\\"SaveVideo\\\",\\\"inputs\\\":{\\\"filename_prefix\\\":\\\"video/synthetic\\\",\\\"video\\\":[\\\"8\\\",0]}}}\", \"workflow\": {\"last_node_id\": 9, \"nodes\": [{\"id\": 3, \"type\": \"KSampler\", \"widgets_values\": [7, \"fixed\", 20, 8.0, \"euler\", \"normal\", 1.0]}, {\"id\": 6, \"type\": \"CLIPTextEncode\", \"widgets_values\": [\"a red fox in snow, detailed fur\"]}, {\"id\": 7, \"type\": \"CLIPTextEncode\", \"widgets_values\": [\"blurry, lowres\"]}], \"links\": [], \"version\": 0.4}}",
+      "ENCODER": "Lavf63.1.102"
+    }
+  }
+}

tests/host_stubs.py

@@ -0,0 +1,29 @@
+"""One stand-in for ComfyUI's ``folder_paths`` module, shared by every test.
+
+Test modules that import the API standalone need a ``folder_paths`` module.
+It must be a real module, not a ``MagicMock``: ``patch("folder_paths.x",
+create=True)`` then adds and removes plain attributes, while a mock would
+hand a later test an auto-created child (or lose the attribute on exit), so
+the outcome depended on which test module discovery imported first.
+"""
+
+import sys
+import tempfile
+import types
+
+
+def install_folder_paths_stub() -> types.ModuleType:
+    """Install the shared ``folder_paths`` stub unless a real module is there.
+
+    A test that installs its own stand-in for a while (and restores it) is
+    left alone; a ``MagicMock`` left behind by older code is replaced.
+    """
+    current = sys.modules.get("folder_paths")
+    if isinstance(current, types.ModuleType):
+        return current
+    stub = types.ModuleType("folder_paths")
+    stub.models_dir = tempfile.gettempdir()
+    stub.base_path = tempfile.gettempdir()
+    stub.get_output_directory = tempfile.gettempdir
+    sys.modules["folder_paths"] = stub
+    return stub

tests/js/api-client.test.js

@@ -296,3 +296,33 @@ test("get forwards an explicit cache mode and keeps the default request otherwis
     fetch.calls[2].resolve(jsonResponse(200, { success: true }));
     await writes;
 });
+
+test("ApiError exposes only a validated current tag set from a tag-conflict body", async () => {
+    const permitted = [{ tags: [], tag_revision: 0 }, { tags: ["warm", "Dune"], tag_revision: 7 }];
+    const denied = [undefined, null, { tags: ["a"] }, { tag_revision: 1 }, { tags: "a", tag_revision: 1 },
+        { tags: [1], tag_revision: 1 }, { tags: ["a"], tag_revision: -1 }, { tags: ["a"], tag_revision: 1.5 },
+        { tags: ["a"], tag_revision: true }, { tags: Array(101).fill("a"), tag_revision: 1 }];
+    assert.equal(new ApiError(409, "old caller").current, null);
+    for (const current of permitted.concat(denied)) {
+        const api = createApiClient(async () => jsonResponse(409, {
+            success: false, error: "changed", code: "tag-conflict", ...(current || {}),
+        }));
+        await assert.rejects(api.put("/x", {}), error => {
+            assert.ok(error instanceof ApiError);
+            assert.deepEqual(error.current, permitted.includes(current) ? current : null);
+            if (permitted.includes(current)) assert.notEqual(error.current.tags, current.tags, "server arrays must be copied");
+            return true;
+        });
+    }
+});
+
+test("keepalive is forwarded only when asked, so a request can outlive its page", async () => {
+    const fetch = fakeFetch();
+    const api = createApiClient(fetch);
+    const writes = [api.del("/x", { keepalive: true }), api.put("/x", {}, { keepalive: true }), api.post("/x", {}, { keepalive: true }),
+        api.del("/x"), api.put("/x", {}), api.get("/x")];
+    assert.deepEqual(fetch.calls.map(call => call.init.keepalive), [true, true, true, undefined, undefined, undefined]);
+    assert.equal("keepalive" in fetch.calls[3].init, false);
+    for (const call of fetch.calls) call.resolve(jsonResponse(200, { success: true }));
+    await Promise.all(writes);
+});

tests/js/helpers/png-bytes.mjs

@@ -0,0 +1,148 @@
+// Byte-exact PNG builders for the local reader tests (real CRC32 and zlib), and
+// File-like blobs: one that records every [start, end) slice over a sparse layout
+// that is never allocated, one that answers slowly, one that fails or shrinks.
+import zlib from "node:zlib";
+
+export const SIGNATURE = Uint8Array.from([0x89, 0x50, 0x4e, 0x47, 0x0d, 0x0a, 0x1a, 0x0a]);
+
+export function u32(n) {
+    return Uint8Array.from([(n >>> 24) & 0xff, (n >>> 16) & 0xff, (n >>> 8) & 0xff, n & 0xff]);
+}
+
+export function latin1(text) {
+    return Buffer.from(text, "latin1");
+}
+
+export function utf8(text) {
+    return Buffer.from(text, "utf8");
+}
+
+export function concat(...parts) {
+    return Uint8Array.from(Buffer.concat(parts.map((part) => Buffer.from(part))));
+}
+
+/**
+ * One chunk with a real CRC32 over type + data. `crc` overrides the checksum and
+ * `declaredLength` the length field (to declare more data than is present).
+ */
+export function chunk(type, data, { crc, declaredLength } = {}) {
+    const typeBytes = latin1(type);
+    const body = Buffer.from(data);
+    const sum = crc ?? zlib.crc32(Buffer.concat([typeBytes, body]));
+    return concat(u32(declaredLength ?? body.length), typeBytes, body, u32(sum));
+}
+
+export function ihdr(width = 1, height = 1, options) {
+    return chunk("IHDR", concat(u32(width), u32(height), [8, 6, 0, 0, 0]), options);
+}
+
+export function idat(data = [0]) {
+    return chunk("IDAT", data);
+}
+
+export function iend(options) {
+    return chunk("IEND", [], options);
+}
+
+/** tEXt: keyword NUL text, Latin-1 (text may be a string or raw bytes). */
+export function tEXt(keyword, text, options) {
+    const body = typeof text === "string" ? latin1(text) : Buffer.from(text);
+    return chunk("tEXt", concat(latin1(keyword), [0], body), options);
+}
+
+/** zTXt: keyword NUL method zlib(text). `compressed` replaces the deflated body. */
+export function zTXt(keyword, text, { method = 0, compressed, level = 6 } = {}) {
+    const body = compressed ?? zlib.deflateSync(typeof text === "string" ? latin1(text) : Buffer.from(text), { level });
+    return chunk("zTXt", concat(latin1(keyword), [0, method], body));
+}
+
+/** iTXt: keyword NUL flag method lang NUL translated NUL text (UTF-8, deflated when flag). */
+export function iTXt(keyword, text, { flag = 0, method = 0, lang = "", translated = "", compressed } = {}) {
+    const raw = typeof text === "string" ? utf8(text) : Buffer.from(text);
+    const body = compressed ?? (flag ? zlib.deflateSync(raw) : raw);
+    const langBytes = typeof lang === "string" ? latin1(lang) : Buffer.from(lang);
+    const translatedBytes = typeof translated === "string" ? utf8(translated) : Buffer.from(translated);
+    return concat(chunk("iTXt", concat(latin1(keyword), [0, flag, method], langBytes, [0], translatedBytes, [0], body)));
+}
+
+/** A whole PNG: signature, IHDR, IDAT, the given chunks, IEND. */
+export function png(...chunks) {
+    return concat(SIGNATURE, ihdr(), idat(), ...chunks, iend());
+}
+
+/** Signature plus exactly the given chunks (no implied IHDR/IDAT/IEND). */
+export function raw(...chunks) {
+    return concat(SIGNATURE, ...chunks);
+}
+
+export function blobOf(bytes) {
+    return new Blob([bytes]);
+}
+
+/**
+ * A File-like object over a sparse layout: `parts` are [{offset, bytes}] placed in a
+ * file of `size` bytes, zero everywhere else; nothing of `size` is allocated. Every
+ * slice is recorded in `ranges` as [start, end).
+ */
+export function sparseBlob(parts, size) {
+    const ranges = [];
+    function bytesOf(start, end) {
+        const out = new Uint8Array(end - start);
+        for (const part of parts) {
+            const from = Math.max(start, part.offset);
+            const to = Math.min(end, part.offset + part.bytes.length);
+            if (from < to) out.set(part.bytes.subarray(from - part.offset, to - part.offset), from - start);
+        }
+        return out;
+    }
+    return {
+        size,
+        ranges,
+        slice(start, end) {
+            ranges.push([start, end]);
+            return { arrayBuffer: async () => bytesOf(start, end).buffer };
+        },
+    };
+}
+
+/** A recording blob over real bytes. */
+export function recordingBlob(bytes) {
+    return sparseBlob([{ offset: 0, bytes }], bytes.length);
+}
+
+/** Every slice resolves after `delayMs`. */
+export function slowBlob(bytes, delayMs) {
+    return {
+        size: bytes.length,
+        slice(start, end) {
+            return {
+                arrayBuffer: () => new Promise((resolve) => setTimeout(() => resolve(bytes.slice(start, end).buffer), delayMs)),
+            };
+        },
+    };
+}
+
+/** Slices fail (as a browser does when the file vanished). */
+export function failingBlob(size) {
+    return {
+        size,
+        slice() {
+            return { arrayBuffer: () => Promise.reject(new Error("NotReadableError")) };
+        },
+    };
+}
+
+/** Reports `size` but every slice beyond `actual` bytes comes back short. */
+export function shrunkBlob(bytes, size) {
+    return {
+        size,
+        slice(start, end) {
+            return { arrayBuffer: async () => bytes.slice(start, Math.min(end, bytes.length)).buffer };
+        },
+    };
+}
+
+/** An A1111 `parameters` record with one settings footer carrying a batch index. */
+export function parametersText(positive, batchIndex, { seedField = "Seed: 1" } = {}) {
+    return `${positive}\nNegative prompt: blurry\nSteps: 20, Sampler: Euler, Batch index: ${batchIndex}, ${seedField}`;
+}

tests/js/helpers/video-bytes.mjs

@@ -0,0 +1,277 @@
+// Byte-exact ISO-BMFF (MP4/MOV/M4V) and Matroska (MKV/WebM) builders for the local
+// video tag reader tests, laid out as ffmpeg's muxers write them (see the dumps in
+// tests/fixtures/local-inspect). The File-like blobs come from png-bytes.mjs.
+import { concat, u32, utf8, latin1 } from "./png-bytes.mjs";
+
+export { concat, u32, utf8, latin1 };
+export { sparseBlob, recordingBlob, slowBlob, failingBlob, blobOf } from "./png-bytes.mjs";
+
+const toBytes = (part) => (typeof part === "string" ? utf8(part) : Buffer.from(part));
+
+function u64(n) {
+    const out = new Uint8Array(8);
+    new DataView(out.buffer).setBigUint64(0, BigInt(n));
+    return out;
+}
+
+// ---------------------------------------------------------------- ISO-BMFF
+
+/** A box: 32-bit size, 4-byte type (string or raw bytes), payload. `size` overrides the field. */
+export function box(type, ...parts) {
+    const options = parts.length && parts[parts.length - 1] && parts[parts.length - 1].__options ? parts.pop() : {};
+    const payload = concat(...parts.map(toBytes));
+    const typeBytes = typeof type === "string" ? latin1(type) : Buffer.from(type);
+    const total = 8 + payload.length;
+    if (options.largesize) return concat(u32(1), typeBytes, u64(options.largesize === true ? total + 8 : options.largesize), payload);
+    return concat(u32(options.size ?? total), typeBytes, payload);
+}
+
+/** Marker for `box` options: box('ftyp', ..., opts({size: 3})). */
+export function opts(options) {
+    return { __options: true, ...options };
+}
+
+export function fullBox(type, ...parts) {
+    return box(type, u32(0), ...parts);
+}
+
+export function ftyp(brand = "isom", compatible = "isomiso2avc1mp41") {
+    return box("ftyp", latin1(brand), u32(512), latin1(compatible));
+}
+
+export function mdat(payload = [0, 0, 0, 0]) {
+    return box("mdat", payload);
+}
+
+export const moov = (...children) => box("moov", ...children);
+export const udta = (...children) => box("udta", ...children);
+export const free = (count = 0) => box("free", new Uint8Array(count));
+
+/** `meta` as ffmpeg writes it: a FullBox (version/flags) whose first child is hdlr. */
+export function meta(...children) {
+    return fullBox("meta", ...children);
+}
+
+export function hdlr(handler = "mdta") {
+    return fullBox("hdlr", u32(0), latin1(handler), new Uint8Array(12), [0]);
+}
+
+/** `keys`: one mdta entry per name (an entry may be [name, namespace]); `count` overrides the count field. */
+export function keys(names, { count } = {}) {
+    const entries = names.map((entry) => {
+        const [name, namespace] = Array.isArray(entry) ? entry : [entry, "mdta"];
+        const bytes = toBytes(name);
+        return concat(u32(8 + bytes.length), latin1(namespace), bytes);
+    });
+    return fullBox("keys", u32(count ?? names.length), ...entries);
+}
+
+export const ilst = (...items) => box("ilst", ...items);
+
+/** A `data` atom: type indicator (1 = UTF-8), locale, payload. */
+export function data(payload, { type = 1, locale = 0, size } = {}) {
+    return box("data", u32(type), u32(locale), toBytes(payload), ...(size === undefined ? [] : [opts({ size })]));
+}
+
+/** An ilst item addressed by its 1-based `keys` index. */
+export function item(index, ...children) {
+    return box(u32(index), ...children);
+}
+
+/** An ilst item addressed by a legacy fourcc such as "©cmt" (Latin-1 bytes). */
+export function legacyItem(fourcc, ...children) {
+    return box(latin1(fourcc), ...children);
+}
+
+/** `moov/udta/meta` carrying per-key mdta tags, as core SaveVideo writes through ffmpeg. */
+export function mdtaUdta(tags) {
+    const names = Object.keys(tags);
+    return udta(meta(hdlr("mdta"), keys(names), ilst(...names.map((name, index) => item(index + 1, data(tags[name]))))));
+}
+
+/** `moov/udta/meta` with legacy ilst atoms: {comment, description} → ©cmt / desc. */
+export function legacyUdta({ comment, description, tool = "Lavf" } = {}) {
+    const items = [legacyItem("©too", data(tool))];
+    if (comment !== undefined) items.push(legacyItem("©cmt", data(comment)));
+    if (description !== undefined) items.push(legacyItem("desc", data(description)));
+    return udta(meta(hdlr("mdir"), ilst(...items)));
+}
+
+/** A whole MP4: ftyp, (moov first when faststart), mdat, moov with the given udta. */
+export function mp4(udtaBox, { brand = "isom", faststart = false, extraMoov = [], payload } = {}) {
+    const movie = moov(...extraMoov, udtaBox);
+    const media = mdat(payload);
+    return faststart ? concat(ftyp(brand), movie, media) : concat(ftyp(brand), free(), media, movie);
+}
+
+export function perKeyMp4(tags, options) {
+    return mp4(mdtaUdta(tags), options);
+}
+
+/** Top-level boxes of `bytes` as [{type, start, end}]. */
+export function topLevelBoxes(bytes) {
+    const out = [];
+    let offset = 0;
+    while (offset + 8 <= bytes.length) {
+        const view = new DataView(bytes.buffer, bytes.byteOffset + offset);
+        let size = view.getUint32(0);
+        if (size === 1) size = Number(view.getBigUint64(8));
+        if (size === 0) size = bytes.length - offset;
+        out.push({ type: latin1Of(bytes.subarray(offset + 4, offset + 8)), start: offset, end: offset + size });
+        offset += size;
+    }
+    return out;
+}
+
+function latin1Of(bytes) {
+    return Buffer.from(bytes).toString("latin1");
+}
+
+// ---------------------------------------------------------------- Matroska
+
+export const ID = {
+    EBML: 0x1a45dfa3,
+    EBMLVersion: 0x4286,
+    DocType: 0x4282,
+    DocTypeVersion: 0x4287,
+    Segment: 0x18538067,
+    SeekHead: 0x114d9b74,
+    Seek: 0x4dbb,
+    SeekID: 0x53ab,
+    SeekPosition: 0x53ac,
+    Void: 0xec,
+    CRC32: 0xbf,
+    Info: 0x1549a966,
+    Tracks: 0x1654ae6b,
+    Tags: 0x1254c367,
+    Tag: 0x7373,
+    Targets: 0x63c0,
+    TargetTypeValue: 0x68ca,
+    TagTrackUID: 0x63c5,
+    SimpleTag: 0x67c8,
+    TagName: 0x45a3,
+    TagString: 0x4487,
+    TagLanguage: 0x447a,
+    TagDefault: 0x4484,
+    TagBinary: 0x4485,
+    Cluster: 0x1f43b675,
+    Timestamp: 0xe7,
+    Cues: 0x1c53bb6b,
+};
+
+/** The ID's bytes (EBML IDs carry their own length marker). */
+export function idBytes(id) {
+    const out = [];
+    let value = id;
+    while (value > 0) {
+        out.unshift(value & 0xff);
+        value = Math.floor(value / 256);
+    }
+    return Uint8Array.from(out);
+}
+
+/** A size VINT: minimal width unless `width` is given; `unknown` writes all ones. */
+export function vint(value, { width, unknown = false } = {}) {
+    let length = width;
+    if (!length) {
+        length = 1;
+        while (length < 8 && value >= 2 ** (7 * length) - 1) length += 1;
+    }
+    const out = new Uint8Array(length);
+    if (unknown) {
+        out.fill(0xff);
+        out[0] = 0xff >> (length - 1);
+        return out;
+    }
+    let rest = BigInt(value);
+    for (let index = length - 1; index >= 0; index -= 1) {
+        out[index] = Number(rest & 0xffn);
+        rest >>= 8n;
+    }
+    out[0] |= 0x80 >> (length - 1);
+    return out;
+}
+
+/** An element: ID, size VINT, payload. `size`/`width`/`unknown` shape the size field. */
+export function el(id, ...parts) {
+    const options = parts.length && parts[parts.length - 1] && parts[parts.length - 1].__options ? parts.pop() : {};
+    const payload = concat(...parts.map(toBytes));
+    const size = vint(options.size ?? payload.length, { width: options.width, unknown: options.unknown });
+    return concat(idBytes(id), size, payload);
+}
+
+export function uint(value, width) {
+    const out = [];
+    let rest = BigInt(value);
+    do {
+        out.unshift(Number(rest & 0xffn));
+        rest >>= 8n;
+    } while (rest > 0n);
+    while (width && out.length < width) out.unshift(0);
+    return Uint8Array.from(out);
+}
+
+export function ebmlHeader(docType = "matroska") {
+    return el(ID.EBML, el(ID.EBMLVersion, [1]), el(ID.DocType, docType), el(ID.DocTypeVersion, [4]));
+}
+
+/** Segment with an 8-byte size field as ffmpeg writes it; `unknown` makes it unknown-sized. */
+export function segment(children, { unknown = false } = {}) {
+    return el(ID.Segment, ...children, opts({ width: 8, unknown }));
+}
+
+/** SeekHead entries: [[id, position]] with positions relative to the Segment payload. */
+export function seekHead(entries) {
+    return el(ID.SeekHead, ...entries.map(([id, position]) => el(ID.Seek, el(ID.SeekID, idBytes(id)), el(ID.SeekPosition, uint(position)))));
+}
+
+export function voidEl(count = 4) {
+    return el(ID.Void, new Uint8Array(count));
+}
+
+/** A SimpleTag with TagName and TagString (`value` undefined omits the string). */
+export function simpleTag(name, value, { lang, def, binary, nested = [], extra = [] } = {}) {
+    const children = [el(ID.TagName, name)];
+    if (lang !== undefined) children.push(el(ID.TagLanguage, lang));
+    if (def !== undefined) children.push(el(ID.TagDefault, [def]));
+    if (value !== undefined) children.push(el(ID.TagString, value));
+    if (binary !== undefined) children.push(el(ID.TagBinary, binary));
+    return el(ID.SimpleTag, ...children, ...nested, ...extra);
+}
+
+/** A Tag whose Targets carry the given track UID (none → global). */
+export function tag(simpleTags, { trackUid, targetType } = {}) {
+    const targets = [];
+    if (targetType !== undefined) targets.push(el(ID.TargetTypeValue, uint(targetType)));
+    if (trackUid !== undefined) targets.push(el(ID.TagTrackUID, uint(trackUid, 8)));
+    return el(ID.Tag, el(ID.Targets, ...targets), ...simpleTags);
+}
+
+export function tagsEl(...tagElements) {
+    return el(ID.Tags, ...tagElements);
+}
+
+/** A Cluster of `size` payload bytes (zeros); `unknown` makes its size unknown. */
+export function cluster(size = 8, { unknown = false, payload } = {}) {
+    return el(ID.Cluster, payload ?? new Uint8Array(size), opts({ unknown }));
+}
+
+/** A Cluster header only, declaring `size` payload bytes (for sparse layouts). */
+export function clusterHeader(size) {
+    return concat(idBytes(ID.Cluster), vint(size, { width: 8 }));
+}
+
+/** A whole Matroska file: EBML header + Segment(children). */
+export function mkv(children, { docType = "matroska", unknown = false } = {}) {
+    return concat(ebmlHeader(docType), segment(children, { unknown }));
+}
+
+/** A global Tags element from {name: value} (ffmpeg upper-cases names at mux time). */
+export function globalTags(tags, options) {
+    return tagsEl(tag(Object.entries(tags).map(([name, value]) => simpleTag(name, value)), options));
+}
+
+/** A Matroska file whose global Tags hold {name: value}, ffmpeg layout. */
+export function perKeyMkv(tags, { docType = "matroska" } = {}) {
+    return mkv([el(ID.Info, new Uint8Array(4)), globalTags(tags), cluster()], { docType });
+}

tests/js/png-local-reader.test.js

@@ -0,0 +1,549 @@
+// Run with: node --test tests/js/
+// The browser PNG File reader is a port of utils/prompt_inspector.py's bounded
+// reader (_png_metadata, _PngRecords, _png_text, _inflate, _batch_normalized). Each
+// rule below names the Python branch it mirrors; tests/test_local_inspect_parity.py
+// proves the two agree on the committed parity corpus.
+const test = require("node:test");
+const assert = require("node:assert/strict");
+const fs = require("node:fs");
+const path = require("node:path");
+const zlib = require("node:zlib");
+const { classifyPrefix, readPngMetadataFile, readImageHeader } = require("../../web/js/png-metadata.js");
+const bytes = require("./helpers/png-bytes.mjs");
+
+const { png, raw, ihdr, idat, iend, tEXt, zTXt, iTXt, chunk, blobOf, sparseBlob, recordingBlob, slowBlob, failingBlob, shrunkBlob, concat, u32, latin1, parametersText } = bytes;
+
+const MiB = 1024 * 1024;
+const ROLE_BUDGET = 2 * MiB;
+const WORKFLOW_BUDGET = 4 * MiB;
+const PNG_READ_CEILING = 19794389;
+const FIXTURES = path.join(__dirname, "..", "fixtures", "local-inspect");
+
+async function read(fileBytes, options) {
+    return readPngMetadataFile(blobOf(fileBytes), options);
+}
+
+function fill(byte, count) {
+    return Buffer.alloc(count, byte);
+}
+
+test("contract: the reader, classifier and header reader are exported", () => {
+    assert.equal(typeof readPngMetadataFile, "function");
+    assert.equal(typeof classifyPrefix, "function");
+    assert.equal(typeof readImageHeader, "function");
+});
+
+test("reads the four retained records and the IHDR size from a core SaveImage PNG", async () => {
+    const result = await read(raw(
+        ihdr(640, 480),
+        idat(),
+        tEXt("prompt", '{"1":{"class_type":"KSampler"}}'),
+        iTXt("workflow", '{"nodes":[]}'),
+        tEXt("parameters", "cat\nSteps: 20, Seed: 1"),
+        zTXt("comfyassets_prompt_pairs", '{"pairs":[]}'),
+        tEXt("Software", "ComfyUI"),
+        iend(),
+    ));
+    assert.deepEqual(result, {
+        status: "complete",
+        records: {
+            prompt: '{"1":{"class_type":"KSampler"}}',
+            workflow: '{"nodes":[]}',
+            parameters: "cat\nSteps: 20, Seed: 1",
+            comfyassets_prompt_pairs: '{"pairs":[]}',
+        },
+        limited: [],
+        width: 640,
+        height: 480,
+    });
+});
+
+test("local PNG skips sparse pixels and reads post-IDAT text", async () => {
+    // Two 1.5 GiB IDAT chunks (each under the 0x7FFFFFFF chunk ceiling) with the
+    // prompt after them: nothing of the 3 GiB is read or allocated.
+    const half = 1.5 * 1024 * 1024 * 1024;
+    const head = concat(bytes.SIGNATURE, ihdr(8192, 8192), u32(half), latin1("IDAT"));
+    const idat1 = { start: bytes.SIGNATURE.length + ihdr().length, end: bytes.SIGNATURE.length + ihdr().length + 8 + half + 4 };
+    const idat2 = { start: idat1.end, end: idat1.end + 8 + half + 4 };
+    const between = concat(u32(0), u32(half), latin1("IDAT"));
+    const tail = concat(u32(0), tEXt("prompt", '{"3":{"class_type":"KSampler"}}'), iend());
+    const size = idat2.end - 4 + tail.length;
+    const file = sparseBlob([
+        { offset: 0, bytes: head },
+        { offset: idat1.end - 4, bytes: between },
+        { offset: idat2.end - 4, bytes: tail },
+    ], size);
+
+    const result = await readPngMetadataFile(file);
+
+    assert.equal(result.status, "complete");
+    assert.deepEqual(result.records, { prompt: '{"3":{"class_type":"KSampler"}}' });
+    assert.deepEqual([result.width, result.height], [8192, 8192]);
+    const insideIdat = file.ranges.filter(([start, end]) => [idat1, idat2].some((region) => start >= region.start && end <= region.end));
+    assert.ok(insideIdat.length >= 2, "the IDAT headers were read");
+    for (const [start, end] of insideIdat) assert.ok(end - start <= 12, `range inside IDAT longer than 12: [${start}, ${end})`);
+    const total = file.ranges.reduce((sum, [start, end]) => sum + (end - start), 0);
+    assert.ok(total <= PNG_READ_CEILING, `read ${total} bytes`);
+});
+
+test("serves the retained prefix without slicing it again", async () => {
+    const whole = png(tEXt("prompt", "p"));
+    const file = recordingBlob(whole);
+    const result = await readPngMetadataFile(file, { prefix: whole.subarray(0, 32) });
+    assert.equal(result.status, "complete");
+    assert.ok(file.ranges.every(([start]) => start >= 32), JSON.stringify(file.ranges));
+    assert.deepEqual(result.records, { prompt: "p" });
+});
+
+test("unknown text keys are skipped by length without reading their body", async () => {
+    const body = fill(0x41, 1000);
+    const whole = png(tEXt("Software", body), tEXt("prompt", "p"));
+    const softwareStart = bytes.SIGNATURE.length + ihdr().length + idat().length;
+    const bodyStart = softwareStart + 8 + "Software".length + 1;
+    const bodyEnd = bodyStart + body.length + 4;
+    const file = recordingBlob(whole);
+    const result = await readPngMetadataFile(file);
+    assert.equal(result.status, "complete");
+    const insideBody = file.ranges.filter(([start, end]) => start < bodyEnd && end > softwareStart + 8 + 80);
+    assert.deepEqual(insideBody, [], "the body past the 80-byte keyword prefix was sliced");
+});
+
+const WF_OVER = WORKFLOW_BUDGET + 1;
+const bigWorkflow = () => tEXt("workflow", fill(0x7b, WF_OVER - "workflow".length - 1));
+
+// One row per Python branch (prompt_inspector.py:91-260).
+const RULES = [
+    { name: "bad IHDR CRC", file: () => raw(ihdr(1, 1, { crc: 1 }), idat(), iend()), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "bad text CRC", file: () => png(tEXt("prompt", "p", { crc: 1 })), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "bad IEND CRC", file: () => raw(ihdr(), idat(), iend({ crc: 1 })), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "first chunk not IHDR", file: () => raw(idat(), ihdr(), iend()), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "IHDR not 13 bytes", file: () => raw(chunk("IHDR", concat(u32(1), u32(1), [8, 6, 0, 0])), idat(), iend()), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "second IHDR", file: () => raw(ihdr(), idat(), ihdr(), iend()), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "zero width", file: () => raw(ihdr(0, 1), idat(), iend()), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "zero height", file: () => raw(ihdr(1, 0), idat(), iend()), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "length 0x80000000", file: () => raw(ihdr(), idat(), chunk("tEXt", latin1("k\0v"), { declaredLength: 0x80000000 }), iend()), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "end past size", file: () => raw(ihdr(), idat(), chunk("tEXt", latin1("k\0v"), { declaredLength: 100 }), iend()), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "file ends inside a chunk header", file: () => raw(ihdr(), idat(), u32(5)), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "file ends before any chunk", file: () => raw(), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "too short for a signature", file: () => bytes.SIGNATURE.subarray(0, 5), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "not a PNG", file: () => latin1("GIF89a................................."), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "IEND without IDAT", file: () => raw(ihdr(), tEXt("prompt", "p"), iend()), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "nonzero IEND length", file: () => raw(ihdr(), idat(), chunk("IEND", [1])), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "duplicate key with a different length", file: () => png(tEXt("prompt", "one"), tEXt("prompt", "other")), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "duplicate key, same length, different bytes", file: () => png(tEXt("prompt", "one"), tEXt("prompt", "two")), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "duplicate key, same bytes, different chunk type", file: () => png(tEXt("prompt", "one"), chunk("zTXt", latin1("prompt\0one"))), expect: { status: "unreadable", reason: "invalid-png" } },
+    {
+        name: "byte-identical duplicate kept once",
+        file: () => png(tEXt("prompt", "same"), tEXt("prompt", "same"), tEXt("prompt", "same")),
+        expect: { status: "complete", records: { prompt: "same" }, limited: [], width: 1, height: 1 },
+    },
+    {
+        name: "parameters copies differing only in Batch index keep the last",
+        file: () => png(tEXt("parameters", parametersText("cat", 0)), tEXt("parameters", parametersText("cat", 1)), tEXt("parameters", parametersText("cat", 12))),
+        expect: { status: "complete", records: { parameters: parametersText("cat", 12) }, limited: [], width: 1, height: 1 },
+    },
+    { name: "parameters copies with any other difference", file: () => png(tEXt("parameters", parametersText("cat", 0)), tEXt("parameters", parametersText("dog", 1))), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "parameters copies without a settings footer", file: () => png(tEXt("parameters", "cat"), tEXt("parameters", "dog")), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "parameters copies with two footers", file: () => png(tEXt("parameters", `${parametersText("cat", 0)}\nSteps: 1, x`), tEXt("parameters", `${parametersText("cat", 1)}\nSteps: 1, x`)), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "parameters copies with two batch indexes", file: () => png(tEXt("parameters", parametersText("cat", 0, { seedField: "Batch index: 0" })), tEXt("parameters", parametersText("cat", 1, { seedField: "Batch index: 1" }))), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "parameters copies whose batch index is not a whole settings field", file: () => png(tEXt("parameters", "Steps: 20, Batch index: 0 trailing"), tEXt("parameters", "Steps: 20, Batch index: 1 trailing")), expect: { status: "unreadable", reason: "invalid-png" } },
+    {
+        name: "workflow 4 MiB + 1 encoded is limited, other records intact",
+        file: () => png(tEXt("prompt", "p"), bigWorkflow(), tEXt("parameters", "cat\nSteps: 1, x")),
+        expect: { status: "complete", records: { prompt: "p", parameters: "cat\nSteps: 1, x" }, limited: ["workflow"], width: 1, height: 1 },
+    },
+    {
+        name: "a duplicate of a limited workflow is skipped",
+        file: () => png(bigWorkflow(), bigWorkflow(), tEXt("prompt", "p")),
+        expect: { status: "complete", records: { prompt: "p" }, limited: ["workflow"], width: 1, height: 1 },
+    },
+    {
+        name: "workflow over 4 MiB decoded (zTXt) is limited",
+        file: () => png(zTXt("workflow", fill(0x7b, WORKFLOW_BUDGET + 1), { level: 9 }), tEXt("prompt", "p")),
+        expect: { status: "complete", records: { prompt: "p" }, limited: ["workflow"], width: 1, height: 1 },
+    },
+    {
+        name: "workflow at exactly 4 MiB decoded is kept",
+        file: () => png(zTXt("workflow", fill(0x7b, WORKFLOW_BUDGET), { level: 9 })),
+        check: (result) => {
+            assert.equal(result.status, "complete");
+            assert.equal(result.records.workflow.length, WORKFLOW_BUDGET);
+            assert.deepEqual(result.limited, []);
+        },
+    },
+    {
+        name: "workflow under 4 MiB inflated but over 4 MiB as UTF-8 is limited",
+        file: () => png(zTXt("workflow", fill(0xe9, WORKFLOW_BUDGET - 10), { level: 9 }), tEXt("prompt", "p")),
+        expect: { status: "complete", records: { prompt: "p" }, limited: ["workflow"], width: 1, height: 1 },
+    },
+    { name: "roles 2 MiB + 1 encoded", file: () => png(tEXt("prompt", fill(0x41, ROLE_BUDGET + 1 - "prompt".length - 1))), expect: { status: "unreadable", reason: "metadata-limit" } },
+    { name: "roles summed over 2 MiB encoded", file: () => png(tEXt("prompt", fill(0x41, 1.5 * MiB)), tEXt("parameters", fill(0x41, 0.5 * MiB))), expect: { status: "unreadable", reason: "metadata-limit" } },
+    { name: "roles 2 MiB + 1 decoded via zTXt", file: () => png(zTXt("prompt", fill(0x41, ROLE_BUDGET + 1), { level: 9 })), expect: { status: "unreadable", reason: "metadata-limit" } },
+    { name: "roles over 2 MiB as UTF-8 from Latin-1 high bytes", file: () => png(tEXt("prompt", fill(0xe9, MiB + 1))), expect: { status: "unreadable", reason: "metadata-limit" } },
+    {
+        name: "roles share the decoded budget across records",
+        file: () => png(zTXt("prompt", fill(0x41, 1.5 * MiB), { level: 9 }), zTXt("parameters", fill(0x41, 0.5 * MiB + 1), { level: 9 })),
+        expect: { status: "unreadable", reason: "metadata-limit" },
+    },
+    {
+        name: "roles at exactly 2 MiB decoded are kept",
+        file: () => png(zTXt("prompt", fill(0x41, 1.5 * MiB), { level: 9 }), zTXt("parameters", fill(0x41, 0.5 * MiB), { level: 9 })),
+        check: (result) => {
+            assert.equal(result.status, "complete");
+            assert.equal(result.records.prompt.length + result.records.parameters.length, ROLE_BUDGET);
+        },
+    },
+    {
+        name: "duplicate workflow bytes capped at 8 MiB: the workflow becomes limited",
+        file: () => {
+            const workflow = tEXt("workflow", fill(0x7b, 3 * MiB));
+            return png(workflow, workflow, workflow, workflow, tEXt("prompt", "p"));
+        },
+        expect: { status: "complete", records: { prompt: "p" }, limited: ["workflow"], width: 1, height: 1 },
+    },
+    {
+        name: "duplicate workflow bytes within 8 MiB keep the workflow",
+        file: () => {
+            const workflow = tEXt("workflow", fill(0x7b, 3 * MiB));
+            return png(workflow, workflow, workflow);
+        },
+        check: (result) => {
+            assert.equal(result.status, "complete");
+            assert.equal(result.records.workflow.length, 3 * MiB);
+            assert.deepEqual(result.limited, []);
+        },
+    },
+    {
+        name: "duplicate role bytes capped at 4 MiB",
+        file: () => {
+            const prompt = tEXt("prompt", fill(0x41, 1.5 * MiB));
+            return png(prompt, prompt, prompt, prompt);
+        },
+        expect: { status: "unreadable", reason: "metadata-limit" },
+    },
+    {
+        name: "duplicate decoded parameters capped at 4 MiB",
+        file: () => {
+            const copy = (index) => zTXt("parameters", parametersText(fill(0x41, 1.5 * MiB).toString("latin1"), index), { level: 9 });
+            return png(copy(0), copy(1), copy(2), copy(3));
+        },
+        expect: { status: "unreadable", reason: "metadata-limit" },
+    },
+    {
+        name: "a longer batch copy draws the growth from the role budget",
+        file: () => {
+            const base = parametersText("", 9);
+            const positive = fill(0x41, ROLE_BUDGET - base.length - 1).toString("latin1");
+            return png(zTXt("parameters", parametersText(positive, 9), { level: 9 }), zTXt("parameters", parametersText(positive, 100), { level: 9 }));
+        },
+        expect: { status: "unreadable", reason: "metadata-limit" },
+    },
+    {
+        name: "a longer batch copy within the role budget is kept",
+        file: () => {
+            const base = parametersText("", 9);
+            const positive = fill(0x41, ROLE_BUDGET - base.length - 2).toString("latin1");
+            return png(zTXt("parameters", parametersText(positive, 9), { level: 9 }), zTXt("parameters", parametersText(positive, 10), { level: 9 }));
+        },
+        check: (result) => {
+            assert.equal(result.status, "complete");
+            assert.ok(result.records.parameters.endsWith("Batch index: 10, Seed: 1"));
+        },
+    },
+    {
+        name: "a batch copy inflating past the role budget",
+        file: () => png(tEXt("parameters", parametersText("cat", 0)), zTXt("parameters", parametersText(fill(0x41, ROLE_BUDGET).toString("latin1"), 1), { level: 9 })),
+        expect: { status: "unreadable", reason: "metadata-limit" },
+    },
+    {
+        name: "10,000 chunks without IEND",
+        file: () => concat(bytes.SIGNATURE, ihdr(), idat(), ...Array.from({ length: 9999 }, () => chunk("xxXx", [])), iend()),
+        expect: { status: "unreadable", reason: "metadata-limit" },
+    },
+    {
+        name: "IEND as the 10,000th chunk is read",
+        file: () => concat(bytes.SIGNATURE, ihdr(), idat(), ...Array.from({ length: 9997 }, () => chunk("xxXx", [])), iend()),
+        expect: { status: "complete", records: {}, limited: [], width: 1, height: 1 },
+    },
+    { name: "zTXt method not 0", file: () => png(zTXt("prompt", "p", { method: 1 })), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "zTXt without a method byte", file: () => png(chunk("zTXt", latin1("prompt\0"))), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "iTXt flag 2", file: () => png(iTXt("prompt", "p", { flag: 2 })), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "iTXt compressed with method 1", file: () => png(iTXt("prompt", "p", { flag: 1, method: 1 })), expect: { status: "unreadable", reason: "invalid-png" } },
+    {
+        name: "iTXt uncompressed ignores the method byte",
+        file: () => png(iTXt("prompt", "p", { flag: 0, method: 7 })),
+        expect: { status: "complete", records: { prompt: "p" }, limited: [], width: 1, height: 1 },
+    },
+    { name: "iTXt shorter than flag and method", file: () => png(chunk("iTXt", latin1("prompt\0\0"))), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "iTXt missing the translated keyword separator", file: () => png(chunk("iTXt", concat(latin1("prompt\0"), [0, 0], latin1("en\0text")))), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "iTXt non-ASCII language tag", file: () => png(iTXt("prompt", "p", { lang: [0xc3, 0xa9] })), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "iTXt invalid UTF-8 translated keyword", file: () => png(iTXt("prompt", "p", { translated: [0xff] })), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "iTXt invalid UTF-8 text", file: () => png(iTXt("prompt", [0xc3, 0x28])), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "iTXt encoded surrogate", file: () => png(iTXt("prompt", [0xed, 0xa0, 0x80])), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "NUL in tEXt text", file: () => png(tEXt("prompt", "a\0b")), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "NUL in inflated zTXt text", file: () => png(zTXt("prompt", "a\0b")), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "NUL in iTXt text", file: () => png(iTXt("prompt", "a\0b")), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "keyword of 0 bytes", file: () => png(chunk("tEXt", latin1("\0v"))), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "keyword of 80 bytes", file: () => png(tEXt("k".repeat(80), "v")), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "keyword without a NUL", file: () => png(chunk("tEXt", latin1("nonul"))), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "empty text chunk", file: () => png(chunk("tEXt", [])), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "malformed keyword on an unknown key still fails", file: () => png(tEXt("prompt", "p"), chunk("tEXt", latin1("x".repeat(100)))), expect: { status: "unreadable", reason: "invalid-png" } },
+    {
+        name: "keyword of 79 bytes is read",
+        file: () => png(tEXt("k".repeat(79), "v"), tEXt("prompt", "p")),
+        expect: { status: "complete", records: { prompt: "p" }, limited: [], width: 1, height: 1 },
+    },
+    {
+        name: "Latin-1 0xE9 is an e-acute in tEXt and zTXt",
+        file: () => png(tEXt("prompt", [0xe9]), zTXt("parameters", [0xe9, 0x0a, 0x53, 0x74, 0x65, 0x70, 0x73, 0x3a, 0x20, 0x31])),
+        expect: { status: "complete", records: { prompt: "é", parameters: "é\nSteps: 1" }, limited: [], width: 1, height: 1 },
+    },
+    {
+        name: "iTXt BOM is preserved, compressed and not",
+        file: () => png(iTXt("prompt", "{}"), iTXt("workflow", "[]", { flag: 1, lang: "en", translated: "wörkflow" })),
+        expect: { status: "complete", records: { prompt: "{}", workflow: "[]" }, limited: [], width: 1, height: 1 },
+    },
+    { name: "truncated zlib data", file: () => png(zTXt("prompt", "x".repeat(500), { compressed: zlib.deflateSync(latin1("x".repeat(500))).subarray(0, 8) })), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "trailing bytes after the zlib stream", file: () => png(zTXt("prompt", "p", { compressed: concat(zlib.deflateSync(latin1("p")), [1, 2, 3]) })), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "empty zlib data", file: () => png(zTXt("prompt", "p", { compressed: [] })), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "garbage zlib data", file: () => png(zTXt("prompt", "p", { compressed: [1, 2, 3, 4, 5, 6, 7, 8] })), expect: { status: "unreadable", reason: "invalid-png" } },
+    { name: "truncated compressed iTXt", file: () => png(iTXt("prompt", "p", { flag: 1, compressed: zlib.deflateSync(latin1("x".repeat(500))).subarray(0, 8) })), expect: { status: "unreadable", reason: "invalid-png" } },
+    {
+        name: "chunks after IEND are ignored",
+        file: () => concat(png(tEXt("prompt", "p")), tEXt("workflow", "w")),
+        expect: { status: "complete", records: { prompt: "p" }, limited: [], width: 1, height: 1 },
+    },
+    {
+        name: "non-ASCII chunk types are skipped",
+        file: () => png(chunk("\xff\xfe\x80\x81", [1, 2, 3]), tEXt("prompt", "p")),
+        expect: { status: "complete", records: { prompt: "p" }, limited: [], width: 1, height: 1 },
+    },
+];
+
+for (const rule of RULES) {
+    test(`server rules, one by one: ${rule.name}`, async () => {
+        const result = await read(rule.file());
+        if (rule.check) rule.check(result);
+        else assert.deepEqual(result, rule.expect);
+    });
+}
+
+test("decompress: undefined with zTXt is decompression-unavailable and yields no records", async () => {
+    const result = await read(png(tEXt("prompt", "p"), zTXt("workflow", "w")), { decompress: undefined });
+    assert.deepEqual(result, { status: "unreadable", reason: "decompression-unavailable" });
+    const plain = await read(png(tEXt("prompt", "p"), zTXt("Comment", "skipped")), { decompress: undefined });
+    assert.deepEqual(plain.records, { prompt: "p" });
+});
+
+function instrumentedStream(transform) {
+    const stats = { writes: [], inFlight: 0, maxInFlight: 0 };
+    return {
+        stats,
+        Stream: class {
+            constructor() {
+                const stream = new TransformStream({
+                    async transform(chunk, controller) {
+                        stats.writes.push(chunk.byteLength);
+                        stats.inFlight += 1;
+                        stats.maxInFlight = Math.max(stats.maxInFlight, stats.inFlight);
+                        await transform(chunk, controller);
+                        stats.inFlight -= 1;
+                    },
+                });
+                this.writable = stream.writable;
+                this.readable = stream.readable;
+            }
+        },
+    };
+}
+
+test("inflation keeps one <=256-byte write in flight", async () => {
+    const body = fill(0x41, 3000);
+    const { stats, Stream } = instrumentedStream(async (chunk, controller) => {
+        await new Promise((resolve) => setTimeout(resolve, 1));
+        controller.enqueue(chunk);
+    });
+    // The fake "inflates" by passing bytes through, so the stored body is the text.
+    const result = await read(png(chunk("zTXt", concat(latin1("prompt\0"), [0], body))), { decompress: Stream });
+    assert.equal(result.status, "complete");
+    assert.equal(result.records.prompt, body.toString("latin1"));
+    assert.ok(stats.writes.every((size) => size <= 256), JSON.stringify(stats.writes));
+    assert.equal(stats.writes.reduce((sum, size) => sum + size, 0), body.length);
+    assert.equal(stats.maxInFlight, 1);
+});
+
+test("inflation settles once on abort with an AbortError and no unhandled rejection", async () => {
+    const unhandled = [];
+    const onUnhandled = (reason) => unhandled.push(reason);
+    process.on("unhandledRejection", onUnhandled);
+    try {
+        const { stats, Stream } = instrumentedStream(() => new Promise(() => {}));
+        const controller = new AbortController();
+        const pending = readPngMetadataFile(blobOf(png(zTXt("prompt", "p".repeat(1000)))), { decompress: Stream, signal: controller.signal });
+        let settled = 0;
+        pending.then(() => settled++, () => settled++);
+        await new Promise((resolve) => setTimeout(resolve, 5));
+        assert.equal(settled, 0, "the read waits on the stream");
+        controller.abort();
+        await assert.rejects(pending, (error) => error.name === "AbortError");
+        await new Promise((resolve) => setTimeout(resolve, 10));
+        assert.equal(settled, 1);
+        assert.equal(stats.maxInFlight, 1);
+        assert.deepEqual(unhandled, []);
+    } finally {
+        process.off("unhandledRejection", onUnhandled);
+    }
+});
+
+test("a stream error under a live signal is still invalid-png, not an abort", async () => {
+    const controller = new AbortController();
+    const result = await read(png(zTXt("prompt", "p", { compressed: [1, 2, 3, 4, 5, 6, 7, 8] })), { signal: controller.signal });
+    assert.deepEqual(result, { status: "unreadable", reason: "invalid-png" });
+    assert.equal(controller.signal.aborted, false);
+});
+
+test("an already aborted signal rejects before any slice", async () => {
+    const controller = new AbortController();
+    controller.abort();
+    const file = recordingBlob(png(tEXt("prompt", "p")));
+    await assert.rejects(readPngMetadataFile(file, { signal: controller.signal }), (error) => error.name === "AbortError");
+    assert.deepEqual(file.ranges, []);
+});
+
+test("aborting between slices rejects with an AbortError", async () => {
+    const controller = new AbortController();
+    const whole = png(tEXt("prompt", "p"));
+    let slices = 0;
+    const file = {
+        size: whole.length,
+        slice(start, end) {
+            return {
+                arrayBuffer: async () => {
+                    if (++slices === 3) controller.abort();
+                    return whole.slice(start, end).buffer;
+                },
+            };
+        },
+    };
+    await assert.rejects(readPngMetadataFile(file, { signal: controller.signal }), (error) => error.name === "AbortError");
+});
+
+test("aborting while inflating a real deflate stream rejects with an AbortError", async () => {
+    const controller = new AbortController();
+    let inflating = false;
+    class Tracked extends DecompressionStream {
+        constructor(format) {
+            super(format);
+            inflating = true;
+        }
+    }
+    // The slow slices make the pacer yield inside the inflater's feed loop; that
+    // yield aborts, so the real stream is torn down mid-inflation.
+    const yieldControl = async () => {
+        if (inflating) controller.abort();
+    };
+    const slow = slowBlob(png(zTXt("prompt", "p", { compressed: require("node:crypto").randomBytes(5000) })), 10);
+    await assert.rejects(readPngMetadataFile(slow, { signal: controller.signal, yieldControl, decompress: Tracked }), (error) => error.name === "AbortError");
+    assert.equal(inflating, true);
+});
+
+test("yields control about every 8 ms while the file answers slowly", async () => {
+    let yields = 0;
+    const yieldControl = async () => {
+        yields += 1;
+    };
+    const result = await readPngMetadataFile(slowBlob(png(tEXt("prompt", "p")), 10), { yieldControl });
+    assert.equal(result.status, "complete");
+    assert.ok(yields >= 1, `yielded ${yields} times`);
+});
+
+test("the default yield lets the event loop turn", async () => {
+    const result = await readPngMetadataFile(slowBlob(png(tEXt("prompt", "p")), 10));
+    assert.deepEqual(result.records, { prompt: "p" });
+});
+
+test("a slice that fails is a read failure", async () => {
+    assert.deepEqual(await readPngMetadataFile(failingBlob(100)), { status: "unreadable", reason: "read-failure" });
+    assert.deepEqual(await readPngMetadataFile(null), { status: "unreadable", reason: "read-failure" });
+});
+
+test("a file that shrank under the reader is a read failure", async () => {
+    const whole = png(tEXt("prompt", "p"));
+    assert.deepEqual(await readPngMetadataFile(shrunkBlob(whole.subarray(0, 40), whole.length)), { status: "unreadable", reason: "read-failure" });
+});
+
+test("classifier", () => {
+    assert.deepEqual(classifyPrefix(bytes.SIGNATURE), { media: "image", kind: "png" });
+    assert.deepEqual(classifyPrefix(Uint8Array.from([0xff, 0xd8, 0xff, 0xe0])), { media: "image", kind: "jpeg" });
+    assert.deepEqual(classifyPrefix(latin1("RIFF\0\0\0\0WEBPVP8 ")), { media: "image", kind: "webp" });
+    assert.deepEqual(classifyPrefix(latin1("\0\0\0\x18ftypisom\0\0\0\0")), { media: "video", kind: "iso-bmff" });
+    assert.deepEqual(classifyPrefix(latin1("\0\0\0\x14ftypqt  ")), { media: "video", kind: "iso-bmff" });
+    assert.deepEqual(classifyPrefix(Uint8Array.from([0x1a, 0x45, 0xdf, 0xa3, 0x01])), { media: "video", kind: "matroska" });
+    assert.deepEqual(classifyPrefix(latin1("RIFF\0\0\0\0AVI LIST")), { media: "video", kind: "avi" });
+    const asf = Uint8Array.from([0x30, 0x26, 0xb2, 0x75, 0x8e, 0x66, 0xcf, 0x11, 0xa6, 0xd9, 0x00, 0xaa, 0x00, 0x62, 0xce, 0x6c, 0, 0]);
+    assert.deepEqual(classifyPrefix(asf), { media: "video", kind: "asf" });
+    assert.deepEqual(classifyPrefix(asf.buffer), { media: "video", kind: "asf" });
+    for (const text of ["plain text file", "<!DOCTYPE html><html>", '<?xml version="1.0"?><svg', "#EXTM3U\n#EXT-X-VERSION:3", "RIFF\0\0\0\0WAVEfmt ", "\0\0\0\x18moov", ""]) {
+        assert.equal(classifyPrefix(latin1(text)), null, text);
+    }
+    assert.equal(classifyPrefix(bytes.SIGNATURE.subarray(0, 7)), null);
+    assert.equal(classifyPrefix(Uint8Array.from([0xff, 0xd8])), null);
+    assert.equal(classifyPrefix("not bytes"), null);
+    assert.equal(classifyPrefix(null), null);
+});
+
+test("image header: JPEG and WebP fixtures match Pillow's dimensions", async () => {
+    // tests/test_local_inspect_parity.py asserts the same two files against Pillow.
+    const jpeg = blobOf(fs.readFileSync(path.join(FIXTURES, "tiny.jpg")));
+    assert.deepEqual(await readImageHeader(jpeg, "jpeg"), { format: "JPEG", width: 48, height: 32 });
+    const webp = blobOf(fs.readFileSync(path.join(FIXTURES, "tiny.webp")));
+    assert.deepEqual(await readImageHeader(webp, "webp"), { format: "WEBP", width: 40, height: 24 });
+});
+
+test("image header: JPEG marker walk skips padding, junk, standalone markers and tables", async () => {
+    const segment = (marker, body) => concat([0xff, marker], u32(body.length + 2).subarray(2), body);
+    const sof = (marker, width, height) => segment(marker, concat([8], u32(height).subarray(2), u32(width).subarray(2), [3]));
+    const sized = concat([0xff, 0xd8], segment(0xe1, fill(0, 20)), [0xff, 0xff, 0xff, 0x00, 0x00, 0x11], [0xff, 0xd0], [0xff, 0x01], segment(0xc4, fill(0, 4)), segment(0xc8, []), segment(0xcc, [1]), sof(0xc2, 300, 200));
+    assert.deepEqual(await readImageHeader(blobOf(sized), "jpeg"), { format: "JPEG", width: 300, height: 200 });
+    for (const marker of [0xc0, 0xc1, 0xc3, 0xc5, 0xc6, 0xc7, 0xc9, 0xca, 0xcb, 0xcd, 0xce, 0xcf]) {
+        assert.deepEqual(await readImageHeader(blobOf(concat([0xff, 0xd8], sof(marker, 7, 9))), "jpeg"), { format: "JPEG", width: 7, height: 9 }, marker.toString(16));
+    }
+    const none = { format: "JPEG", width: null, height: null };
+    assert.deepEqual(await readImageHeader(blobOf(concat([0xff, 0xd8], segment(0xda, [1]))), "jpeg"), none, "SOS before SOF");
+    assert.deepEqual(await readImageHeader(blobOf(concat([0xff, 0xd8], [0xff])), "jpeg"), none, "ends on a marker byte");
+    assert.deepEqual(await readImageHeader(blobOf(concat([0xff, 0xd8], [0xff, 0xe1, 0x00])), "jpeg"), none, "ends inside a length");
+    assert.deepEqual(await readImageHeader(blobOf(concat([0xff, 0xd8], [0xff, 0xc0, 0x00, 0x08, 8, 0])), "jpeg"), none, "ends inside SOF");
+    assert.deepEqual(await readImageHeader(blobOf(concat([0xff, 0xd8], segment(0xe1, []))), "jpeg"), none, "no SOF at all");
+    assert.deepEqual(await readImageHeader(blobOf(Uint8Array.from([0xff])), "jpeg"), none, "too short");
+    assert.deepEqual(await readImageHeader(blobOf(latin1("RIFF")), "jpeg"), none, "not a JPEG");
+    // A SOF past the 64 KiB read gives no dimensions rather than reading on.
+    assert.deepEqual(await readImageHeader(blobOf(concat([0xff, 0xd8], segment(0xe1, fill(0, 65000)), segment(0xe2, fill(0, 1000)), sof(0xc0, 5, 5))), "jpeg"), none, "SOF beyond 64 KiB");
+});
+
+test("image header: WebP VP8, VP8L and VP8X headers", async () => {
+    const riff = (...rest) => concat(latin1("RIFF"), u32(0), latin1("WEBP"), ...rest);
+    const chunkHeader = (type, size) => concat(latin1(type), Uint8Array.from([size & 0xff, (size >>> 8) & 0xff, (size >>> 16) & 0xff, 0]));
+    const vp8 = riff(chunkHeader("VP8 ", 100), [0x30, 0x0b, 0x00, 0x9d, 0x01, 0x2a], [40, 0x40], [24, 0x80], fill(0, 20));
+    assert.deepEqual(await readImageHeader(blobOf(vp8), "webp"), { format: "WEBP", width: 40, height: 24 }, "scale bits are masked");
+    // 14-bit width-1 = 63 and height-1 = 47, little-endian bit-packed after the 0x2f signature.
+    const vp8l = riff(chunkHeader("VP8L", 100), [0x2f], [0x3f, 0xc0, 0x0b, 0x00, 0x00], fill(0, 20));
+    assert.deepEqual(await readImageHeader(blobOf(vp8l), "webp"), { format: "WEBP", width: 64, height: 48 });
+    const vp8x = riff(chunkHeader("VP8X", 10), [0x10, 0, 0, 0], [0xff, 0x00, 0x01], [0x2f, 0x00, 0x00], fill(0, 20));
+    assert.deepEqual(await readImageHeader(blobOf(vp8x), "webp"), { format: "WEBP", width: 65792, height: 48 });
+    const none = { format: "WEBP", width: null, height: null };
+    assert.deepEqual(await readImageHeader(blobOf(riff(chunkHeader("VP8 ", 100), [0x30, 0x0b, 0x00, 0x00, 0x01, 0x2a], fill(0, 24))), "webp"), none, "bad start code");
+    assert.deepEqual(await readImageHeader(blobOf(riff(chunkHeader("VP8L", 100), [0x00], fill(0, 24))), "webp"), none, "bad VP8L signature");
+    assert.deepEqual(await readImageHeader(blobOf(riff(chunkHeader("ALPH", 100), fill(0, 24))), "webp"), none, "unknown first chunk");
+    assert.deepEqual(await readImageHeader(blobOf(riff(chunkHeader("VP8 ", 100))), "webp"), none, "too short");
+    assert.deepEqual(await readImageHeader(blobOf(latin1("RIFF\0\0\0\0WAVEfmt ".padEnd(40, "\0"))), "webp"), none, "not WebP");
+});
+
+test("image header: other kinds give no format and the read stays within 64 KiB", async () => {
+    assert.deepEqual(await readImageHeader(blobOf(bytes.SIGNATURE), "png"), { format: null, width: null, height: null });
+    const file = recordingBlob(concat([0xff, 0xd8], fill(0, 200000)));
+    await readImageHeader(file, "jpeg");
+    assert.deepEqual(file.ranges, [[0, 65536]]);
+    const small = recordingBlob(concat([0xff, 0xd8, 0xff]));
+    await readImageHeader(small, "jpeg");
+    assert.deepEqual(small.ranges, [[0, 3]]);
+});

tests/js/support-module-urls.test.js

@@ -0,0 +1,39 @@
+// The Node oracles that Python drives (tests/js/support/*.mjs) import product
+// modules dynamically. A dynamic import takes a URL or a bare/relative specifier,
+// never a filesystem path: on Windows `D:\...\x.mjs` parses as the `d:` scheme
+// (ERR_UNSUPPORTED_ESM_URL_SCHEME). Every such import goes through moduleUrl, and
+// this test proves the Windows form on any host so Linux catches a regression.
+import test from "node:test";
+import assert from "node:assert/strict";
+import fs from "node:fs";
+import path from "node:path";
+import { fileURLToPath } from "node:url";
+
+const here = path.dirname(fileURLToPath(import.meta.url));
+const support = path.join(here, "support");
+const m = await import("./support/module-url.mjs").catch((error) => {
+    if (error.code !== "ERR_MODULE_NOT_FOUND") throw error;
+    return {};
+});
+
+test("moduleUrl turns a filesystem path into a file: URL on every platform", async () => {
+    assert.equal(typeof m.moduleUrl, "function", "moduleUrl is exported");
+    assert.equal(m.moduleUrl("D:\\x\\y z.mjs", { windows: true }), "file:///D:/x/y%20z.mjs");
+    assert.equal(m.moduleUrl("C:\\repo\\web\\js\\workspace\\local-video.mjs", { windows: true }), "file:///C:/repo/web/js/workspace/local-video.mjs");
+    assert.equal(m.moduleUrl("/repo/web/js/a b.mjs", { windows: false }), "file:///repo/web/js/a%20b.mjs");
+    const own = m.moduleUrl(path.join(here, "support", "module-url.mjs"));
+    assert.ok(own.startsWith("file:///"), own);
+    assert.equal((await import(own)).moduleUrl, m.moduleUrl, "the URL imports the same module instance");
+});
+
+test("every dynamic import in the Node oracles is a file URL, never a path", () => {
+    const scripts = fs.readdirSync(support).filter((name) => name.endsWith(".mjs"));
+    assert.ok(scripts.includes("local-inspect-oracle.mjs"));
+    for (const name of scripts) {
+        const source = fs.readFileSync(path.join(support, name), "utf8").split("\n").filter((line) => !line.trim().startsWith("//")).join("\n");
+        for (const match of source.matchAll(/\bimport\(\s*([^)]*?)[,)]/g)) {
+            const argument = match[1].trim();
+            assert.ok(/^(moduleUrl\(|pathToFileURL\(|["'`]\.{1,2}\/)/.test(argument), `${name}: import(${argument}) must be a file URL or a relative specifier`);
+        }
+    }
+});

tests/js/support/local-inspect-oracle.mjs

@@ -0,0 +1,59 @@
+// The browser's local readers and projection run over files on disk, for
+// tests/test_local_inspect_parity.py, which compares every answer with the server's
+// utils.prompt_inspector.read_inspection_file and utils.inspection_response._metadata
+// on the same file. Each file is opened as a Blob (fs.openAsBlob) so the reader
+// slices it exactly as it slices a File.
+// stdin: [{name, path} | {name, tags, limited?} | {name, records, record_formats?, carrier_text?, limited?}]
+// stdout: {name: {classified, result, metadata}} where result is readPngMetadataFile's
+//         answer for a PNG, readImageHeader's for another image, for a video
+//         {read: readVideoTags(...), selection: selectVideoCarrier(read.tags, {limited})}
+//         (selection null unless the read completed), else null; metadata is
+//         projectMetadata(...) of the local read (the server's _metadata) for every
+//         classified file, else null.
+//         A {tags} entry answers {classified: null, result: {selection}, metadata} for
+//         the carrier rule alone, as the server selects from ffprobe's tag map.
+//         A {records} entry (records null = an unreadable file) answers
+//         {classified: null, result: null, metadata} for the projection alone.
+import fs from 'node:fs';
+import path from 'node:path';
+import { createRequire } from 'node:module';
+import { fileURLToPath } from 'node:url';
+import { moduleUrl } from './module-url.mjs';
+
+const here = path.dirname(fileURLToPath(import.meta.url));
+const require = createRequire(import.meta.url);
+const pngMetadata = require(path.join(here, '..', '..', '..', 'web', 'js', 'png-metadata.js'));
+const web = path.join(here, '..', '..', '..', 'web', 'js', 'workspace');
+// ES modules import by file URL (a Windows path would parse as a URL scheme).
+const { selectVideoCarrier } = await import(moduleUrl(path.join(web, 'local-video.mjs')));
+const { projectMetadata } = await import(moduleUrl(path.join(web, 'local-projection.mjs')));
+const { readLocalFile, toInspectionRead } = await import(moduleUrl(path.join(web, 'local-inspection.mjs')));
+
+const selectionRead = (selection) => selection.status === 'complete' ?
+    { records: selection.records, record_formats: selection.record_formats, carrierText: selection.carrierText, limited: selection.limited } : null;
+
+const chunks = [];
+for await (const chunk of process.stdin) chunks.push(chunk);
+const output = {};
+for (const entry of JSON.parse(Buffer.concat(chunks).toString('utf8'))) {
+    const { name } = entry;
+    if (entry.records !== undefined) {
+        const read = entry.records === null ? null : {
+            records: entry.records,
+            record_formats: entry.record_formats ?? Object.fromEntries(Object.keys(entry.records).map((key) => [key, 'original-text'])),
+            carrierText: entry.carrier_text ?? null, limited: entry.limited ?? [],
+        };
+        output[name] = { classified: null, result: null, metadata: projectMetadata(read) };
+        continue;
+    }
+    if (entry.path === undefined) {
+        const selection = selectVideoCarrier(entry.tags, { limited: entry.limited || [] });
+        output[name] = { classified: null, result: { selection }, metadata: projectMetadata(selectionRead(selection)) };
+        continue;
+    }
+    const file = await fs.openAsBlob(entry.path);
+    const { classification, result } = await readLocalFile(file, { pngMetadata });
+    const metadata = classification === null ? null : projectMetadata(toInspectionRead(classification, result).read);
+    output[name] = { classified: classification, result, metadata };
+}
+process.stdout.write(JSON.stringify(output));

tests/js/support/module-url.mjs

@@ -0,0 +1,9 @@
+// A dynamic import() takes a URL, never a filesystem path: on Windows a path such
+// as D:\repo\x.mjs parses as the "d:" scheme (ERR_UNSUPPORTED_ESM_URL_SCHEME).
+// Every oracle in this directory imports product modules through moduleUrl.
+import { pathToFileURL } from 'node:url';
+
+/** The file: URL for a filesystem path; `windows` forces the Windows form on any host (for tests). */
+export function moduleUrl(filePath, { windows = process.platform === 'win32' } = {}) {
+    return pathToFileURL(filePath, { windows }).href;
+}

tests/js/workspace-file-inspector.test.js

@@ -86,3 +86,34 @@ test('Show in folder outcomes use the D38 toasts',async()=>{
     for(const error of [{status:403,code:null},{status:409,code:'source-changed'},{status:415,code:'unsupported-media-type'},{status:429,code:null},{status:500},new TypeError('offline')])
         assert.equal(revealOutcome(error),'The folder could not be opened. Try again.');
 });
+
+// #249 PR C Task C3 (D88): one chip row, the file's current image tags first, then D21's linked tags not already shown.
+function tagged(imageTags, library = {status:'linked', tags:['Minimal', 'warm_light', 'editorial'], tags_scope:'positive'}) {
+    return {...record(), imageTags, library};
+}
+test('renders current image tags before linked tags without duplicates', async () => {
+    const {inspectorTagChips} = await load();
+    assert.equal(typeof inspectorTagChips, 'function', 'inspectorTagChips is exported');
+    assert.deepEqual(inspectorTagChips(tagged({state:'current', tags:['dunes', 'minimal']}), {filterable:true}),
+        [{name:'dunes', filter:true}, {name:'minimal', filter:true}, {name:'warm_light', filter:true}, {name:'editorial', filter:true}]);
+    // D21's rule still decides the linked chips: a recipe's own tags stay plain, an image tag stays a filter.
+    assert.deepEqual(inspectorTagChips(tagged({state:'current', tags:['Editorial']}, {status:'linked', tags:['editorial', 'cover'], tags_scope:'pair'}), {filterable:true}),
+        [{name:'Editorial', filter:true}, {name:'cover', filter:false}]);
+    // An unlinked file shows only its own tags; a facet-less record (history, local) only the linked ones.
+    assert.deepEqual(inspectorTagChips(tagged({state:'current', tags:['loose']}, {status:'none', tags:[], tags_scope:null}), {filterable:true}), [{name:'loose', filter:true}]);
+    assert.deepEqual(inspectorTagChips(tagged(null), {filterable:true}).map(chip => chip.name), ['Minimal', 'warm_light', 'editorial']);
+    assert.deepEqual(inspectorTagChips(tagged({state:'current', tags:[]}, {status:'ambiguous', tags:['x'], tags_scope:'positive'}), {filterable:true}), []);
+});
+test('image tag chips filter in the panel and are plain in the viewer', async () => {
+    const {inspectorTagChips} = await load();
+    const source = tagged({state:'current', tags:['dunes']});
+    assert.deepEqual(inspectorTagChips(source, {filterable:true}).map(chip => chip.filter), [true, true, true, true]);
+    assert.deepEqual(inspectorTagChips(source, {filterable:false}).map(chip => chip.filter), [false, false, false, false]);
+});
+test('none and unavailable render no image chips', async () => {
+    const {inspectorTagChips} = await load();
+    for (const state of ['none', 'unavailable']) {
+        assert.deepEqual(inspectorTagChips(tagged({state, tags:[]}), {filterable:true}).map(chip => chip.name), ['Minimal', 'warm_light', 'editorial']);
+        assert.deepEqual(inspectorTagChips(tagged({state, tags:[]}, {status:'none', tags:[], tags_scope:null}), {filterable:true}), []);
+    }
+});

tests/js/workspace-image-layout.test.js

@@ -149,3 +149,15 @@ test('a narrow gallery caps the concept columns at the smallest readable card',
     }
     assert.equal(layoutImages([{}], 990, 'list', {columns:3}).columns, 1);
 });
+
+test('the card info stacks below the narrowest bounded card, a list row beside its art too (D31)', async () => {
+    const {stacksCardInfo} = await source;
+    assert.equal(typeof stacksCardInfo, 'function', 'stacksCardInfo is not implemented');
+    assert.equal(stacksCardInfo(119, 'grid'), true);
+    assert.equal(stacksCardInfo(120, 'grid'), false);
+    assert.equal(stacksCardInfo(119, 'masonry'), true);
+    // A list row's info is what its art (160px, at most 45%) leaves: 195px at 200% zoom on a phone leaves 107px.
+    assert.equal(stacksCardInfo(195, 'list'), true);
+    assert.equal(stacksCardInfo(280, 'list'), false);
+    assert.equal(stacksCardInfo(358, 'list'), false);
+});

tests/js/workspace-image-tagging-state.test.js

@@ -0,0 +1,769 @@
+// Run with: node --test tests/js/workspace-image-tagging-state.test.js
+const test = require('node:test');
+const assert = require('node:assert/strict');
+const { createApiClient } = require('../../web/js/api-client.js');
+
+// Only an absent module is tolerated, so RED is an assertion failure, not a crash.
+const source = import('../../web/js/workspace/image-tagging-state.mjs').catch(error => {
+    if (error.code !== 'ERR_MODULE_NOT_FOUND') throw error;
+    return {};
+});
+async function load() {
+    const module = await source;
+    assert.equal(typeof module.createImageTaggingController, 'function', 'image-tagging-state.mjs is not implemented');
+    return module;
+}
+
+const EPOCH = '11111111-1111-4111-8111-111111111111';
+const SCOPE = 'c'.repeat(64);
+const SHA = 'd'.repeat(64);
+const TOKEN = 'capture-token_1';
+const TOKEN2 = 'capture-token_2';
+const TARGET = Object.freeze({ root_index: 0, root_scope: SCOPE, path: 'dunes/dunes_00001.png' });
+const OBSERVE = '/prompt_manager/image-annotations/observe';
+const SUGGEST = '/prompt_manager/image-annotations/suggest';
+const TAGS = '/prompt_manager/image-annotations/tags';
+const MODELS = '/prompt_manager/autotag/models';
+const CAPTURE = '/prompt_manager/image-annotations/capture/';
+const DEFAULT_PROMPT = 'Generate only comma-separated Danbooru tags (lowercase_underscores, no spaces).';
+
+const flush = () => new Promise(resolve => setImmediate(resolve));
+async function settle(times = 8) { for (let i = 0; i < times; i++) await flush(); }
+function deferred() {
+    let resolve, reject;
+    const promise = new Promise((yes, no) => { resolve = yes; reject = no; });
+    return { promise, resolve, reject };
+}
+function response(status, body) {
+    return { ok: status >= 200 && status < 300, status, statusText: `Status ${status}`, json: async () => body };
+}
+const fail = (status, code, error = `failure ${code}`, extra = {}) => response(status, { success: false, code, error, ...extra });
+
+function captureInfo(overrides = {}) {
+    return {
+        token: TOKEN, expires_at: '2026-10-22T12:00:00+00:00', preview_url: CAPTURE + (overrides.token || TOKEN),
+        sha256: SHA, size: 1234, width: 832, height: 1216, media_type: 'image',
+        generate: { supported: true, reason: null }, ...overrides,
+    };
+}
+function observed(overrides = {}, capture = {}) {
+    return {
+        success: true, library_epoch: EPOCH, target: { ...TARGET }, capture: captureInfo(capture),
+        title: 'Desert pavilion', filename: 'dunes_00001.png', tags: ['Minimal', 'warm_light'], tag_revision: 3,
+        ...overrides,
+    };
+}
+const models = (overrides = {}) => ({
+    success: true, models: { 'wd14-swinv2': { downloaded: true } }, default_prompt: DEFAULT_PROMPT,
+    wd14_general_threshold: 0.35, wd14_character_threshold: 0.85, suggestion_limit: 10, ...overrides,
+});
+const SANDSTONE = [
+    { name: 'minimal', confidence: 0.99 }, { name: 'editorial', confidence: 0.97 },
+    { name: 'sandstone', confidence: 0.981 }, { name: 'Editorial', confidence: 0.5 },
+];
+const suggested = (rows = SANDSTONE, overrides = {}) => ({ success: true, model_type: 'wd14-swinv2', options: {}, suggestions: rows, ...overrides });
+const receipt = (tags, overrides = {}) => ({ success: true, library_epoch: EPOCH, sha256: SHA, tags, tag_revision: 4, replayed: false, ...overrides });
+
+/** Real createApiClient over a fetch whose answers come from routes[METHOD URL](call). */
+function harness(routes = {}) {
+    const calls = [];
+    const fetch = (url, init = {}) => {
+        const call = { url, method: init.method, init, body: init.body === undefined ? undefined : JSON.parse(init.body), aborted: false, signal: init.signal };
+        calls.push(call);
+        return new Promise((resolve, reject) => {
+            init.signal?.addEventListener('abort', () => {
+                call.aborted = true;
+                reject(new DOMException('The operation was aborted.', 'AbortError'));
+            });
+            const key = `${call.method} ${url.startsWith(CAPTURE) ? CAPTURE : url}`;
+            const route = routes[key];
+            Promise.resolve().then(() => {
+                if (!route) return response(200, { success: true });
+                return route(call);
+            }).then(resolve, reject);
+        });
+    };
+    return {
+        api: createApiClient(fetch), calls,
+        of: (method, url) => calls.filter(call => call.method === method && (url === CAPTURE ? call.url.startsWith(CAPTURE) : call.url === url)),
+    };
+}
+function defaultRoutes(extra = {}) {
+    return {
+        [`POST ${OBSERVE}`]: () => response(200, observed()),
+        [`GET ${MODELS}`]: () => response(200, models()),
+        [`POST ${SUGGEST}`]: () => response(200, suggested()),
+        [`PUT ${TAGS}`]: call => response(200, receipt(call.body.tags)),
+        ...extra,
+    };
+}
+async function opened(routes, options = {}) {
+    const { createImageTaggingController } = await load();
+    const h = harness(defaultRoutes(routes));
+    const c = createImageTaggingController({ api: h.api, target: TARGET, source: 'images', ...options });
+    await c.open();
+    await settle();
+    return { c, h };
+}
+
+test('MODEL_OPTIONS lists legacy engines in mock order with D84 labels', async () => {
+    const { MODEL_OPTIONS } = await load();
+    assert.deepEqual(MODEL_OPTIONS.map(option => [option.id, option.label]), [
+        ['wd14-swinv2', 'WD14 SwinV2 · ~400 MB'],
+        ['wd14-vit', 'WD14 ViT · ~400 MB'],
+        ['gguf', 'GGUF JoyCaption (Recommended) · ~4.5 GB'],
+        ['hf', 'HuggingFace JoyCaption · ~16 GB'],
+    ]);
+    assert.ok(Object.isFrozen(MODEL_OPTIONS) && Object.isFrozen(MODEL_OPTIONS[0]));
+});
+
+test('dedupeTags trims, drops blanks and keeps the first spelling case-insensitively', async () => {
+    const { dedupeTags, parseManualTags } = await load();
+    const input = [' Minimal ', 'minimal', '', 'warm_light', 'WARM_LIGHT', 7, 'dune'];
+    assert.deepEqual(dedupeTags(input), ['Minimal', 'warm_light', 'dune']);
+    assert.deepEqual(input[0], ' Minimal ', 'input is not mutated');
+    assert.deepEqual(dedupeTags(undefined), []);
+    assert.deepEqual(parseManualTags(' a, ,B , b,c '), ['a', 'B', 'c']);
+    assert.deepEqual(parseManualTags(null), []);
+});
+
+test('the controller rejects a malformed target, source or api', async () => {
+    const { createImageTaggingController } = await load();
+    const { api } = harness();
+    assert.throws(() => createImageTaggingController({ api, target: { ...TARGET, root_index: true } }), TypeError);
+    assert.throws(() => createImageTaggingController({ api, target: { ...TARGET, root_scope: 'x' } }), TypeError);
+    assert.throws(() => createImageTaggingController({ api, target: { ...TARGET, path: '' } }), TypeError);
+    assert.throws(() => createImageTaggingController({ api, target: TARGET, source: 'elsewhere' }), TypeError);
+    assert.throws(() => createImageTaggingController({ api: {}, target: TARGET }), TypeError);
+    assert.throws(() => createImageTaggingController(), TypeError);
+});
+
+test('opens with saved tags joined by comma-space', async () => {
+    const { createImageTaggingController } = await load();
+    const h = harness(defaultRoutes());
+    const c = createImageTaggingController({ api: h.api, target: TARGET, source: 'images' });
+    const seen = [];
+    const unsubscribe = c.subscribe(state => seen.push(state.phase));
+    assert.equal(c.state.phase, 'observing');
+    const pending = c.open();
+    assert.equal(c.state.phase, 'observing');
+    await pending;
+    await settle();
+    assert.equal(c.state.phase, 'ready');
+    assert.equal(c.state.manual, 'Minimal, warm_light');
+    assert.deepEqual(c.state.savedTags, ['Minimal', 'warm_light']);
+    assert.equal(c.state.image.filename, 'dunes_00001.png');
+    assert.equal(c.state.image.title, 'Desert pavilion');
+    assert.equal(c.state.image.dimensions, '832 × 1216');
+    assert.equal(c.state.image.previewUrl, CAPTURE + TOKEN);
+    assert.equal(c.state.model, 'wd14-swinv2');
+    assert.equal(c.state.countText, 'Not generated');
+    assert.equal(c.state.primaryLabel, 'Save manual tags');
+    assert.equal(c.state.suggestionsText, 'Generate suggestions with the selected model before approving.');
+    assert.equal(c.state.showScopeNote, false);
+    assert.equal(c.state.canSave, true);
+    assert.equal(c.state.canGenerate, true);
+    assert.ok(Object.isFrozen(c.state));
+    assert.deepEqual(h.of('POST', OBSERVE)[0].body, { target: { ...TARGET } });
+    assert.ok(seen.includes('ready'));
+    unsubscribe();
+    c.setManual('x');
+    assert.equal(seen.at(-1), 'ready', 'an unsubscribed listener is not called');
+    assert.throws(() => c.subscribe('nope'), TypeError);
+});
+
+test('an image without known dimensions shows none and an unseen file starts empty', async () => {
+    const { c } = await opened({
+        [`POST ${OBSERVE}`]: () => response(200, observed({ tags: [], tag_revision: 0, title: null }, { width: null, height: null })),
+    });
+    assert.equal(c.state.image.dimensions, null);
+    assert.equal(c.state.manual, '');
+    assert.equal(c.state.image.title, null);
+});
+
+test('instructions default to the server prompt', async () => {
+    const { c } = await opened({
+        [`GET ${MODELS}`]: () => response(200, models({ suggestion_limit: 7, wd14_general_threshold: 0.5 })),
+    });
+    assert.equal(c.state.options.instructions, DEFAULT_PROMPT);
+    assert.equal(c.state.options.suggestion_limit, 7);
+    assert.equal(c.state.options.general_threshold, 0.5);
+    assert.equal(c.state.options.character_threshold, 0.85);
+    c.setModel('gguf');
+    assert.equal(c.state.options.instructions, DEFAULT_PROMPT);
+    c.setOption('instructions', 'only colours');
+    c.setModel('hf');
+    assert.equal(c.state.options.instructions, DEFAULT_PROMPT, 'a model change re-renders the defaults like the mock');
+    assert.equal(c.state.options.suggestion_limit, 7, 'the suggestion count survives a model change');
+});
+
+test('an unreadable or malformed models read falls back to the documented defaults', async () => {
+    for (const route of [() => fail(500, null), () => response(200, models({ default_prompt: 4, suggestion_limit: 0, wd14_general_threshold: 2, wd14_character_threshold: 'x' }))]) {
+        const { c } = await opened({ [`GET ${MODELS}`]: route });
+        assert.equal(c.state.phase, 'ready');
+        assert.deepEqual({ ...c.state.options }, { general_threshold: 0.35, character_threshold: 0.85, instructions: '', suggestion_limit: 10 });
+    }
+});
+
+test('observe failures and malformed answers end in observe-error and retry observes again', async () => {
+    const { createImageTaggingController } = await load();
+    let answer = () => fail(409, 'source-changed', 'The source changed. Refresh the image collection before tagging this file.');
+    const h = harness(defaultRoutes({ [`POST ${OBSERVE}`]: () => answer() }));
+    const c = createImageTaggingController({ api: h.api, target: TARGET });
+    await c.open();
+    assert.equal(c.state.phase, 'observe-error');
+    assert.equal(c.state.statusText, 'The source changed. Refresh the image collection before tagging this file.');
+    assert.equal(c.state.canSave, false);
+    assert.equal(c.state.canGenerate, false);
+    assert.equal(c.state.canRetry, true);
+    answer = () => response(200, observed({ tag_revision: -1 }));
+    await c.retry();
+    assert.equal(c.state.phase, 'observe-error');
+    assert.equal(c.state.statusText, 'The image could not be opened for tagging.');
+    assert.equal(h.of('DELETE', CAPTURE).length, 1, 'a malformed answer still releases its capture');
+    answer = () => Promise.reject(new TypeError('Failed to fetch'));
+    await c.retry();
+    assert.equal(c.state.statusText, 'The image could not be opened for tagging.');
+    answer = () => response(200, observed());
+    await c.retry();
+    await settle();
+    assert.equal(c.state.phase, 'ready');
+    assert.equal(h.of('POST', OBSERVE).length, 4);
+});
+
+test('model change clears suggestions count and status but keeps manual', async () => {
+    const { c } = await opened();
+    c.setManual('Minimal, dune');
+    await c.generate();
+    assert.equal(c.state.countText, '3 suggestions');
+    c.setModel('wd14-vit');
+    assert.deepEqual(c.state.suggestions, []);
+    assert.equal(c.state.countText, 'Not generated');
+    assert.equal(c.state.statusText, '');
+    assert.equal(c.state.suggestionsText, 'Generate suggestions with the selected model before approving.');
+    assert.equal(c.state.manual, 'Minimal, dune');
+    assert.equal(c.state.model, 'wd14-vit');
+    c.setModel('wd14-vit');
+    assert.equal(c.state.model, 'wd14-vit');
+    assert.throws(() => c.setModel('ollama'), TypeError);
+    assert.throws(() => c.setOption('model', 1), TypeError);
+});
+
+test('generate sends the model options and stages scored chips', async () => {
+    const { c, h } = await opened();
+    c.setOption('general_threshold', '0.4');
+    c.setOption('suggestion_limit', '3');
+    assert.equal(await c.generate(), true);
+    const sent = h.of('POST', SUGGEST)[0].body;
+    assert.deepEqual(sent, {
+        target: { ...TARGET }, capture_token: TOKEN, expected_epoch: EPOCH, model_type: 'wd14-swinv2',
+        options: { general_threshold: 0.4, character_threshold: 0.85, suggestion_limit: 3 },
+    });
+    assert.deepEqual(c.state.suggestions.map(row => [row.name, row.label]), [['minimal', '99%'], ['editorial', '97%'], ['sandstone', '98%']]);
+    assert.equal(c.state.statusText, 'Suggestions from WD14 SwinV2 · ~400 MB. Approve tags to save.');
+    assert.equal(c.state.suggestionsText, '');
+    c.setModel('gguf');
+    await c.generate();
+    assert.deepEqual(h.of('POST', SUGGEST)[1].body.options, { instructions: DEFAULT_PROMPT, suggestion_limit: 3 });
+});
+
+test('language models label chips Suggested and the count is capped at the requested limit', async () => {
+    const rows = Array.from({ length: 12 }, (_, i) => ({ name: `tag${i}`, confidence: null }));
+    const { c } = await opened({ [`POST ${SUGGEST}`]: () => response(200, suggested(rows, { model_type: 'gguf' })) });
+    c.setModel('gguf');
+    await c.generate();
+    assert.equal(c.state.suggestions.length, 10);
+    assert.ok(c.state.suggestions.every(row => row.label === 'Suggested' && row.confidence === null));
+    assert.equal(c.state.countText, '10 suggestions');
+});
+
+test('an empty result shows the threshold copy and a malformed one is a generate error', async () => {
+    let answer = () => response(200, suggested([]));
+    const { c } = await opened({ [`POST ${SUGGEST}`]: () => answer() });
+    await c.generate();
+    assert.equal(c.state.suggestionsText, 'No suggestions met this threshold. Adjust it or enter tags manually.');
+    assert.equal(c.state.countText, '0 suggestions');
+    assert.equal(c.state.primaryLabel, 'Approve selected tags');
+    answer = () => response(200, suggested([{ name: 4, confidence: 2 }]));
+    await c.generate();
+    assert.equal(c.state.phase, 'generate-error');
+    assert.equal(c.state.statusText, 'Auto-tagging failed. Check server logs for details.');
+    answer = () => response(200, suggested([{ name: 'one', confidence: 1 }]));
+    assert.equal(await c.retry(), true, 'retry after a generate error generates again');
+    assert.equal(c.state.countText, '1 suggestion');
+    assert.equal(c.state.suggestions[0].label, '100%');
+});
+
+test('invalid options send nothing', async () => {
+    const { c, h } = await opened();
+    const { optionError } = await load();
+    for (const [key, value] of [['general_threshold', '1.5'], ['general_threshold', ''], ['character_threshold', 'abc'], ['suggestion_limit', '0'], ['suggestion_limit', '2.5'], ['suggestion_limit', '101']]) {
+        c.setOption('general_threshold', 0.35);
+        c.setOption('character_threshold', 0.85);
+        c.setOption('suggestion_limit', 10);
+        c.setOption(key, value);
+        assert.equal(await c.generate(), false, `${key}=${value}`);
+        assert.equal(optionError(c.state.model, c.state.options).field, key);
+    }
+    c.setModel('gguf');
+    c.setOption('instructions', '   ');
+    assert.equal(await c.generate(), false);
+    c.setOption('instructions', 'x'.repeat(4001));
+    assert.equal(await c.generate(), false);
+    assert.equal(optionError('gguf', c.state.options).field, 'instructions');
+    assert.equal(h.of('POST', SUGGEST).length, 0);
+    assert.equal(optionError('wd14-vit', { general_threshold: 0, character_threshold: 1, suggestion_limit: 100 }), null);
+});
+
+test('late generation for an old serial, closed or expired dialog is ignored', async () => {
+    const pending = [];
+    const { c, h } = await opened({ [`POST ${SUGGEST}`]: () => { const d = deferred(); pending.push(d); return d.promise; } });
+    const first = c.generate();
+    assert.equal(c.state.phase, 'generating');
+    assert.equal(c.state.statusText, 'Generating suggestions…');
+    assert.equal(c.state.canSave, false);
+    c.setModel('wd14-vit');
+    assert.equal(c.state.phase, 'ready');
+    assert.equal(h.of('POST', SUGGEST)[0].aborted, true);
+    await settle();
+    pending[0].resolve(response(200, suggested()));
+    assert.equal(await first, false);
+    assert.equal(c.state.phase, 'ready', 'the superseded request neither stages chips nor reports its abort');
+    assert.equal(c.state.statusText, '');
+    assert.deepEqual(c.state.suggestions, []);
+    assert.equal(c.state.countText, 'Not generated');
+
+    const second = c.generate();
+    c.cancel();
+    await settle();
+    pending[1].resolve(response(200, suggested()));
+    assert.equal(await second, false);
+    assert.deepEqual(c.state.suggestions, []);
+    assert.equal(c.state.closed, true);
+    assert.equal(await c.generate(), false, 'a closed dialog never generates');
+
+    const expired = await opened({ [`POST ${SUGGEST}`]: () => fail(409, 'capture-expired', 'The preview expired. Open the Tag dialog again.') });
+    await expired.c.generate();
+    assert.equal(expired.c.state.phase, 'capture-expired');
+    assert.equal(expired.c.state.canSave, false);
+    assert.equal(expired.c.state.canGenerate, false);
+});
+
+test('a reopen while observing ignores the older observation and releases its capture', async () => {
+    const { createImageTaggingController } = await load();
+    const answers = [deferred(), deferred()];
+    let n = 0;
+    const h = harness(defaultRoutes({ [`POST ${OBSERVE}`]: () => answers[n++].promise }));
+    const c = createImageTaggingController({ api: h.api, target: TARGET });
+    const one = c.open();
+    const two = c.open();
+    answers[1].resolve(response(200, observed({}, { token: TOKEN2 })));
+    await two;
+    answers[0].resolve(response(200, observed()));
+    await one;
+    await settle();
+    assert.equal(c.state.image.previewUrl, CAPTURE + TOKEN2);
+    assert.deepEqual(h.of('DELETE', CAPTURE).map(call => call.url), [CAPTURE + TOKEN]);
+    c.dispose();
+    assert.deepEqual(h.of('DELETE', CAPTURE).map(call => call.url), [CAPTURE + TOKEN, CAPTURE + TOKEN2]);
+});
+
+test('primary label toggles Save manual tags / Approve selected tags', async () => {
+    const { c } = await opened();
+    assert.equal(c.state.primaryLabel, 'Save manual tags');
+    await c.generate();
+    assert.equal(c.state.primaryLabel, 'Approve selected tags'); c.setModel('wd14-vit'); assert.equal(c.state.primaryLabel, 'Save manual tags');
+    assert.equal(c.state.countText, 'Not generated');
+});
+
+test('remove drops one chip and the last removal shows the all-removed copy', async () => {
+    const { c } = await opened();
+    await c.generate();
+    c.remove('editorial');
+    assert.deepEqual(c.state.suggestions.map(row => row.name), ['minimal', 'sandstone']);
+    assert.equal(c.state.countText, '2 suggestions');
+    c.remove('missing');
+    assert.equal(c.state.suggestions.length, 2);
+    c.remove('minimal');
+    c.remove('sandstone');
+    assert.deepEqual(c.state.suggestions, []);
+    assert.equal(c.state.suggestionsText, 'All suggestions removed. Generate again or enter tags manually.');
+    assert.equal(c.state.primaryLabel, 'Approve selected tags');
+});
+
+test('save sends manual then remaining chips deduplicated case-insensitively keeping the manual spelling', async () => {
+    const { createImageTaggingController } = await load();
+    const h = harness(defaultRoutes({
+        [`POST ${SUGGEST}`]: () => response(200, suggested([
+            { name: 'minimal', confidence: 0.99 }, { name: 'editorial', confidence: 0.97 },
+            { name: 'sandstone', confidence: 0.96 }, { name: 'desert', confidence: 0.9 },
+        ])),
+    }));
+    const c = createImageTaggingController({ api: h.api, target: TARGET, source: 'review' });
+    await c.open();
+    await settle();
+    assert.equal(c.state.showScopeNote, true);
+    await c.generate();
+    c.remove('desert');
+    assert.equal(c.state.primaryLabel, 'Approve selected tags');
+    assert.equal(await c.save(), true);
+    const sent = h.of('PUT', TAGS)[0].body;
+    assert.deepEqual(sent.tags, ['Minimal', 'warm_light', 'editorial', 'sandstone']); // manual "Minimal", chip "minimal" dropped
+    assert.deepEqual(sent, {
+        target: { ...TARGET }, capture_token: TOKEN, sha256: SHA, expected_epoch: EPOCH,
+        expected_tag_revision: 3, tags: sent.tags,
+    });
+    assert.equal(c.state.phase, 'saved');
+    assert.deepEqual(c.state.savedTags, sent.tags);
+    assert.equal(c.state.tagRevision, 4);
+    assert.equal(c.state.toast, '4 tags saved for dunes_00001.png. Prompt association still awaits review.');
+    assert.equal(c.state.canSave, false);
+    assert.equal(await c.save(), false, 'a saved dialog does not save twice');
+});
+
+test('toast matches the mock with the review suffix', async () => {
+    const { c } = await opened({ [`PUT ${TAGS}`]: () => response(200, receipt(['one'])) });
+    c.setManual('one');
+    await c.save();
+    assert.equal(c.state.toast, '1 tags saved for dunes_00001.png.', 'the mock always says "tags"');
+    const many = await opened({}, { source: 'library' });
+    await many.c.save();
+    assert.equal(many.c.state.toast, '2 tags saved for dunes_00001.png.');
+});
+
+test('empty manual and no chips sends the acknowledged empty set', async () => {
+    const { c, h } = await opened({ [`PUT ${TAGS}`]: () => response(200, receipt([])) });
+    c.setManual('  ,  ');
+    await c.save();
+    assert.deepEqual(h.of('PUT', TAGS)[0].body.tags, []);
+    assert.equal(c.state.toast, '0 tags saved for dunes_00001.png.');
+});
+
+test('conflict keeps draft and exposes current saved tags', async () => {
+    let answer = () => fail(409, 'tag-conflict', 'The tags of this image changed since the dialog opened; review them first.',
+        { tags: ['night', 'rain'], tag_revision: 5 });
+    const { c, h } = await opened({ [`PUT ${TAGS}`]: call => answer(call) });
+    await c.generate();
+    c.setManual('Minimal, dune');
+    await c.save();
+    assert.equal(c.state.phase, 'conflict');
+    assert.deepEqual(c.state.currentTags, ['night', 'rain']);
+    assert.equal(c.state.manual, 'night, rain, Minimal, dune');
+    assert.equal(c.state.suggestions.length, 3);
+    assert.equal(c.state.statusText, 'The tags of this image changed since the dialog opened; review them first. Saved now: night, rain.');
+    assert.equal(c.state.canSave, true);
+    answer = call => response(200, receipt(call.body.tags, { tag_revision: 6 }));
+    await c.save();
+    assert.equal(h.of('PUT', TAGS)[1].body.expected_tag_revision, 5, 'the second save is against the revision the user has seen');
+    assert.equal(c.state.phase, 'saved');
+    assert.equal(c.state.currentTags, null);
+});
+
+test('lost response keeps draft and retry resends the identical body once', async () => {
+    const answers = [() => Promise.reject(new TypeError('Failed to fetch')), () => fail(504, null, 'Gateway Timeout'),
+        () => response(200, { success: true, tags: 'broken' })];
+    for (const lost of answers) {
+        let answer = lost;
+        const { c, h } = await opened({ [`PUT ${TAGS}`]: call => answer(call) });
+        c.setManual('Minimal, dune');
+        await c.save();
+        assert.equal(c.state.phase, 'unknown-outcome');
+        assert.equal(c.state.statusText, 'The save could not be confirmed. Retry sends the same tags again.');
+        assert.equal(c.state.canRetry, true);
+        assert.equal(c.state.canSave, false);
+        assert.equal(c.state.manual, 'Minimal, dune');
+        c.setManual('changed meanwhile');
+        const gate = deferred();
+        answer = call => gate.promise.then(() => response(200, receipt(call.body.tags, { replayed: true })));
+        const retried = c.retry();
+        assert.equal(await c.retry(), false, 'a retry in flight is not sent twice');
+        gate.resolve();
+        assert.equal(await retried, true);
+        const puts = h.of('PUT', TAGS);
+        assert.equal(puts.length, 2);
+        assert.deepEqual(puts[1].body, puts[0].body);
+        assert.equal(c.state.phase, 'saved');
+        assert.equal(await c.retry(), false);
+    }
+});
+
+test('write errors keep the draft and allow another save', async () => {
+    const { c } = await opened({ [`PUT ${TAGS}`]: () => fail(400, 'tag-limit', 'An image can carry at most 100 tags.') });
+    await c.save();
+    assert.equal(c.state.phase, 'write-error');
+    assert.equal(c.state.statusText, 'An image can carry at most 100 tags.');
+    assert.equal(c.state.canSave, true);
+    assert.equal(c.state.canRetry, false);
+    const generic = await opened({ [`PUT ${TAGS}`]: () => fail(500, 'save-failed', '') });
+    await generic.c.save();
+    assert.equal(generic.c.state.phase, 'write-error');
+    assert.equal(generic.c.state.statusText, 'The tags could not be saved.');
+});
+
+test('storage-changed and file-changed disable save until reopen', async () => {
+    for (const [code, phase] of [['storage-changed', 'storage-changed'], ['file-changed', 'file-changed'], ['source-changed', 'file-changed'], ['capture-expired', 'capture-expired']]) {
+        const message = `server says ${code}`;
+        let answer = () => fail(409, code, message);
+        const { c, h } = await opened({ [`PUT ${TAGS}`]: () => answer() });
+        c.setManual('kept draft');
+        await c.save();
+        assert.equal(c.state.phase, phase);
+        assert.equal(c.state.statusText, message);
+        assert.equal(c.state.canSave, false);
+        assert.equal(c.state.canGenerate, false);
+        assert.equal(await c.save(), false);
+        assert.equal(h.of('PUT', TAGS).length, 1);
+        await c.open();
+        await settle();
+        assert.equal(c.state.phase, 'ready', 'reopen observes again');
+        assert.equal(c.state.manual, 'Minimal, warm_light');
+        assert.equal(h.of('DELETE', CAPTURE).length, 1, 'reopen releases the old capture');
+    }
+});
+
+test('video disables generate with the still-image message', async () => {
+    const reason = 'Auto-tagging is available for supported still images only.';
+    const { c, h } = await opened({
+        [`POST ${OBSERVE}`]: () => response(200, observed({ filename: 'clip.mp4' }, { media_type: 'video', generate: { supported: false, reason } })),
+    });
+    assert.equal(c.state.image.mediaType, 'video');
+    assert.equal(c.state.canGenerate, false);
+    assert.equal(c.state.canSave, true);
+    assert.equal(c.state.statusText, reason);
+    c.setModel('gguf');
+    assert.equal(c.state.statusText, reason, 'the reason survives a model change');
+    assert.equal(await c.generate(), false);
+    assert.equal(h.of('POST', SUGGEST).length, 0);
+});
+
+test('model-unavailable and model-busy set the status line only', async () => {
+    let answer = () => response(200, suggested());
+    const { c } = await opened({ [`POST ${SUGGEST}`]: () => answer() });
+    await c.generate();
+    const before = c.state;
+    for (const [code, message] of [
+        ['model-unavailable', 'WD14 SwinV2 is not downloaded. Download it from the legacy Auto-tag page.'],
+        ['model-busy', 'A tagging model is busy with another operation. Try again when it finishes.'],
+        ['unsupported-media', 'Images larger than 4096 px per side or 16,777,216 px are not supported.'],
+    ]) {
+        answer = () => fail(409, code, message);
+        assert.equal(await c.generate(), false);
+        assert.equal(c.state.statusText, message);
+        assert.deepEqual(c.state.suggestions, before.suggestions);
+        assert.equal(c.state.countText, before.countText);
+        assert.equal(c.state.primaryLabel, before.primaryLabel);
+        assert.equal(c.state.manual, before.manual);
+        assert.equal(c.state.canSave, true);
+        assert.equal(c.state.canGenerate, true);
+    }
+    answer = () => Promise.reject(new TypeError('Failed to fetch'));
+    await c.generate();
+    assert.equal(c.state.statusText, 'Auto-tagging failed. Check server logs for details.');
+    c.setModel('wd14-vit');
+    assert.equal(c.state.phase, 'ready');
+});
+
+test('a storage restore during generate requires a reopen', async () => {
+    const { c } = await opened({ [`POST ${SUGGEST}`]: () => fail(409, 'storage-changed', 'Prompt storage was restored; reopen the Tag dialog.') });
+    c.setManual('kept');
+    await c.generate();
+    assert.equal(c.state.phase, 'storage-changed');
+    assert.equal(c.state.canSave, false);
+    assert.equal(c.state.canGenerate, false);
+    await c.retry();
+    await settle();
+    assert.equal(c.state.phase, 'ready', 'Retry observes again (D116)');
+    assert.equal(c.state.manual, 'kept');
+});
+
+test('dispose aborts requests and releases the capture once', async () => {
+    const pending = deferred();
+    const { c, h } = await opened({ [`POST ${SUGGEST}`]: () => pending.promise });
+    const token = TOKEN;
+    const releases = [];
+    const generating = c.generate();
+    let notified = 0;
+    c.subscribe(() => { notified += 1; });
+    c.dispose();
+    c.dispose();
+    c.cancel();
+    for (const call of h.of('DELETE', CAPTURE)) releases.push(decodeURIComponent(call.url.slice(CAPTURE.length)));
+    assert.deepEqual(releases, [token]);
+    assert.equal(h.of('POST', SUGGEST)[0].aborted, true);
+    pending.resolve(response(200, suggested()));
+    assert.equal(await generating, false);
+    assert.equal(notified, 0, 'a disposed controller never notifies');
+    c.setManual('ignored');
+    c.remove('x');
+    assert.notEqual(c.state.manual, 'ignored');
+    assert.equal(await c.save(), false);
+    assert.equal(await c.open(), undefined);
+    assert.equal(h.of('POST', OBSERVE).length, 1);
+});
+
+test('dispose while observing aborts the models read and releases a late capture', async () => {
+    const { createImageTaggingController } = await load();
+    const late = deferred();
+    const h = harness(defaultRoutes({ [`POST ${OBSERVE}`]: () => late.promise, [`GET ${MODELS}`]: () => new Promise(() => {}) }));
+    const c = createImageTaggingController({ api: h.api, target: TARGET });
+    const opening = c.open();
+    c.dispose();
+    assert.equal(h.of('GET', MODELS)[0].aborted, true);
+    late.resolve(response(200, observed()));
+    await opening;
+    await settle();
+    assert.deepEqual(h.of('DELETE', CAPTURE).map(call => call.url), [CAPTURE + TOKEN]);
+    assert.equal(c.state.phase, 'observing');
+});
+
+test('a failing capture release is not reported as a dialog error', async () => {
+    const { c, h } = await opened({ [`DELETE ${CAPTURE}`]: () => Promise.reject(new TypeError('offline')) });
+    c.dispose();
+    await settle();
+    assert.equal(h.of('DELETE', CAPTURE).length, 1);
+    assert.equal(c.state.phase, 'ready');
+});
+
+test('the opening sentence is the concept on-screen copy and the chip hint shows only with chips (D83)', async () => {
+    const { createImageTaggingController } = await load();
+    const h = harness(defaultRoutes());
+    const c = createImageTaggingController({ api: h.api, target: TARGET });
+    assert.equal(c.state.suggestionsText, 'Generate suggestions with the selected model before approving.');
+    assert.equal(c.state.chipHint, '');
+    await c.open();
+    await settle();
+    assert.equal(c.state.suggestionsText, 'Generate suggestions with the selected model before approving.');
+    assert.equal(c.state.chipHint, '');
+    await c.generate();
+    assert.equal(c.state.chipHint, 'Remove anything you do not want to add.');
+    for (const row of [...c.state.suggestions]) c.remove(row.name);
+    assert.equal(c.state.chipHint, '');
+});
+
+test('the option fields follow the server defaults: open and a model change bump the revision, typing does not', async () => {
+    const { c, h } = await opened({
+        [`GET ${MODELS}`]: () => response(200, models({ suggestion_limit: 7, wd14_general_threshold: 0.5, wd14_character_threshold: 0.9 })),
+    });
+    const opening = c.state.optionsRevision;
+    assert.ok(opening > 0, 'the observation that applies the server defaults bumps the revision');
+    assert.deepEqual({ ...c.state.options }, { general_threshold: 0.5, character_threshold: 0.9, instructions: DEFAULT_PROMPT, suggestion_limit: 7 });
+    c.setOption('general_threshold', '0.6');
+    assert.equal(c.state.optionsRevision, opening, 'typing never re-renders the fields');
+    c.setModel('wd14-vit');
+    assert.equal(c.state.optionsRevision, opening + 1);
+    assert.equal(c.state.options.general_threshold, 0.5);
+    await c.generate();
+    assert.deepEqual(h.of('POST', SUGGEST)[0].body.options, { general_threshold: 0.5, character_threshold: 0.9, suggestion_limit: 7 });
+});
+
+test('a conflict shows the current set and merges it ahead of the draft, so a second Save never drops it unseen', async () => {
+    let answer = () => fail(409, 'tag-conflict', 'The tags of this image changed since the dialog opened; review them first.',
+        { tags: ['changed_elsewhere', 'Minimal'], tag_revision: 5 });
+    const { c, h } = await opened({ [`PUT ${TAGS}`]: call => answer(call) });
+    const draft = c.state.draftRevision;
+    c.setManual('minimal, dune');
+    await c.save();
+    assert.equal(c.state.phase, 'conflict');
+    assert.equal(c.state.statusText,
+        'The tags of this image changed since the dialog opened; review them first. Saved now: changed_elsewhere, Minimal.');
+    assert.equal(c.state.manual, 'changed_elsewhere, Minimal, dune', 'the current set leads, the draft follows, repeats dropped');
+    assert.equal(c.state.draftRevision, draft + 1, 'the dialog rewrites its field');
+    answer = call => response(200, receipt(call.body.tags, { tag_revision: 6 }));
+    await c.save();
+    const second = h.of('PUT', TAGS)[1].body;
+    assert.deepEqual(second.tags, ['changed_elsewhere', 'Minimal', 'dune']);
+    assert.equal(second.expected_tag_revision, 5);
+    const empty = await opened({ [`PUT ${TAGS}`]: () => fail(409, 'tag-conflict', 'Changed.', { tags: [], tag_revision: 9 }) });
+    empty.c.setManual('a');
+    await empty.c.save();
+    assert.equal(empty.c.state.statusText, 'Changed. Saved now: no tags.');
+    assert.equal(empty.c.state.manual, 'a');
+});
+
+test('the capture release and the save survive leaving the page (keepalive)', async () => {
+    const { c, h } = await opened();
+    await c.save();
+    assert.equal(h.of('PUT', TAGS)[0].init.keepalive, true);
+    const other = await opened();
+    other.c.dispose();
+    const [release] = other.h.of('DELETE', CAPTURE);
+    assert.equal(release.init.keepalive, true);
+    assert.equal(h.of('POST', OBSERVE)[0].init.keepalive, undefined, 'reads are not kept alive');
+});
+
+test('closing during a save releases the capture only after the PUT settles and reports the outcome', async () => {
+    const outcomes = [];
+    const gate = deferred();
+    const { c, h } = await opened({ [`PUT ${TAGS}`]: call => gate.promise.then(() => response(200, receipt(call.body.tags))) },
+        { onLateSave: outcome => outcomes.push(outcome) });
+    c.setManual('dune');
+    const saving = c.save();
+    c.dispose();
+    await settle();
+    assert.equal(h.of('DELETE', CAPTURE).length, 0, 'the capture stays while its save is in flight');
+    gate.resolve();
+    assert.equal(await saving, true);
+    await settle();
+    const order = h.calls.filter(call => call.method !== 'GET').map(call => call.method);
+    assert.deepEqual(order, ['POST', 'PUT', 'DELETE']);
+    assert.deepEqual(outcomes, [{ saved: true, filename: 'dunes_00001.png', tags: ['dune'], toast: '1 tags saved for dunes_00001.png.' }]);
+    assert.equal(c.state.closed, true);
+
+    for (const [route, message] of [
+        [() => fail(409, 'tag-conflict', 'Changed elsewhere.', { tags: ['x'], tag_revision: 5 }), 'Changed elsewhere.'],
+        [() => fail(400, 'tag-limit', 'Too many.'), 'Too many.'],
+        [() => Promise.reject(new TypeError('Failed to fetch')), 'The tag save could not be confirmed. Open Tag image again to check the saved tags.'],
+    ]) {
+        const late = [];
+        const held = deferred();
+        const run = await opened({ [`PUT ${TAGS}`]: () => held.promise.then(route) }, { onLateSave: outcome => late.push(outcome) });
+        const pending = run.c.save();
+        run.c.cancel();
+        held.resolve();
+        assert.equal(await pending, false);
+        await settle();
+        assert.deepEqual(late, [{ saved: false, message }]);
+        assert.equal(run.h.of('DELETE', CAPTURE).length, 1);
+    }
+    const quiet = await opened();
+    quiet.c.dispose();
+    assert.equal(quiet.h.of('DELETE', CAPTURE).length, 1, 'with no save in flight the capture is released at once');
+    const { createImageTaggingController } = await load();
+    assert.throws(() => createImageTaggingController({ api: h.api, target: TARGET, onLateSave: 'x' }), TypeError);
+});
+
+test('retry after a changed storage, file or expired capture observes again and keeps the draft', async () => {
+    for (const code of ['storage-changed', 'file-changed', 'capture-expired']) {
+        let revision = 3;
+        const { c, h } = await opened({
+            [`PUT ${TAGS}`]: () => fail(409, code, `server says ${code}`),
+            [`POST ${OBSERVE}`]: () => response(200, observed({ tag_revision: revision, tags: revision === 3 ? ['Minimal', 'warm_light'] : ['night'] })),
+        });
+        c.setManual('kept draft');
+        await c.save();
+        assert.equal(c.state.canRetry, true);
+        const draft = c.state.draftRevision;
+        await c.retry();
+        await settle();
+        assert.equal(c.state.phase, 'ready');
+        assert.equal(c.state.manual, 'kept draft', 'an unchanged tag set keeps the draft as typed');
+        assert.equal(c.state.draftRevision, draft + 1);
+        assert.equal(h.of('POST', OBSERVE).length, 2);
+        revision = 4;
+        await c.save();
+        await c.retry();
+        await settle();
+        assert.equal(c.state.manual, 'night, kept draft', 'a set changed meanwhile is shown ahead of the draft');
+        assert.equal(c.state.statusText, 'The tags of this image changed since the dialog opened. Saved now: night.');
+        assert.equal(c.state.tagRevision, 4);
+    }
+    // A first observation that failed has no draft to keep: its retry fills the saved tags.
+    const { createImageTaggingController } = await load();
+    let first = true;
+    const h = harness(defaultRoutes({ [`POST ${OBSERVE}`]: () => (first ? (first = false, fail(500, 'observe-failed')) : response(200, observed())) }));
+    const c = createImageTaggingController({ api: h.api, target: TARGET });
+    await c.open();
+    await c.retry();
+    await settle();
+    assert.equal(c.state.manual, 'Minimal, warm_light');
+});

tests/js/workspace-image-tagging.test.js

@@ -0,0 +1,28 @@
+import test from 'node:test';
+import assert from 'node:assert/strict';
+
+// Guarded: before the dialog module exists every test fails on an assertion, not on the import.
+const loaded = await import('../../web/js/workspace/image-tagging.mjs').catch(() => null);
+const module = () => { assert.ok(loaded, 'image-tagging.mjs is not implemented'); return loaded; };
+const SCOPE = 'a'.repeat(64);
+
+test('a listed file gives its tagging target, anything else none', () => {
+    const { imageTarget } = module();
+    assert.deepEqual(imageTarget({ root_index: 1, root_scope: SCOPE, path: 'a/b.png', filename: 'b.png' }),
+        { root_index: 1, root_scope: SCOPE, path: 'a/b.png' });
+    for (const image of [null, {}, { root_index: -1, root_scope: SCOPE, path: 'a.png' }, { root_index: 0, path: 'a.png' },
+        { root_index: 0, root_scope: 'A'.repeat(64), path: 'a.png' }, { root_index: 0, root_scope: SCOPE, path: '' },
+        { root_index: 1.5, root_scope: SCOPE, path: 'a.png' }]) {
+        assert.equal(imageTarget(image), null);
+    }
+});
+
+test('tag save listeners subscribe and unsubscribe once', () => {
+    const { onImageTagsSaved, imageTagsListenerCount } = module();
+    const before = imageTagsListenerCount();
+    const off = onImageTagsSaved(() => {});
+    assert.equal(imageTagsListenerCount(), before + 1);
+    off(); off();
+    assert.equal(imageTagsListenerCount(), before);
+    assert.throws(() => onImageTagsSaved('nope'), TypeError);
+});

tests/js/workspace-images-lifecycle.test.js

@@ -21,3 +21,12 @@ test('Images mounted after its parent is aborted creates no DOM, requests or obs
         for(const [key,descriptor] of originals) {if(descriptor) Object.defineProperty(globalThis,key,descriptor);else delete globalThis[key];}
     }
 });
+
+test('Images mounted after its parent is aborted follows no tag saves', async () => {
+    const tagging = await import('../../web/js/workspace/image-tagging.mjs').catch(() => null);
+    assert.ok(tagging, 'image-tagging.mjs is not implemented');
+    const before = tagging.imageTagsListenerCount();
+    const parent = createScope(); parent.dispose();
+    mountImages({}, {get() { assert.fail('Request started after parent disposal'); }}, parent, {embedded: true, fetchPage() {}}).dispose();
+    assert.equal(tagging.imageTagsListenerCount(), before);
+});

tests/js/workspace-images-state.test.js

@@ -352,3 +352,129 @@ test('selectionLabel counts the selected images as the concept bar does', async
     assert.equal(selectionLabel(2), '2 selected');
     assert.equal(selectionLabel(1234), '1,234 selected');
 });
+
+test('refresh re-reads the loaded range in place and publishes once, replacing the items', async () => {
+    const {createImageStream,imageKey}=await source;const requests=[];const states=[];
+    const stream=createImageStream({api:{get:(url,{signal})=>{const d=deferred();requests.push({...d,url,signal});return d.promise;}},onChange:(s,hint)=>states.push([s,hint]),batchSize:2});
+    assert.equal(typeof stream.refresh,'function','refresh is part of the stream');
+    const first=stream.reset({q:'tag'});requests[0].resolve(page([image('a'),image('b')],5,'c1'));await first;
+    const second=stream.loadMore();requests[1].resolve(page([image('c'),image('d')],5,'c2'));await second;
+    const before=stream.state;const published=states.length;
+    const refresh=stream.refresh();
+    // The current collection stays shown while the re-read runs; Load more waits for it.
+    assert.equal(stream.state,before);assert.equal(states.length,published);
+    assert.equal(await stream.loadMore(),0);assert.equal(requests.length,3);
+    const query=new URL(requests[2].url,'http://local').searchParams;
+    // The loaded range is re-read in pages as large as it (at most 200), not in the listing's batches (M2).
+    assert.equal(query.get('q'),'tag');assert.equal(query.has('cursor'),false);assert.equal(query.get('limit'),'4');
+    requests[2].resolve(page([image('a'),image('new'),image('c')],6,'n1'));
+    await new Promise(resolve=>setImmediate(resolve));
+    const next=new URL(requests[3].url,'http://local').searchParams;
+    assert.equal(next.get('cursor'),'n1');assert.equal(next.get('limit'),'2','the rest, at least one batch');
+    assert.equal(stream.state,before);
+    requests[3].resolve(page([image('d')],6,'n2'));
+    assert.equal(await refresh,4);
+    assert.deepEqual(stream.state.items.map(imageKey),['0:a','0:new','0:c','0:d']);
+    assert.deepEqual([stream.state.total,stream.state.cursor,stream.state.hasMore,stream.state.loaded,stream.state.loading],[6,'n2',true,4,false]);
+    assert.equal(states.length,published+1);assert.deepEqual(states.at(-1)[1],{replaced:true});
+    // Load more continues from the refreshed cursor.
+    const more=stream.loadMore();assert.equal(new URL(requests[4].url,'http://local').searchParams.get('cursor'),'n2');
+    requests[4].resolve(page([image('e')],6));await more;
+    assert.deepEqual(stream.state.items.map(imageKey),['0:a','0:new','0:c','0:d','0:e']);
+});
+
+test('refresh stops early when the result shrank, and a reset or failure leaves the collection as it was', async () => {
+    const {createImageStream,imageKey}=await source;const requests=[];
+    const stream=createImageStream({api:{get:(url,{signal})=>{const d=deferred();requests.push({...d,url,signal});return d.promise;}},batchSize:2});
+    assert.equal(typeof stream.refresh,'function','refresh is part of the stream');
+    const first=stream.reset({tags:['x']});requests[0].resolve(page([image('a'),image('b')],4,'c1'));await first;
+    const shrink=stream.refresh();requests[1].resolve(page([image('b')],1));
+    assert.equal(await shrink,1);assert.deepEqual(stream.state.items.map(imageKey),['0:b']);assert.equal(stream.state.hasMore,false);
+    const before=stream.state;
+    const failing=stream.refresh();requests[2].reject(new Error('offline'));
+    await assert.rejects(failing,/offline/,'a failed re-read is reported to the page');assert.equal(stream.state,before);
+    const stale=stream.refresh();const fresh=stream.reset({tags:['y']});
+    assert.equal(requests[3].signal.aborted,true);
+    requests[4].resolve(page([image('y')]));await fresh;
+    requests[3].resolve(page([image('stale')]));assert.equal(await stale,0);
+    assert.deepEqual(stream.state.items.map(imageKey),['0:y']);
+    stream.dispose();assert.equal(await stream.refresh(),0);
+});
+
+test('a refresh asked while a page loads runs once that page settles (I6)', async () => {
+    const {createImageStream,imageKey}=await source;const requests=[];const states=[];
+    const stream=createImageStream({api:{get:(url,{signal})=>{const d=deferred();requests.push({...d,url,signal});return d.promise;}},onChange:(s,hint)=>states.push(hint),batchSize:2});
+    const first=stream.reset({tags:['x']});requests[0].resolve(page([image('a'),image('b')],4,'c1'));await first;
+    for (const settle of ['resolve','reject']) {
+        const more=stream.loadMore();const sent=requests.length;
+        assert.equal(await stream.refresh(),0,'nothing is read beside the page in flight');
+        assert.equal(await stream.refresh(),0);
+        assert.equal(requests.length,sent);
+        if (settle==='resolve') requests[sent-1].resolve(page([image('c'),image('d')],4,'c2'));
+        else requests[sent-1].reject(new Error('offline'));
+        await more;
+        await new Promise(resolve=>setImmediate(resolve));
+        assert.equal(requests.length,sent+1,'exactly one pending refresh follows');
+        assert.equal(new URL(requests[sent].url,'http://local').searchParams.has('cursor'),false);
+        requests[sent].resolve(page([image('a'),image('b'),image('c'),image('d')],6,'r1'));
+        await new Promise(resolve=>setImmediate(resolve));
+        assert.deepEqual(states.at(-1),{replaced:true});
+        assert.equal(states.filter(hint=>hint?.replaced).length,settle==='resolve'?1:2);
+    }
+    // A reset drops a pending refresh.
+    const more=stream.loadMore();const sent=requests.length;
+    assert.equal(await stream.refresh(),0);
+    const fresh=stream.reset({tags:['y']});
+    requests[sent].resolve(page([image('y')]));await fresh;
+    requests[sent-1].resolve(page([image('late')],1));await more;
+    await new Promise(resolve=>setImmediate(resolve));
+    assert.equal(requests.length,sent+1);assert.deepEqual(stream.state.items.map(imageKey),['0:y']);
+    assert.equal(stream.refreshing,false);
+});
+
+test('refresh reads a long loaded range in at most 200-file pages and exposes refreshing (M2, M3)', async () => {
+    const {createImageStream}=await source;const requests=[];
+    const stream=createImageStream({api:{get:(url,{signal})=>{const d=deferred();requests.push({...d,url,signal});return d.promise;}},batchSize:150});
+    const many=(prefix,n)=>Array.from({length:n},(_,i)=>image(`${prefix}${i}`));
+    const first=stream.reset({q:'x'});requests[0].resolve(page(many('a',150),1000,'c1'));await first;
+    const second=stream.loadMore();requests[1].resolve(page(many('b',150),1000,'c2'));await second;
+    const refresh=stream.refresh();
+    assert.equal(stream.refreshing,true);
+    assert.equal(new URL(requests[2].url,'http://local').searchParams.get('limit'),'200');
+    requests[2].resolve(page(many('r',200),1000,'n1'));
+    await new Promise(resolve=>setImmediate(resolve));
+    assert.equal(new URL(requests[3].url,'http://local').searchParams.get('limit'),'150');
+    requests[3].resolve(page(many('s',150),1000,'n2'));
+    assert.equal(await refresh,350);assert.equal(requests.length,4);
+    assert.equal(stream.refreshing,false);
+});
+
+test('a second refresh supersedes the first, repeats in a re-read are dropped, and dispose abandons one in flight', async () => {
+    const {createImageStream,imageKey}=await source;const requests=[];const states=[];
+    const stream=createImageStream({api:{get:(url,{signal})=>{const d=deferred();requests.push({...d,url,signal});return d.promise;}},onChange:(s,hint)=>states.push(hint),batchSize:2});
+    const first=stream.reset({q:'x'});requests[0].resolve(page([image('a'),image('b')],2));await first;
+    const older=stream.refresh();const newer=stream.refresh();
+    assert.equal(requests[1].signal.aborted,true,'the older re-read is aborted');
+    requests[2].resolve(page([image('b'),image('b'),image('a')],2));
+    assert.equal(await newer,2);
+    requests[1].resolve(page([image('old')],1));assert.equal(await older,0);
+    assert.deepEqual(stream.state.items.map(imageKey),['0:b','0:a']);
+    assert.equal(states.filter(hint=>hint?.replaced).length,1);
+    const pending=stream.refresh();
+    stream.dispose();
+    assert.equal(requests[3].signal.aborted,true);
+    requests[3].resolve(page([image('late')],1));
+    assert.equal(await pending,0);assert.deepEqual(stream.state.items.map(imageKey),['0:b','0:a']);
+});
+
+test('a refresh deferred behind a page load is dropped when the stream is disposed meanwhile', async () => {
+    const {createImageStream}=await source;const requests=[];
+    const stream=createImageStream({api:{get:(url,{signal})=>{const d=deferred();requests.push({...d,url,signal});return d.promise;}},batchSize:2});
+    const first=stream.reset({q:'x'});requests[0].resolve(page([image('a'),image('b')],4,'c1'));await first;
+    const more=stream.loadMore();
+    assert.equal(await stream.refresh(),0);
+    stream.dispose();
+    requests[1].resolve(page([image('c')],4,'c2'));await more;
+    await new Promise(resolve=>setImmediate(resolve));
+    assert.equal(requests.length,2,'no refresh runs after dispose');
+});

tests/js/workspace-inspection-state.test.js

@@ -702,3 +702,57 @@ test('a workflow the legacy parser throws on keeps the server result instead of
     assert.deepEqual(record.prompts.positive, { status: 'resolved', text: 'exact\npositive', source: 'execution' });
     assert.equal(record.prompts.negative.status, 'ambiguous');
 });
+
+// #249 PR C Task C3: the B3 envelope's `image_tags` facet, the file's own current tags (S1).
+test('image_tags facet is kept exactly and absent stays null', () => {
+    const entry = image(), project = contract('toInspectionRecord');
+    assert.equal(project(envelope(entry), entry, extractGenerationFacts).imageTags, null);
+    for (const facet of [{ state: 'current', tags: ['dunes', 'warm_light'] }, { state: 'current', tags: [] }, { state: 'none', tags: [] }, { state: 'unavailable', tags: [] }]) {
+        const data = { ...envelope(entry), image_tags: facet };
+        const record = project(data, entry, extractGenerationFacts);
+        assert.deepEqual(record.imageTags, facet);
+        assert.notEqual(record.imageTags.tags, facet.tags);
+    }
+    const hundred = Array.from({ length: 100 }, (_, index) => `tag_${index}`);
+    assert.deepEqual(project({ ...envelope(entry), image_tags: { state: 'current', tags: hundred } }, entry, extractGenerationFacts).imageTags.tags, hundred);
+});
+
+test('rejects a malformed image_tags facet', () => {
+    const entry = image();
+    for (const facet of [null, [], 'current', { state: 'current' }, { tags: [] }, { state: 'current', tags: ['a'], extra: 1 },
+        { state: 'stale', tags: [] }, { state: 'none', tags: ['a'] }, { state: 'unavailable', tags: ['a'] }, { state: 'current', tags: 'a' },
+        { state: 'current', tags: [1] }, { state: 'current', tags: [''] }, { state: 'current', tags: [' a'] }, { state: 'current', tags: ['a,b'] },
+        { state: 'current', tags: ['a\nb'] }, { state: 'current', tags: ['x'.repeat(51)] },
+        { state: 'current', tags: Array.from({ length: 101 }, (_, index) => `tag_${index}`) }]) {
+        assert.throws(() => module.toInspectionRecord({ ...envelope(entry), image_tags: facet }, entry, extractGenerationFacts), TypeError, JSON.stringify(facet));
+    }
+    assert.equal(module.toInspectionRecord({ ...envelope(entry), image_tags: { state: 'current', tags: ['é'.repeat(50)] } }, entry, extractGenerationFacts).imageTags.tags[0].length, 50);
+});
+
+test('refresh re-reads the selected file and keeps the shown record until the new read lands', async () => {
+    const { api, requests } = pendingApi(), updates = [];
+    const controller = contract('createInspectionController')({ api, extractFacts: extractGenerationFacts });
+    assert.equal(typeof controller.refresh, 'function', 'refresh is part of the controller');
+    controller.subscribe(state => updates.push(state));
+    const load = controller.select(image()); requests[0].resolve(envelope()); await load;
+    const shown = controller.state, count = updates.length;
+    const refresh = controller.refresh();
+    assert.equal(requests.length, 2);
+    // No loading state is published: the panel keeps its content.
+    assert.equal(controller.state, shown); assert.equal(updates.length, count);
+    requests[1].resolve({ ...envelope(), image_tags: { state: 'current', tags: ['fresh'] } }); await refresh;
+    assert.deepEqual(controller.state.record.imageTags, { state: 'current', tags: ['fresh'] });
+    assert.equal(controller.state.loading, false); assert.equal(updates.length, count + 1);
+    // A failed quiet re-read reports the failure as a read would.
+    const failing = controller.refresh(); requests[2].reject(new Error('offline')); await failing;
+    assert.equal(controller.state.error, true);
+    // Without a shown record a refresh is an ordinary retry (loading published).
+    const again = controller.refresh(); assert.equal(controller.state.loading, true);
+    requests[3].resolve(envelope()); await again; assert.equal(controller.state.error, false);
+    // A newer selection supersedes a refresh in flight.
+    const stale = controller.refresh(); const other = controller.select({ ...image(), path: 'folder/b.png', filename: 'b.png', url: image().url.replace('a%20%23%25', 'b') });
+    requests[5].resolve(envelope({ ...image(), path: 'folder/b.png', filename: 'b.png', url: image().url.replace('a%20%23%25', 'b') })); await other;
+    requests[4].resolve(envelope()); await stale;
+    assert.equal(controller.state.record.identity.path, 'folder/b.png');
+    controller.dispose(); assert.equal(await controller.refresh(), undefined);
+});

tests/js/workspace-library-state.test.js

@@ -47,6 +47,22 @@ test('restored shelves keep filters, loaded parents and continuation without a f
  const next=createLibraryStream({api:{get:async url=>{urls.push(url);return page([item(2)]);}},restore:{filters:{text:'retained',sort:'name'},state:old.state}});assert.equal(next.state.items.length,1);await next.loadMore();assert.match(urls[0],/text=retained/);assert.match(urls[0],/cursor=next/);assert.equal(next.state.items.length,2);next.dispose();
  assert.throws(()=>createLibraryStream({api:{},restore:{filters:{},state:{items:[item(1),item(1)],total:2,variantTotal:0,cursor:null,hasMore:false}}}));
 });
+test('a shelf left before its first page answered or after it failed restores and reads page one',async()=>{
+ const {createLibraryStream}=await source;
+ // The shelf saves whatever state it holds when the route changes; a first page still pending or failed has no variant total yet.
+ let pending;const left=createLibraryStream({api:{get:()=>new Promise(resolve=>{pending=resolve;})}});const first=left.reset({text:'kept'});
+ assert.equal(left.state.loading,true);assert.equal(left.state.variantTotal,null);const saved=left.state;left.dispose();pending(page([item(9)]));await first;
+ const failed=createLibraryStream({api:{get:async()=>{throw Error('offline');}}});await failed.reset();assert.ok(failed.state.error);failed.dispose();
+ for(const state of [saved,failed.state]) {
+  const urls=[];const next=createLibraryStream({api:{get:async url=>{urls.push(url);return page([item(1)]);}},restore:{filters:{text:'kept'},state}});
+  assert.deepEqual([next.state.items.length,next.state.loading,next.state.error,next.state.variantTotal,next.state.hasMore],[0,false,null,null,true]);
+  assert.equal(await next.loadMore(),1);assert.equal(new URL(urls[0],'https://x').searchParams.get('text'),'kept');assert.equal(new URL(urls[0],'https://x').searchParams.has('cursor'),false);
+  assert.equal(next.state.variantTotal,6);next.dispose();
+ }
+ // A pending first page is only the empty, cursorless state: anything else without a variant total stays invalid.
+ for(const state of [{items:[item(1)],total:1,variantTotal:null,cursor:null,hasMore:false},{items:[],total:0,variantTotal:null,cursor:'next',hasMore:true},{items:[],total:3,variantTotal:null,cursor:null,hasMore:true},{items:[],total:0,variantTotal:null,cursor:null,hasMore:false}])
+  assert.throws(()=>createLibraryStream({api:{},restore:{filters:{},state}}),/Invalid library page/);
+});
 test('parent shelf window bounds DOM and pins focused rows without changing logical order',async()=>{
  const {libraryWindow}=await source;const first=libraryWindow(500,2,0,900);assert.equal(first[0],0);assert.ok(first.length<24);
  const middle=libraryWindow(500,2,50000,900,0);assert.equal(middle[0],0);assert.ok(middle.some(index=>index>150));assert.ok(middle.length<24);assert.deepEqual(middle,[...middle].sort((a,b)=>a-b));assert.deepEqual(libraryWindow(3,2,0,900),[0,1,2]);

tests/js/workspace-local-inspection.test.js

@@ -0,0 +1,308 @@
+// web/js/workspace/local-inspection.mjs turns one chosen File's local reads into
+// the inspection wire the workspace already renders, fits it into the 8 MiB
+// response budget exactly as the server does, and plugs into the shared
+// inspection controller, inspector and viewer through localFileAdapter.
+import test from "node:test";
+import assert from "node:assert/strict";
+import fs from "node:fs";
+import path from "node:path";
+import { createRequire } from "node:module";
+import { fileURLToPath } from "node:url";
+import { raw, ihdr, idat, iend, tEXt, iTXt, utf8 } from "./helpers/png-bytes.mjs";
+import { createInspectionController, toInspectionRecord, projectInspectionPayload } from "../../web/js/workspace/inspection-state.mjs";
+import { inspectionDownloads } from "../../web/js/workspace/file-inspector.mjs";
+
+const here = path.dirname(fileURLToPath(import.meta.url));
+const require = createRequire(import.meta.url);
+const PngMetadata = require("../../web/js/png-metadata.js");
+const ComfyMetadata = require("../../web/js/comfy-metadata.js");
+const { extractGenerationFacts } = ComfyMetadata;
+// The page loads both as scripts; the legacy fill and the PNG reader find them on the global.
+globalThis.PngMetadata = PngMetadata;
+globalThis.ComfyMetadata = ComfyMetadata;
+const m = await import("../../web/js/workspace/local-inspection.mjs").catch((error) => {
+    if (error.code !== "ERR_MODULE_NOT_FOUND") throw error;
+    return {};
+});
+
+const FIXTURES = path.join(here, "..", "fixtures", "local-inspect");
+const PARITY = path.join(here, "..", "fixtures", "inspection_parity");
+const PROMPT = '{"3":{"class_type":"KSampler","inputs":{"seed":7,"steps":20,"cfg":8.0,"sampler_name":"euler","scheduler":"normal","denoise":1.0,"model":["4",0],"positive":["6",0],"negative":["7",0],"latent_image":["5",0]}},"4":{"class_type":"CheckpointLoaderSimple","inputs":{"ckpt_name":"synthetic-v1.safetensors"}},"5":{"class_type":"EmptyLatentImage","inputs":{"width":512,"height":512,"batch_size":1}},"6":{"class_type":"CLIPTextEncode","inputs":{"text":"a red fox in snow, detailed fur","clip":["4",1]}},"7":{"class_type":"CLIPTextEncode","inputs":{"text":"blurry, lowres","clip":["4",1]}}}';
+const WORKFLOW = '{"last_node_id":7,"nodes":[{"id":3,"type":"KSampler","widgets_values":[7,"fixed",20,8.0,"euler","normal",1.0]}],"links":[],"version":0.4}';
+const CONTAINER_WORKFLOW = "Original workflow JSON text is unavailable for this container record. Download metadata or the original file.";
+const LIMIT = 8388608;
+const plain = (value) => JSON.parse(JSON.stringify(value));
+const bytes = (value) => Buffer.byteLength(JSON.stringify(value), "utf8");
+
+function pngFile(name, chunks, { lastModified = 1_700_000_000_123 } = {}) {
+    return new File([raw(ihdr(640, 480), idat(), ...chunks, iend())], name, { type: "image/png", lastModified });
+}
+const fixtureFile = (name, directory = FIXTURES) => new File([fs.readFileSync(path.join(directory, name))], name, { lastModified: 1_700_000_000_000 });
+const classify = async (file) => PngMetadata.classifyPrefix(new Uint8Array(await file.slice(0, 32).arrayBuffer()));
+async function handleFor(file, { url = "blob:http://host.test/11111111-2222-3333-4444-555555555555", filename } = {}) {
+    return m.createLocalFileHandle({ file, url, classification: await classify(file), filename });
+}
+async function recordFor(file, options) {
+    const handle = await handleFor(file, options);
+    const wire = await m.localFileAdapter.read(handle);
+    return { handle, wire, record: m.localFileAdapter.toRecord(wire, handle, extractGenerationFacts) };
+}
+
+test("contract", () => {
+    for (const name of ["buildLocalPayload", "fitLocalPayload", "createLocalFileHandle", "readLocalFile", "toInspectionRead", "sanitizeLocalFilename"]) {
+        assert.equal(typeof m[name], "function", `${name} is exported`);
+    }
+    assert.equal(typeof m.localFileAdapter, "object");
+    for (const name of ["validate", "key", "scope", "inspectionPath", "read", "toRecord", "originalDownload", "mediaUrl"]) {
+        assert.equal(typeof m.localFileAdapter[name], "function", `adapter.${name}`);
+    }
+    assert.ok(Object.isFrozen(m.localFileAdapter));
+});
+
+test("a PNG File becomes the server-shaped wire with local provenance", async () => {
+    const file = pngFile("Fox Render.PNG", [tEXt("prompt", PROMPT), tEXt("workflow", WORKFLOW)]);
+    const { wire, record, handle } = await recordFor(file);
+    assert.deepEqual(wire.file, { observation: "selected-local-file", source_label: "Selected local file", media_type: "image", extension: ".png", format: "PNG",
+        size: file.size, width: 640, height: 480, modified_time: 1_700_000_000.123 });
+    assert.equal(wire.success, true);
+    assert.equal(wire.inspection_version, 1);
+    assert.deepEqual(wire.prompts, { positive: { status: "unresolved" }, negative: { status: "unresolved" } });
+    assert.deepEqual(wire.resolution, { state: "unresolved", sampler_id: null, reason: "Resolved in this browser from the workflow.", reason_code: null, node: null });
+    assert.deepEqual(wire.library, { status: "unavailable", scope: "saved-library", tags: [], category_scope: null, tags_scope: null });
+    assert.equal(wire.metadata.status, "available");
+    assert.deepEqual(wire.metadata.raw, { prompt: PROMPT, workflow: WORKFLOW });
+    assert.deepEqual(wire.metadata.record_formats, { prompt: "original-text", workflow: "original-text" });
+    assert.equal(wire.metadata.prompt_graph.status, "parsed");
+    assert.equal(plain(wire.metadata.prompt_graph.value)["3"].inputs.cfg, "8.0");
+    assert.equal(wire.metadata.workflow_graph.status, "parsed");
+    assert.equal(wire.metadata.parameters, null);
+    for (const key of Object.keys(wire)) assert.ok(["success", "inspection_version", "file", "prompts", "resolution", "metadata", "library"].includes(key), key);
+    // The existing projection renders it: facts from the API graph, legacy fill for the roles, no library link.
+    assert.deepEqual(record.identity, { kind: "local-file", key: handle.key, path: "Fox Render.PNG" });
+    assert.equal(record.file.observation, "selected-local-file");
+    assert.equal(record.generation.seed.value, "7");
+    assert.equal(record.generation.cfg.value, "8.0");
+    assert.equal(record.prompts.positive.text, "a red fox in snow, detailed fur");
+    assert.equal(record.prompts.positive.source, "graph-inferred");
+    assert.equal(record.prompts.negative.text, "blurry, lowres");
+    assert.equal(record.historyTarget, null);
+    assert.deepEqual(record.download, { url: "blob:http://host.test/11111111-2222-3333-4444-555555555555", filename: "Fox Render.PNG" });
+    const downloads = inspectionDownloads(record);
+    assert.equal(downloads.metadata.available, true);
+    assert.equal(downloads.workflow.bodyText, WORKFLOW);
+    assert.equal(downloads.metadata.filename, "Fox Render.metadata.json");
+    const exported = JSON.parse(downloads.metadata.bodyText);
+    assert.equal(exported.identity.kind, "local-file");
+    assert.equal(exported.file.source_label, "Selected local file");
+});
+
+test("file facts follow the server: extension like Path.suffix, null when unknown", async () => {
+    const cases = [["a.b.PnG", ".png"], [".hidden", ""], ["noext", ""], ["dot.", ""], ["x.tar.gz", ".gz"]];
+    for (const [name, extension] of cases) {
+        const wire = m.buildLocalPayload({ read: { status: "complete", records: {}, limited: [], width: 2, height: 3 }, file: { name, size: 10, lastModified: 0 }, classification: { media: "image", kind: "png" } });
+        assert.equal(wire.file.extension, extension, name);
+        assert.equal(wire.file.modified_time, 0);
+    }
+    const sparse = m.buildLocalPayload({ read: { status: "complete", records: {}, limited: [], width: 2, height: 3 }, file: { name: "a.png" }, classification: { media: "image", kind: "png" } });
+    assert.deepEqual([sparse.file.size, sparse.file.modified_time], [null, null]);
+    assert.equal(m.buildLocalPayload({ read: { status: "complete", records: {}, limited: [], width: 2, height: 3 }, file: { name: "a.png", size: -1, lastModified: NaN }, classification: { media: "image", kind: "png" } }).file.size, null);
+    assert.throws(() => m.buildLocalPayload({ read: null, file: { name: "a" }, classification: null }), TypeError);
+    assert.throws(() => m.buildLocalPayload({ read: null, file: { name: "a" }, classification: { media: "text", kind: "svg" } }), TypeError);
+});
+
+test("sanitized filenames carry no separators or control characters", () => {
+    assert.equal(m.sanitizeLocalFilename("../a/b\\c\u0000\u001f\u007f.png"), ".._a_b_c.png");
+    assert.equal(m.sanitizeLocalFilename("  spaced .png  "), "spaced .png");
+    assert.equal(m.sanitizeLocalFilename(""), "file");
+    assert.equal(m.sanitizeLocalFilename(undefined), "file");
+    assert.equal(m.sanitizeLocalFilename("\u0000"), "file");
+    assert.equal([...m.sanitizeLocalFilename("😀".repeat(300))].length, 255);
+});
+
+test("unreadable PNG metadata is unreadable, not absent", async () => {
+    const { wire, record } = await recordFor(pngFile("bad.png", [tEXt("prompt", PROMPT, { crc: 0 })]));
+    assert.equal(wire.metadata.status, "unreadable");
+    assert.deepEqual([wire.file.format, wire.file.width, wire.file.height], [null, null, null]);
+    assert.equal(record.generation.seed.reason, "metadata-unreadable");
+    assert.equal(record.prompts.positive.status, "unresolved");
+    assert.equal(JSON.parse(inspectionDownloads(record).metadata.bodyText).metadata.status, "unreadable");
+    const limited = await recordFor(pngFile("big.png", [tEXt("prompt", PROMPT), tEXt("workflow", "[" + "0,".repeat(2_097_153) + "]")]));
+    assert.deepEqual(limited.wire.metadata.record_formats, { prompt: "original-text", workflow: "limited" });
+    assert.equal(limited.wire.metadata.workflow_graph.status, "limited");
+    assert.equal(limited.record.metadata.record_formats.workflow, "limited");
+});
+
+test("JPEG and WebP previews have header facts and absent metadata", async () => {
+    for (const [name, format, width, height] of [["tiny.jpg", "JPEG", 48, 32], ["tiny.webp", "WEBP", 40, 24]]) {
+        const { wire, record } = await recordFor(fixtureFile(name));
+        assert.deepEqual([wire.file.format, wire.file.width, wire.file.height, wire.file.media_type], [format, width, height, "image"], name);
+        assert.equal(wire.metadata.status, "absent");
+        assert.equal(record.generation.seed.status, "not-recorded");
+    }
+});
+
+test("video payloads", async () => {
+    const legacy = await recordFor(fixtureFile("legacy-comment.mp4"));
+    assert.deepEqual([legacy.wire.file.media_type, legacy.wire.file.format, legacy.wire.file.width, legacy.wire.file.extension], ["video", null, null, ".mp4"]);
+    assert.equal(legacy.wire.metadata.record_formats.workflow, "reserialized-json");
+    assert.equal(legacy.wire.metadata.record_formats.prompt, "original-text");
+    assert.equal(legacy.wire.metadata.workflow_graph.status, "parsed");
+    assert.equal(inspectionDownloads(legacy.record).workflow.available, false);
+    assert.equal(inspectionDownloads(legacy.record).workflow.reason, CONTAINER_WORKFLOW);
+    assert.equal(legacy.record.prompts.positive.text, "a red fox in snow, detailed fur");
+    // Ruling M1: a Matroska COMMENT tag is not the exact-case `comment` envelope, so it is absent, as Images reports it.
+    const webm = await recordFor(fixtureFile("legacy-comment.webm"));
+    assert.equal(webm.wire.metadata.status, "absent");
+    const tags = await recordFor(fixtureFile("core-tags.webm"));
+    assert.deepEqual(tags.wire.metadata.record_formats, { prompt: "original-text", workflow: "original-text" });
+    assert.equal(tags.record.generation.seed.value, "7");
+    // AVI and ASF are recognised, previewable, and never read.
+    const avi = new File([utf8("RIFF"), new Uint8Array([0, 0, 0, 0]), utf8("AVI LIST"), new Uint8Array(64)], "clip.avi");
+    const aviRecord = await recordFor(avi);
+    assert.equal(aviRecord.wire.metadata.status, "unreadable");
+    assert.equal(aviRecord.wire.file.media_type, "video");
+    assert.equal(aviRecord.record.generation.seed.reason, "metadata-unreadable");
+    const asf = new File([new Uint8Array([0x30, 0x26, 0xb2, 0x75, 0x8e, 0x66, 0xcf, 0x11, 0xa6, 0xd9, 0x00, 0xaa, 0x00, 0x62, 0xce, 0x6c]), new Uint8Array(64)], "clip.wmv");
+    assert.equal((await recordFor(asf)).wire.metadata.status, "unreadable");
+    // A pre-classified video needs no PNG reader; an image without one is an error, not a silent absence.
+    const clip = fixtureFile("core.mkv");
+    assert.equal((await m.readLocalFile(clip, { classification: { media: "video", kind: "matroska" }, pngMetadata: null })).result.read.status, "complete");
+    await assert.rejects(m.readLocalFile(clip, { pngMetadata: null }), /PNG reader/);
+    await assert.rejects(m.readLocalFile(fixtureFile("tiny.jpg"), { classification: { media: "image", kind: "jpeg" }, pngMetadata: null }), /PNG reader/);
+    await assert.rejects(m.readLocalFile("not a blob"), TypeError);
+    assert.deepEqual(await m.readLocalFile(new File([utf8("<svg/>")], "x.png"), { pngMetadata: PngMetadata }), { classification: null, result: null });
+    // A container that fails to parse is unreadable too.
+    const broken = new File([utf8("\0\0\0\x08ftypisom"), utf8("\0\0\0\x01moov")], "broken.mp4");
+    assert.equal((await recordFor(broken)).wire.metadata.status, "unreadable");
+});
+
+test("local cap measures escaped UTF-8 atomically", async () => {
+    // A 1 MiB Latin-1 prompt of 'é' (2 UTF-8 bytes, 1 code unit each) and a 4 MiB UTF-8 workflow of
+    // emoji (4 bytes, 2 units): the UTF-16 length of the wire is well under the limit while its UTF-8
+    // size is not, so a cap counting code units would keep too much.
+    const prompt = `{"1":{"class_type":"A","inputs":{"t":"${"é".repeat(1_000_000)}"}}}`;
+    const workflow = `{"nodes":[],"u":"${"😀".repeat(1_000_000)}"}`;
+    const file = pngFile("wide.png", [iTXt("workflow", workflow), tEXt("prompt", prompt)]);
+    const handle = await handleFor(file);
+    const raw = await m.readLocalFile(handle.file, { classification: handle.classification, pngMetadata: PngMetadata });
+    assert.equal(raw.result.status, "complete");
+    const wire = m.buildLocalPayload({ read: raw.result, file: { name: file.name, size: file.size, lastModified: file.lastModified }, classification: handle.classification });
+    const text = JSON.stringify(wire);
+    assert.ok(text.length < LIMIT && bytes(wire) > LIMIT, `${text.length} code units, ${bytes(wire)} bytes`);
+    const fitted = await m.fitLocalPayload(wire);
+    assert.ok(bytes(fitted) <= LIMIT);
+    assert.equal(fitted.metadata.workflow_graph.status, "limited", "the workflow graph goes first");
+    assert.equal(fitted.metadata.prompt_graph.status, "parsed");
+    assert.deepEqual(Object.keys(fitted.metadata.raw).sort(), ["prompt", "workflow"], "raw is kept");
+    const record = m.localFileAdapter.toRecord(fitted, handle, extractGenerationFacts);
+    assert.equal(inspectionDownloads(record).metadata.available, true);
+    assert.equal(inspectionDownloads(record).workflow.available, false);
+    // Whole-raw omission is atomic and marks the response-size limit; the inspector offers no metadata download.
+    const squeezed = await m.fitLocalPayload(wire, bytes(fitted) - 1);
+    assert.deepEqual([squeezed.metadata.workflow_graph.status, squeezed.metadata.prompt_graph.status], ["limited", "limited"]);
+    const tiny = await m.fitLocalPayload(wire, 2000);
+    assert.equal(tiny.metadata.status, "limited");
+    assert.equal(tiny.metadata.omitted_reason, "response-size-limit");
+    assert.deepEqual(tiny.metadata.raw, {});
+    assert.equal(tiny.metadata.parameters, null);
+    assert.ok(bytes(tiny) <= 2000);
+    const omitted = m.localFileAdapter.toRecord(tiny, handle, extractGenerationFacts);
+    assert.equal(inspectionDownloads(omitted).metadata.available, false);
+    assert.equal(omitted.generation.seed.reason, "projection-limited");
+    await assert.rejects(m.fitLocalPayload(wire, 100), /exceeds/);
+    assert.equal(wire.metadata.status, "available", "the input wire is not mutated");
+    assert.equal((await m.localFileAdapter.read(handle)).metadata.workflow_graph.status, "limited", "adapter.read returns the fitted wire");
+});
+
+test("adapter rejects foreign blob URLs and never resolves owned ones", async () => {
+    const file = pngFile("a.png", [tEXt("prompt", PROMPT)]);
+    const classification = await classify(file);
+    const owned = m.createLocalFileHandle({ file, url: "blob:http://host.test/0b0e", classification });
+    assert.ok(Object.isFrozen(owned));
+    assert.deepEqual(m.localFileAdapter.originalDownload(owned), { url: "blob:http://host.test/0b0e", filename: "a.png" });
+    assert.equal(m.localFileAdapter.mediaUrl(owned), "blob:http://host.test/0b0e");
+    // The controller and the viewer session spread-copy descriptors: a copy is the same selection.
+    const copy = { ...owned, occurrence: {}, reference: {} };
+    assert.equal(m.localFileAdapter.key(copy), owned.key);
+    assert.equal(m.localFileAdapter.mediaUrl(copy), "blob:http://host.test/0b0e");
+    // A tampered URL, key, File or name, or a descriptor this module never made, is foreign.
+    const tampered = [{ ...owned, url: "blob:http://host.test/other" }, { ...owned, key: "local:1" }, { ...owned, file: new File([], "x.png") }, { ...owned, filename: "other.png" },
+        { kind: "local-file", key: owned.key, filename: owned.filename, url: owned.url, file, is_video: false, classification }, null, "local:1"];
+    for (const foreign of tampered) {
+        for (const method of ["validate", "originalDownload", "mediaUrl", "key", "scope", "inspectionPath"]) assert.throws(() => m.localFileAdapter[method](foreign), TypeError, method);
+        await assert.rejects(m.localFileAdapter.read(foreign), TypeError);
+        assert.throws(() => m.localFileAdapter.toRecord({ success: true, inspection_version: 1 }, foreign, extractGenerationFacts), TypeError);
+    }
+    for (const url of ["http://host.test/x.png", "/prompt_manager/images/serve/a.png", "javascript:alert(1)", "data:image/png;base64,AA==", "", 5]) {
+        assert.throws(() => m.createLocalFileHandle({ file, url, classification }), TypeError, String(url));
+    }
+    assert.throws(() => m.createLocalFileHandle({ file: "not a blob", url: "blob:http://host.test/1", classification }), TypeError);
+    assert.throws(() => m.createLocalFileHandle({ file, url: "blob:http://host.test/1", classification: null }), TypeError);
+    assert.throws(() => m.createLocalFileHandle({ file, url: "blob:http://host.test/1", classification: { media: "text", kind: "svg" } }), TypeError);
+    // A proxy-mounted host resolves API URLs; blob URLs are session-owned and bypass it.
+    const before = Object.getOwnPropertyDescriptor(globalThis, "window");
+    Object.defineProperty(globalThis, "window", { configurable: true, value: { location: { href: "https://example.test/proxy/prompt_manager/workspace" }, ApiClient: { resolveUrl: (url) => `/proxy${url}` } } });
+    try {
+        assert.equal(m.localFileAdapter.mediaUrl(owned), "blob:http://host.test/0b0e");
+        assert.equal(m.localFileAdapter.originalDownload(owned).url, "blob:http://host.test/0b0e");
+    } finally {
+        if (before) Object.defineProperty(globalThis, "window", before); else delete globalThis.window;
+    }
+    // Without an object URL the record has no download and the viewer cannot show the original.
+    const previewless = m.createLocalFileHandle({ file, url: null, classification });
+    assert.equal(m.localFileAdapter.mediaUrl(previewless), null);
+    assert.throws(() => m.localFileAdapter.originalDownload(previewless), TypeError);
+    const record = m.localFileAdapter.toRecord(await m.localFileAdapter.read(previewless), previewless, extractGenerationFacts);
+    assert.equal(record.download, null);
+    assert.equal(inspectionDownloads(record).metadata.filename, "a.metadata.json");
+});
+
+test("select reuses the same local key", async () => {
+    const file = pngFile("a.png", [tEXt("prompt", PROMPT)]);
+    const classification = await classify(file);
+    const first = m.createLocalFileHandle({ file, url: "blob:http://host.test/1", classification });
+    const second = m.createLocalFileHandle({ file, url: "blob:http://host.test/2", classification });
+    assert.notEqual(first.key, second.key, "the same bytes chosen twice are two selections");
+    assert.equal(m.localFileAdapter.scope(first), "local");
+    assert.equal(m.localFileAdapter.inspectionPath(first), `local-file:${first.key}`);
+    assert.equal(m.localFileAdapter.key(first), first.key);
+    const reads = [];
+    // Until the loader seam lands (PR L2 Task 4) the controller reaches the adapter through api.get.
+    const api = { get: (sentinel) => { reads.push(sentinel); return m.localFileAdapter.read(reads.length === 1 || sentinel === `local-file:${first.key}` ? first : second); } };
+    const controller = createInspectionController({ api, extractFacts: extractGenerationFacts, adapter: m.localFileAdapter });
+    const pending = controller.select(first);
+    assert.equal(controller.select(first), pending, "the same handle returns the pending read");
+    await pending;
+    assert.deepEqual(reads, [`local-file:${first.key}`]);
+    assert.equal(controller.state.record.identity.key, first.key);
+    assert.equal(controller.state.record.prompts.positive.text, "a red fox in snow, detailed fur");
+    await controller.select(second);
+    assert.equal(reads.length, 2);
+    assert.equal(controller.state.record.identity.key, second.key);
+    await controller.select({ ...second, occurrence: {}, reference: {} });
+    assert.equal(reads.length, 2, "the controller's own copy is the same selection");
+    await controller.select({ ...second, url: "blob:http://host.test/evil" });
+    assert.equal(controller.state.error, true, "a foreign handle is published as an error, never read");
+    assert.equal(reads.length, 2);
+    controller.dispose();
+});
+
+test("fileFor admits selected-local-file only for a local identity", async () => {
+    const file = pngFile("a.png", [tEXt("prompt", PROMPT)]);
+    const { wire, handle } = await recordFor(file);
+    const rooted = { root_index: 0, path: "folder/a.png", filename: "a.png", url: "/prompt_manager/images/serve/folder/a.png?root=0" };
+    assert.throws(() => toInspectionRecord({ ...wire, image: { root_index: 0, path: "folder/a.png", filename: "a.png" } }, rooted, extractGenerationFacts), /Invalid inspection file/);
+    assert.throws(() => projectInspectionPayload({ ...wire, file: { ...wire.file, observation: "current-file" } }, { kind: "local-file", key: handle.key, path: "a.png" }, null, extractGenerationFacts), /Invalid inspection file/);
+    assert.throws(() => m.localFileAdapter.toRecord({ ...wire, success: false }, handle, extractGenerationFacts), TypeError);
+    assert.throws(() => m.localFileAdapter.toRecord({ ...wire, inspection_version: 2 }, handle, extractGenerationFacts), TypeError);
+    assert.throws(() => m.localFileAdapter.toRecord(null, handle, extractGenerationFacts), TypeError);
+});
+
+test("the parity corpus PNGs read into renderable records", async () => {
+    for (const name of ["f02.png", "f06a.png", "f23.png"]) {
+        const { wire, record } = await recordFor(fixtureFile(name, PARITY));
+        assert.equal(wire.metadata.status, "available", name);
+        assert.equal(record.file.format, "PNG");
+        assert.ok(["parsed", "limited", "malformed"].includes(wire.metadata.prompt_graph.status));
+    }
+});

tests/js/workspace-local-projection.test.js

@@ -0,0 +1,249 @@
+// web/js/workspace/local-projection.mjs ports utils/inspection_response.py: the
+// strict Decimal/duplicate-key parse, the prompt/workflow/carrier projection, the
+// depth and node ceilings, _metadata and the 8 MiB omission steps. These units pin
+// the rules the plan names; tests/test_local_inspect_parity.py proves byte parity
+// with the server over the committed corpus.
+import test from "node:test";
+import assert from "node:assert/strict";
+import fs from "node:fs";
+import path from "node:path";
+import { fileURLToPath } from "node:url";
+
+const here = path.dirname(fileURLToPath(import.meta.url));
+const m = await import("../../web/js/workspace/local-projection.mjs").catch((error) => {
+    if (error.code !== "ERR_MODULE_NOT_FOUND") throw error;
+    return {};
+});
+const CASES = JSON.parse(fs.readFileSync(path.join(here, "..", "fixtures", "local-inspect", "projection-cases.json"), "utf8")).cases;
+
+const plain = (value) => JSON.parse(JSON.stringify(value));
+const bytes = (value) => Buffer.byteLength(JSON.stringify(value), "utf8");
+const graphOf = (text, kind = "workflow") => plain(m.projectGraph(text, kind));
+const nested = (depth, close = "]", open = "[") => open.repeat(depth) + close.repeat(depth);
+const nodes = (count) => `{${Array.from({ length: count }, (_, i) => `"${i}":{"class_type":"A","inputs":{}}`).join(",")}}`;
+
+test("contract", () => {
+    for (const name of ["projectGraph", "projectMetadata", "fitInspectionPayload", "inspectionBytes", "parseStrictJson"]) {
+        assert.equal(typeof m[name], "function", `${name} is exported`);
+    }
+    assert.equal(m.MAX_RESPONSE_BYTES, 8388608);
+    assert.equal(m.MAX_PROJECTION_DEPTH, 64);
+    assert.equal(m.GRAPH_NODE_CEILING, 20000);
+});
+
+test("projection spells numbers like the server", () => {
+    // literal -> Python str(Decimal(literal)) (or the int when it is a safe integer)
+    const table = [
+        ["1e3", "1E+3"], ["1e0", "1"], ["-0.0", "-0.0"], ["6.50", "6.50"], ["0.0100", "0.0100"], ["1.0e2", "1.0E+2"],
+        ["0e5", "0E+5"], ["0.0000001", "1E-7"], ["0.000001", "0.000001"], ["1E+3", "1E+3"], ["1e-0", "1"], ["0.5e1", "5"],
+        ["0.05e1", "0.5"], ["0.0e0", "0.0"], ["1.5e+300", "1.5E+300"], ["123e-8", "0.00000123"], ["2.5e-3", "0.0025"],
+        ["1.0", "1.0"], ["-1.50", "-1.50"], ["0.0", "0.0"], ["100.0", "100.0"], ["1234567e-6", "1.234567"], ["0.5E-6", "5E-7"],
+        ["0.5E-5", "0.000005"], ["1e22", "1E+22"], ["0.000000", "0.000000"], ["-0.000000e-3", "-0E-9"], ["7.0e-6", "0.0000070"],
+        ["7.0e-7", "7.0E-7"], ["12345678901234567890.123", "12345678901234567890.123"], ["9007199254740992.0", "9007199254740992.0"],
+        ["1e999999999999999999", "1E+999999999999999999"], ["1e-1999999999999999997", "1E-1999999999999999997"],
+        ["0.0e1000000000000000000", "0E+999999999999999999"], ["1.5e999999999999999999", "1.5E+999999999999999999"],
+        ["0.1e-1999999999999999996", "1E-1999999999999999997"],
+        ["9007199254740991", 9007199254740991], ["-9007199254740991", -9007199254740991], ["9007199254740992", "9007199254740992"],
+        ["-9007199254740992", "-9007199254740992"], ["10000000000000000", "10000000000000000"], ["123456789012345678901234567890", "123456789012345678901234567890"],
+        ["0", 0], ["-0", 0], ["42", 42], ["-7", -7],
+    ];
+    for (const [literal, expected] of table) {
+        const graph = graphOf(`{"n":${literal}}`);
+        assert.equal(graph.status, "parsed", literal);
+        assert.deepEqual(graph.value, { n: expected }, literal);
+    }
+    assert.deepEqual(graphOf('{"n":[NaN,Infinity,-Infinity]}').value, { n: [null, null, null] });
+    // A 4300-digit integer is Python's int() limit; one more digit fails the parse.
+    assert.deepEqual(graphOf(`{"n":${"1".repeat(4300)}}`).value, { n: "1".repeat(4300) });
+    assert.equal(graphOf(`{"n":${"1".repeat(4301)}}`).status, "malformed");
+    assert.equal(graphOf(`{"n":-${"1".repeat(4301)}}`).status, "malformed");
+    // Exact Decimal construction fails past the C library's exponent range.
+    for (const literal of ["1e1000000000000000000", "15e999999999999999999", "1e-1999999999999999998", "1.5e-1999999999999999997",
+        "0e1000000000000000000", "1e9223372036854775808", "0.0e1000000000000000001", "1e99999999999999999999999999"]) {
+        assert.equal(graphOf(`{"n":${literal}}`).status, "malformed", literal);
+    }
+    // Escapes: surrogate pairs join, lone surrogates stay, control characters decode.
+    const text = graphOf('{"s":"\\u00E9\\uD83D\\uDE00\\uD800\\u0000\\u001f\\/\\\\\\"\\b\\f\\n\\r\\t\\uD800\\u0041"}').value.s;
+    assert.equal(text, "é\u{1f600}\ud800\u0000\u001f/\\\"\b\f\n\r\t\ud800A");
+    assert.equal(JSON.stringify(m.projectGraph('{"s":"\\ud800"}', "workflow").value), '{"s":"\\ud800"}');
+});
+
+test("projection rejects what Python's json rejects and nothing else", () => {
+    const rejected = ["{}", "\f1", "  1", '{"a":1,}', "[1,]", "{'a':1}", '{"a"}', "[1 2]", "tru", "nul", "-NaN", "1.", ".5", "01", "+1",
+        "1e", "-", '"\\x"', '"abc', "", "  ", '"\u0001"', '{"a":1}x', '"\\u12"', '{"a":1,"a":2}', '{"a":{"b":1,"b":1}}', "[1,[2,{\"k\":1,\"k\":2}]]",
+        '{"a":1 "b":2}', "[,1]", "{,}", '{"a":}', "[1}", "{1:2}", "
[]", "[]\u000b", "0x10", "1_000", "--1", "1e+", "1E-", '"\\uD800\\uDC"'];
+    for (const text of rejected) assert.equal(graphOf(text).status, "malformed", JSON.stringify(text));
+    for (const text of [" \t\r\n{}\n\t ", '{"a":[1,{"b":null}],"c":"d"}', '{"\\u0000":1}', '{"a":1,"A":1}', '{"":[]}']) {
+        assert.equal(graphOf(text).status, "parsed", JSON.stringify(text));
+    }
+    // Top-level values that are not objects parse but are not graphs.
+    for (const text of ["[]", '"x"', "1", "null", "true", "NaN", "1.5", "-Infinity"]) assert.equal(graphOf(text).status, "malformed", text);
+    assert.equal(graphOf(undefined).status, "malformed");
+    assert.equal(graphOf(null).status, "malformed");
+});
+
+test("projection limits", () => {
+    // A workflow's root is depth 1; the 64th nested container is kept, the 65th is limited.
+    assert.equal(graphOf(`{"a":${nested(63)}}`).status, "parsed");
+    assert.equal(graphOf(`{"a":${nested(64)}}`).status, "limited");
+    // Prompt: root 1, node 2, inputs 3; a value chain reaching depth 64 is kept, 65 is limited.
+    assert.equal(graphOf(`{"1":{"class_type":"A","inputs":{"v":${nested(61)}}}}`, "prompt").status, "parsed");
+    assert.equal(graphOf(`{"1":{"class_type":"A","inputs":{"v":${nested(62)}}}}`, "prompt").status, "limited");
+    // Dropped node keys are still validated for depth.
+    assert.equal(graphOf(`{"1":{"class_type":"A","inputs":{},"_meta":{"v":${nested(61)}}}}`, "prompt").status, "parsed");
+    assert.equal(graphOf(`{"1":{"class_type":"A","inputs":{},"_meta":{"v":${nested(62)}}}}`, "prompt").status, "limited");
+    // A workflow is normalized before its shape is checked, so a deep non-object is limited; a
+    // prompt checks its root first. Depth never overflows the parser (100,000 nested arrays).
+    assert.equal(graphOf(nested(10)).status, "malformed");
+    assert.equal(graphOf(nested(5000)).status, "limited");
+    assert.equal(graphOf(nested(5000), "prompt").status, "malformed");
+    assert.equal(graphOf(`{"a":${nested(100000)}}`).status, "limited");
+    // Node ceilings: 20,000 API nodes parse, 20,001 are limited.
+    assert.equal(graphOf(nodes(20000), "prompt").status, "parsed");
+    assert.equal(graphOf(nodes(20001), "prompt").status, "limited");
+    const list = (count) => `[${Array.from({ length: count }, () => "{}").join(",")}]`;
+    assert.equal(graphOf(`{"nodes":${list(20000)}}`).status, "parsed");
+    assert.equal(graphOf(`{"nodes":${list(20001)}}`).status, "limited");
+    assert.equal(graphOf(`{"nodes":${list(10001)},"definitions":{"subgraphs":[{"nodes":${list(10000)}},5,{"nodes":3}]}}`).status, "limited");
+    assert.equal(graphOf(`{"nodes":${list(10000)},"definitions":{"subgraphs":[{"nodes":${list(10000)}}]}}`).status, "parsed");
+    assert.equal(graphOf(`{"nodes":{"${"x".repeat(5)}":1},"definitions":[]}`).status, "parsed");
+    // Duplicate keys anywhere are malformed, including inside dropped node keys.
+    assert.equal(graphOf('{"1":{"class_type":"A","inputs":{},"_meta":{"t":1,"t":2}}}', "prompt").status, "malformed");
+    assert.equal(graphOf('{"1":{"class_type":"A","inputs":{"a":1,"a":1}}}', "prompt").status, "malformed");
+});
+
+test("prompt nodes keep only class_type and inputs on null-prototype objects", () => {
+    const text = '{"3":{"class_type":"KSampler","inputs":{"seed":9007199254740993,"cfg":8.0,"n":NaN},"_meta":{"title":"x"},"is_changed":[1]},"4":5,"5":[1,2.50],"__proto__":{"class_type":"P","inputs":{}},"constructor":"c"}';
+    const graph = m.projectGraph(text, "prompt");
+    assert.equal(graph.status, "parsed");
+    assert.equal(Object.getPrototypeOf(graph.value), null);
+    assert.equal(Object.getPrototypeOf(graph.value["3"]), null);
+    assert.equal(Object.getPrototypeOf(graph.value["3"].inputs), null);
+    // JSON.parse, like the projection, makes "__proto__" an own key; an object literal would not.
+    assert.deepEqual(plain(graph.value), JSON.parse('{"3":{"class_type":"KSampler","inputs":{"seed":"9007199254740993","cfg":"8.0","n":null}},"4":5,"5":[1,"2.50"],"__proto__":{"class_type":"P","inputs":{}},"constructor":"c"}'));
+    assert.ok(Object.hasOwn(graph.value, "__proto__"));
+    assert.deepEqual(plain(graph.value["__proto__"]), { class_type: "P", inputs: {} });
+    assert.equal(graph.value.constructor, "c");
+    for (const text of ["[]", '"3"', "5", "null", '{"nodes":[]}x']) assert.equal(m.projectGraph(text, "prompt").status, text === '{"nodes":[]}x' ? "malformed" : "malformed", text);
+    // A workflow keeps everything, normalized.
+    assert.deepEqual(graphOf('{"nodes":[{"id":1,"widgets_values":[7,"fixed",8.0,1e3]}],"links":[],"version":0.4}').value,
+        { nodes: [{ id: 1, widgets_values: [7, "fixed", "8.0", "1E+3"] }], links: [], version: "0.4" });
+});
+
+test("projectMetadata mirrors _metadata for records, formats and carriers", () => {
+    const prompt = '{"1":{"class_type":"A","inputs":{"x":1.0}}}', workflow = '{"nodes":[],"version":0.4}';
+    const read = (records, record_formats, carrierText = null, limited = []) => ({ records, record_formats, carrierText, limited });
+    assert.deepEqual(plain(m.projectMetadata(null)), { status: "unreadable", omitted_reason: null, raw: {}, record_formats: {}, parameters: null,
+        prompt_graph: { status: "absent", value: null }, workflow_graph: { status: "absent", value: null } });
+    assert.deepEqual(plain(m.projectMetadata(read({}, {}))), { status: "absent", omitted_reason: null, raw: {}, record_formats: {}, parameters: null,
+        prompt_graph: { status: "absent", value: null }, workflow_graph: { status: "absent", value: null } });
+    const full = plain(m.projectMetadata(read({ prompt, workflow, parameters: "cat\nSteps: 1", comfyassets_prompt_pairs: "{}", other: "x" },
+        { prompt: "original-text", workflow: "original-text", parameters: "original-text", comfyassets_prompt_pairs: "original-text", other: "original-text" })));
+    assert.deepEqual(full, { status: "available", omitted_reason: null,
+        raw: { prompt, workflow, parameters: "cat\nSteps: 1", comfyassets_prompt_pairs: "{}" },
+        record_formats: { prompt: "original-text", workflow: "original-text", parameters: "original-text", comfyassets_prompt_pairs: "original-text" },
+        parameters: "cat\nSteps: 1", prompt_graph: { status: "parsed", value: { 1: { class_type: "A", inputs: { x: "1.0" } } } },
+        workflow_graph: { status: "parsed", value: { nodes: [], version: "0.4" } } });
+    // limited records lose their text; invalid is malformed; reserialized without a carrier is limited.
+    const limited = plain(m.projectMetadata(read({ prompt, workflow }, { prompt: "original-text" }, null, ["workflow"])));
+    assert.deepEqual(limited.raw, { prompt });
+    assert.deepEqual(limited.record_formats, { prompt: "original-text", workflow: "limited" });
+    assert.equal(limited.workflow_graph.status, "limited");
+    assert.equal(plain(m.projectMetadata(read({ prompt }, { prompt: "original-text", workflow: "invalid" }))).workflow_graph.status, "malformed");
+    assert.equal(plain(m.projectMetadata(read({ workflow }, { workflow: "reserialized-json" }))).workflow_graph.status, "limited");
+    assert.equal(plain(m.projectMetadata(read({ prompt: "{" }, { prompt: "original-text" }))).prompt_graph.status, "malformed");
+    // Unknown formats are dropped from record_formats but the raw text still projects.
+    const odd = plain(m.projectMetadata(read({ prompt }, { prompt: "weird" })));
+    assert.deepEqual(odd.record_formats, {});
+    assert.equal(odd.prompt_graph.status, "parsed");
+    // A non-text record vanishes from raw and its named graph is malformed.
+    assert.equal(plain(m.projectMetadata(read({ prompt: 5 }, { prompt: "original-text" }))).prompt_graph.status, "malformed");
+    // Carriers: strings re-parse by kind, objects were projected at depth 2; a bad carrier fails every present graph.
+    const envelope = JSON.stringify({ prompt, workflow: { nodes: [], v: 1 } });
+    const carried = plain(m.projectMetadata(read({ prompt, workflow: '{"nodes": [], "v": 1}' }, { prompt: "original-text", workflow: "reserialized-json" }, envelope)));
+    assert.deepEqual(carried.prompt_graph, { status: "parsed", value: { 1: { class_type: "A", inputs: { x: "1.0" } } } });
+    assert.deepEqual(carried.workflow_graph, { status: "parsed", value: { nodes: [], v: 1 } });
+    const broken = plain(m.projectMetadata(read({ prompt, workflow }, { prompt: "original-text", workflow: "original-text" }, "[1]")));
+    assert.deepEqual([broken.prompt_graph.status, broken.workflow_graph.status], ["malformed", "malformed"]);
+    const deep = plain(m.projectMetadata(read({ prompt }, { prompt: "original-text" }, `{"title":${nested(70)},"prompt":${JSON.stringify(prompt)}}`)));
+    assert.equal(deep.prompt_graph.status, "limited");
+    const partial = plain(m.projectMetadata(read({ prompt, workflow }, { prompt: "original-text", workflow: "original-text" }, '{"prompt":"{","workflow":5}')));
+    assert.deepEqual([partial.prompt_graph.status, partial.workflow_graph.status], ["malformed", "malformed"]);
+    const missing = plain(m.projectMetadata(read({ prompt }, { prompt: "original-text" }, '{"workflow":{}}')));
+    assert.deepEqual([missing.prompt_graph.status, missing.workflow_graph.status], ["malformed", "absent"]);
+    // The carrier's own prompt object is projected: depth 2, node keys filtered.
+    const objectPrompt = plain(m.projectMetadata(read({ prompt: "x" }, { prompt: "reserialized-json" }, '{"prompt":{"1":{"class_type":"A","inputs":{},"_meta":{}}}}')));
+    assert.deepEqual(objectPrompt.prompt_graph, { status: "parsed", value: { 1: { class_type: "A", inputs: {} } } });
+});
+
+test("fitInspectionPayload omits the workflow graph, then the prompt graph, then all raw atomically", async () => {
+    const payload = {
+        success: true, inspection_version: 1, metadata: {
+            status: "available", omitted_reason: null, raw: { prompt: "é".repeat(100), workflow: "{}" }, record_formats: { prompt: "original-text", workflow: "original-text" },
+            parameters: null, prompt_graph: { status: "parsed", value: { 1: { class_type: "A", inputs: { t: "\ud800".repeat(50) } } } },
+            workflow_graph: { status: "parsed", value: { nodes: Array.from({ length: 50 }, (_, i) => ({ id: i })) } },
+        },
+    };
+    const text = JSON.stringify(payload);
+    // UTF-8 bytes, not UTF-16 code units: 'é' is 2 bytes, an escaped lone surrogate 6.
+    assert.equal(m.inspectionBytes(payload), Buffer.byteLength(text, "utf8"));
+    assert.ok(m.inspectionBytes(payload) > text.length);
+    // Measured in bounded chunks: pairs and 2-byte characters across every chunk edge count once.
+    for (const shift of [0, 1, 2, 3]) {
+        const wide = { s: "x".repeat(shift) + "é😀".repeat(200000) + "\ud800", t: "a😀".repeat(70000) };
+        assert.equal(m.inspectionBytes(wide), Buffer.byteLength(JSON.stringify(wide), "utf8"), `shift ${shift}`);
+    }
+    const whole = await m.fitInspectionPayload(payload, bytes(payload));
+    assert.equal(whole, payload, "a payload within the limit is returned as is");
+    const first = await m.fitInspectionPayload(payload, bytes(payload) - 1);
+    assert.equal(first.metadata.workflow_graph.status, "limited");
+    assert.equal(first.metadata.prompt_graph.status, "parsed");
+    assert.deepEqual(first.metadata.raw, payload.metadata.raw);
+    assert.ok(bytes(first) <= bytes(payload) - 1);
+    const second = await m.fitInspectionPayload(payload, bytes(first) - 1);
+    assert.deepEqual([second.metadata.workflow_graph.status, second.metadata.prompt_graph.status], ["limited", "limited"]);
+    assert.equal(second.metadata.status, "available");
+    assert.deepEqual(second.metadata.raw, payload.metadata.raw);
+    const third = await m.fitInspectionPayload(payload, bytes(second) - 1);
+    assert.deepEqual(third.metadata, { ...second.metadata, raw: {}, parameters: null, status: "limited", omitted_reason: "response-size-limit" });
+    assert.ok(bytes(third) <= bytes(second) - 1);
+    await assert.rejects(m.fitInspectionPayload(payload, bytes(third) - 1), /exceeds/);
+    // The caller's payload is never mutated and absent graphs are left alone.
+    assert.equal(JSON.stringify(payload), text);
+    const absent = { ...payload, metadata: { ...payload.metadata, workflow_graph: { status: "absent", value: null } } };
+    const fitted = await m.fitInspectionPayload(absent, bytes(absent) - 1);
+    assert.equal(fitted.metadata.workflow_graph.status, "absent");
+    assert.equal(fitted.metadata.prompt_graph.status, "limited");
+    assert.equal((await m.fitInspectionPayload(payload)).metadata.status, "available", "the default limit is 8 MiB");
+    // Every omission step after the first measure hands the event loop a turn, and an
+    // abort between steps rejects with AbortError; the answer itself is unchanged.
+    let turns = 0;
+    const counted = await m.fitInspectionPayload(payload, bytes(second) - 1, { yieldControl: async () => { turns += 1; } });
+    assert.equal(JSON.stringify(counted), JSON.stringify(third));
+    assert.equal(turns, 3);
+    assert.equal(await m.fitInspectionPayload(payload, bytes(payload), { yieldControl: async () => { turns += 1; } }), payload);
+    assert.equal(turns, 3, "a payload within the limit is measured once, without a turn");
+    const controller = new AbortController();
+    await assert.rejects(m.fitInspectionPayload(payload, bytes(second) - 1, { signal: controller.signal, yieldControl: async () => controller.abort() }), { name: "AbortError" });
+});
+
+test("the committed projection corpus has every outcome and parses with the oracle entry shape", () => {
+    assert.ok(CASES.length >= 40);
+    const names = new Set(), statuses = new Set(), graphs = new Set();
+    for (const entry of CASES) {
+        assert.equal(typeof entry.name, "string");
+        assert.ok(!names.has(entry.name), `duplicate case ${entry.name}`);
+        names.add(entry.name);
+        const read = entry.records === null ? null : { records: entry.records, record_formats: entry.record_formats ?? Object.fromEntries(Object.keys(entry.records).map((key) => [key, "original-text"])),
+            carrierText: entry.carrier_text ?? null, limited: entry.limited ?? [] };
+        const metadata = m.projectMetadata(read);
+        statuses.add(metadata.status);
+        graphs.add(metadata.prompt_graph.status);
+        graphs.add(metadata.workflow_graph.status);
+        if (entry.expect) {
+            for (const [field, status] of Object.entries(entry.expect)) assert.equal(metadata[field].status, status, `${entry.name} ${field}`);
+        }
+    }
+    assert.deepEqual([...statuses].sort(), ["absent", "available", "unreadable"]);
+    assert.deepEqual([...graphs].sort(), ["absent", "limited", "malformed", "parsed"]);
+});

tests/js/workspace-local-video.test.js

@@ -0,0 +1,622 @@
+// Run with: node --test tests/js/
+// The browser video tag readers mirror what the server reads through ffprobe
+// (utils/prompt_inspector.py _video_metadata) and the carrier rule of
+// utils/video_metadata.py select_video_carrier. Each rule names its server twin;
+// tests/test_local_inspect_parity.py proves the two agree on the committed fixtures
+// and on recorded tag maps.
+import test from "node:test";
+import assert from "node:assert/strict";
+import fs from "node:fs";
+import path from "node:path";
+import { fileURLToPath } from "node:url";
+import * as v from "./helpers/video-bytes.mjs";
+
+const here = path.dirname(fileURLToPath(import.meta.url));
+const FIXTURES = path.join(here, "..", "fixtures", "local-inspect");
+const RECORDED = JSON.parse(fs.readFileSync(path.join(FIXTURES, "video-tags-v1.json"), "utf8")).fixtures;
+const m = await import("../../web/js/workspace/local-video.mjs").catch((error) => {
+    if (error.code !== "ERR_MODULE_NOT_FOUND") throw error;
+    return {};
+});
+const { readVideoTags, selectVideoCarrier, pythonJsonDumps, pythonJsonLoads } = m;
+
+const MiB = 1024 * 1024;
+const GiB = 1024 * MiB;
+const ROLE_BUDGET = 2 * MiB;
+const WORKFLOW_BUDGET = 4 * MiB;
+const TAG_READ_BYTES = 12 * MiB + 64 * 1024;
+const PER_KEY = new Set(["prompt", "workflow", "parameters", "comfyassets_prompt_pairs"]);
+const GRAPH = '{"3":{"class_type":"KSampler","inputs":{"seed":7}}}';
+const WORKFLOW = '{"nodes":[],"version":0.4}';
+
+// The tags ffprobe reports to the server: the six probed names, ASCII case folded.
+function knownTags(map) {
+    return Object.fromEntries(Object.entries(map).filter(([name]) => PER_KEY.has(name.toLowerCase()) || ["comment", "description"].includes(name.toLowerCase())));
+}
+
+function fixture(name) {
+    return new Uint8Array(fs.readFileSync(path.join(FIXTURES, name)));
+}
+
+async function read(bytes, options = {}) {
+    return readVideoTags(v.blobOf(bytes), options);
+}
+
+function fill(byte, count) {
+    return Buffer.alloc(count, byte);
+}
+
+function complete(tags, extra = {}) {
+    return { status: "complete", tags, limited: [], ...extra };
+}
+
+test("contract: the readers, the carrier rule and the json port are exported", () => {
+    assert.equal(typeof readVideoTags, "function");
+    assert.equal(typeof selectVideoCarrier, "function");
+    assert.equal(typeof pythonJsonDumps, "function");
+    assert.equal(typeof pythonJsonLoads, "function");
+});
+
+test("iso-bmff reader finds per-key tags with moov after mdat", async () => {
+    for (const [name, recorded] of Object.entries(RECORDED)) {
+        const blob = v.recordingBlob(fixture(name));
+        const result = await readVideoTags(blob, { prefix: fixture(name).subarray(0, 32) });
+        assert.equal(result.status, "complete", name);
+        assert.deepEqual(result.tags, knownTags(recorded), name);
+        assert.equal(result.container, name.endsWith(".mkv") || name.endsWith(".webm") ? "matroska" : "iso-bmff", name);
+        assert.ok(["isom", "qt", "M4V", "matroska", "webm"].includes(result.format), `${name}: ${result.format}`);
+        assert.deepEqual(result.limited, [], name);
+        assert.ok(blob.ranges.length > 0, name);
+    }
+    // core-moov-end.mp4 behind a 3 GiB mdat: only the box headers and the tag bytes are sliced.
+    const bytes = fixture("core-moov-end.mp4");
+    const boxes = v.topLevelBoxes(bytes);
+    const mdat = boxes.find((box) => box.type === "mdat");
+    const moov = boxes.find((box) => box.type === "moov");
+    const mdatSize = 3 * GiB + 16;
+    const header = v.concat(v.u32(1), v.latin1("mdat"), new Uint8Array(8));
+    new DataView(header.buffer).setBigUint64(8, BigInt(mdatSize));
+    const moovOffset = mdat.start + mdatSize;
+    const blob = v.sparseBlob(
+        [
+            { offset: 0, bytes: bytes.subarray(0, mdat.start) },
+            { offset: mdat.start, bytes: header },
+            { offset: moovOffset, bytes: bytes.subarray(moov.start, moov.end) },
+        ],
+        moovOffset + (moov.end - moov.start),
+    );
+    const result = await readVideoTags(blob, { prefix: bytes.subarray(0, 32) });
+    assert.deepEqual(result.tags, knownTags(RECORDED["core-moov-end.mp4"]));
+    const inside = blob.ranges.filter(([start]) => start >= mdat.start && start < moovOffset);
+    assert.ok(inside.every(([start, end]) => start === mdat.start && end - start <= 16), JSON.stringify(inside));
+    const total = blob.ranges.reduce((sum, [start, end]) => sum + (end - start), 0);
+    assert.ok(total <= m.VIDEO_READ_CEILING && total < 64 * 1024, `${total}`);
+});
+
+test("iso-bmff: legacy ilst atoms, moov/meta, sticky mdta handler and data types", async () => {
+    // ©cmt and desc map to comment and description, as ffprobe reports them.
+    const legacy = await read(v.mp4(v.legacyUdta({ comment: '{"prompt":"{}"}', description: "note" })));
+    assert.deepEqual(legacy, complete({ comment: '{"prompt":"{}"}', description: "note" }, { container: "iso-bmff", format: "isom" }));
+    // moov/meta (ISO style) is read like moov/udta/meta.
+    const isoMeta = await read(v.mp4(v.box("meta", v.u32(0), v.hdlr("mdta"), v.keys(["prompt"]), v.ilst(v.item(1, v.data(GRAPH))))));
+    assert.deepEqual(isoMeta.tags, { prompt: GRAPH });
+    // meta without the version/flags word (QuickTime) still finds hdlr.
+    const qtMeta = await read(v.mp4(v.udta(v.box("meta", v.hdlr("mdta"), v.keys(["prompt"]), v.ilst(v.item(1, v.data(GRAPH)))))));
+    assert.deepEqual(qtMeta.tags, { prompt: GRAPH });
+    // Index items need an mdta handler and a keys box; with a legacy handler they are skipped.
+    const noKeys = await read(v.mp4(v.udta(v.meta(v.hdlr("mdir"), v.ilst(v.item(1, v.data(GRAPH)))))));
+    assert.deepEqual(noKeys.tags, {});
+    // The mdta handler is sticky across the file, as ffmpeg's found_hdlr_mdta.
+    const sticky = await read(v.mp4(v.udta(v.meta(v.hdlr("mdta"), v.keys(["prompt"])), v.meta(v.hdlr("mdir"), v.ilst(v.item(1, v.data(GRAPH)))))));
+    assert.deepEqual(sticky.tags, { prompt: GRAPH });
+    // As ffprobe reports them: consecutive data children join with ";", integer data
+    // types print as decimals, other types are not tags, text types 0/1/4 are read, NUL
+    // truncates, and the run of data children stops at the first other child.
+    const mixed = await read(v.mp4(v.udta(v.meta(
+        v.hdlr("mdta"),
+        v.keys(["prompt", "workflow", "parameters", "comfyassets_prompt_pairs", "comment", "description"]),
+        v.ilst(
+            v.item(1, v.data("first"), v.data("second"), v.box("name", "x"), v.data("not read")),
+            v.item(2, v.data([0, 0, 0, 7], { type: 21 }), v.data([0xff, 0xfe], { type: 21 }), v.data([0xff, 0xfe], { type: 22 })),
+            v.item(3, v.data("cut\0off")),
+            v.item(4, v.data("sort", { type: 4 }), v.data("mac", { type: 0 })),
+            v.item(5, v.data([0, 0, 0, 0, 0, 0, 0], { type: 21 })),
+            v.item(6, v.data("float", { type: 23 })),
+        ),
+    ))));
+    assert.deepEqual(mixed.tags, { prompt: "first;second", workflow: "7;-2;65534", parameters: "cut", comfyassets_prompt_pairs: "sort;mac" });
+    // Keys outside the mdta namespace address nothing; unknown names are skipped unread.
+    const blob = v.recordingBlob(v.mp4(v.udta(v.meta(
+        v.hdlr("mdta"),
+        v.keys([["prompt", "udta"], "encoder", "workflow"]),
+        v.ilst(v.item(1, v.data(GRAPH)), v.item(2, v.data(fill(0x41, 5000))), v.item(3, v.data(WORKFLOW))),
+    ))));
+    const result = await readVideoTags(blob);
+    assert.deepEqual(result.tags, { workflow: WORKFLOW });
+    assert.ok(blob.ranges.every(([start, end]) => end - start < 5000), "the encoder value was never sliced");
+    // Legacy fourccs are mapped even in mdta mode, and a repeated name joins with ";".
+    const repeated = await read(v.mp4(v.udta(v.meta(
+        v.hdlr("mdta"),
+        v.keys(["prompt", "prompt", "comment"]),
+        v.ilst(v.item(1, v.data("one")), v.item(2, v.data("two")), v.item(3, v.data("mdta")), v.legacyItem("©cmt", v.data("c"))),
+    ))));
+    assert.deepEqual(repeated.tags, { prompt: "one;two", comment: "mdta;c" });
+    // Apple's namespaced keys are reported verbatim, so they are never an envelope.
+    const apple = await read(v.perKeyMp4({ "com.apple.quicktime.description": '{"prompt":"{}"}', prompt: GRAPH }));
+    assert.deepEqual(apple.tags, { prompt: GRAPH });
+    // An item whose first child is not a data atom, or a data atom shorter than its
+    // 16-byte header, is skipped as ffmpeg skips it.
+    const junk = await read(v.mp4(v.udta(v.meta(v.hdlr("mdta"), v.keys(["prompt"]), v.ilst(v.item(1, v.box("name", "x"), v.data(GRAPH)))))));
+    assert.deepEqual(junk.tags, {});
+    const short = await read(v.mp4(v.udta(v.meta(v.hdlr("mdta"), v.keys(["prompt"]), v.ilst(v.item(1, v.data(GRAPH, { size: 12 })))))));
+    assert.deepEqual(short.tags, {});
+    // A box declared with size 0 runs to the end of its parent; faststart puts moov first.
+    const fast = await read(v.perKeyMp4({ prompt: GRAPH }, { faststart: true, brand: "mp42" }));
+    assert.deepEqual(fast, complete({ prompt: GRAPH }, { container: "iso-bmff", format: "mp42" }));
+    const toEnd = await read(v.concat(v.ftyp(), v.mdat(), v.moov(v.mdtaUdta({ prompt: GRAPH }), v.opts({ size: 0 }))));
+    assert.deepEqual(toEnd.tags, { prompt: GRAPH });
+});
+
+test("matroska reads global Tags and skips clusters", async () => {
+    const { ID } = v;
+    // ffmpeg layout: SeekHead, Info, Tracks, Tags, Cluster, Cues; names upper-cased.
+    const plain = await read(v.perKeyMkv({ PROMPT: GRAPH, WORKFLOW, ENCODER: "Lavf" }, { docType: "webm" }));
+    assert.deepEqual(plain, complete({ PROMPT: GRAPH, WORKFLOW }, { container: "matroska", format: "webm" }));
+    // Tags after a 3 GiB Cluster, located through the SeekHead: no slice inside the cluster.
+    const tags = v.globalTags({ PROMPT: GRAPH });
+    const clusterSize = 3 * GiB;
+    // The Tags position (relative to the Segment payload) is a 4-byte uint either way,
+    // so the SeekHead's length can be measured with a placeholder first.
+    const seekLength = v.seekHead([[ID.Tags, clusterSize]]).length;
+    const seekHead = v.seekHead([[ID.Tags, seekLength + 12 + clusterSize]]);
+    assert.equal(seekHead.length, seekLength);
+    const head = v.concat(v.ebmlHeader(), v.idBytes(ID.Segment), v.vint(seekHead.length + 12 + clusterSize + tags.length, { width: 8 }), seekHead, v.clusterHeader(clusterSize));
+    const clusterStart = v.ebmlHeader().length + 12 + seekHead.length;
+    const tagsOffset = head.length + clusterSize;
+    const blob = v.sparseBlob([{ offset: 0, bytes: head }, { offset: tagsOffset, bytes: tags }], tagsOffset + tags.length);
+    const seeked = await readVideoTags(blob, { prefix: head.subarray(0, 32) });
+    assert.deepEqual(seeked.tags, { PROMPT: GRAPH });
+    const inside = blob.ranges.filter(([start]) => start >= clusterStart && start < tagsOffset);
+    assert.ok(inside.every(([start, end]) => start === clusterStart && end - start <= 16), JSON.stringify(inside));
+    assert.ok(blob.ranges.reduce((sum, [start, end]) => sum + (end - start), 0) <= m.VIDEO_READ_CEILING);
+    // ffprobe reads level-1 elements up to the first Cluster: Tags behind it are found
+    // only through a SeekHead, so without one they are not read (nor sliced).
+    const behind = v.recordingBlob(v.mkv([v.cluster(64), v.globalTags({ PROMPT: GRAPH }), el(ID.Cues, new Uint8Array(2))]));
+    const sequential = await readVideoTags(behind);
+    assert.deepEqual(sequential.tags, {});
+    assert.ok(behind.ranges.every(([, end]) => end <= v.ebmlHeader().length + 12 + 12 + 32), JSON.stringify(behind.ranges));
+    // Track-targeted Tags are ignored; chapter and attachment targets too.
+    const targeted = await read(v.mkv([
+        v.tagsEl(
+            v.tag([v.simpleTag("PROMPT", "track")], { trackUid: 1 }),
+            v.tag([v.simpleTag("WORKFLOW", WORKFLOW)], { targetType: 50 }),
+            v.tag([v.simpleTag("PROMPT", "zero uid is global")], { trackUid: 0 }),
+        ),
+        v.cluster(),
+    ]));
+    assert.deepEqual(targeted.tags, { WORKFLOW, PROMPT: "zero uid is global" });
+    // An unknown-size Cluster (live recordings) is the same stop; a SeekHead still works.
+    const hidden = await read(v.mkv([v.cluster(8, { unknown: true }), v.globalTags({ PROMPT: GRAPH })]));
+    assert.deepEqual(hidden, complete({}, { container: "matroska", format: "matroska" }));
+    const before = await read(v.mkv([v.globalTags({ PROMPT: GRAPH }), v.cluster(8, { unknown: true })]));
+    assert.deepEqual(before.tags, { PROMPT: GRAPH });
+    const info = el(ID.Info, new Uint8Array(4));
+    const clusterUnknown = v.cluster(8, { unknown: true });
+    const seekToTags = v.seekHead([[ID.Tags, 0x7f]]); // a one-byte position, like the real one
+    const viaSeek = await read(v.mkv([v.seekHead([[ID.Tags, seekToTags.length + info.length + clusterUnknown.length]]), info, clusterUnknown, v.globalTags({ PROMPT: GRAPH })]));
+    assert.deepEqual(viaSeek.tags, { PROMPT: GRAPH });
+    // An unknown-size Segment is fine (live recordings); other unknown sizes are invalid.
+    const liveSegment = await read(v.mkv([v.globalTags({ PROMPT: GRAPH }), v.cluster()], { unknown: true }));
+    assert.deepEqual(liveSegment.tags, { PROMPT: GRAPH });
+    const unknownTags = await read(v.mkv([el(ID.Tags, v.tag([v.simpleTag("PROMPT", GRAPH)]), v.opts({ unknown: true }))]));
+    assert.deepEqual(unknownTags, { status: "unreadable", reason: "invalid-container" });
+    // ffmpeg's tag naming (checked against ffprobe 9.0.2): a language other than und
+    // files the tag under NAME-lang only, unless TagDefault is set; nested SimpleTags
+    // become PARENT/CHILD (never ours); a string-less tag deletes the key; a repeated
+    // name keeps the last value; NUL ends the string; Void and CRC-32 are skipped.
+    const naming = await read(v.mkv([
+        v.voidEl(8),
+        v.tagsEl(
+            el(ID.CRC32, new Uint8Array(4)),
+            v.tag([
+                v.simpleTag("PROMPT", "eng only", { lang: "eng" }),
+                v.simpleTag("WORKFLOW", "eng only", { lang: "eng", def: 0 }),
+                v.simpleTag("workflow", "default and eng", { lang: "eng", def: 1 }),
+                v.simpleTag("PARAMETERS", "und", { lang: "und", nested: [v.simpleTag("prompt", "nested")] }),
+                v.simpleTag("comfyassets_prompt_pairs", "cut\0here"),
+                v.simpleTag("description", undefined, { binary: [1, 2, 3] }),
+                v.simpleTag("comment", "note"),
+                v.simpleTag("comment", undefined),
+                v.simpleTag("prompt", "first"),
+                v.simpleTag("prompt", "last", { binary: [1] }),
+            ]),
+        ),
+        v.cluster(),
+    ]));
+    assert.deepEqual(naming.tags, { workflow: "default and eng", PARAMETERS: "und", comfyassets_prompt_pairs: "cut", prompt: "last" });
+    // Long names are never ours and are skipped without reading their strings.
+    const longName = v.recordingBlob(v.mkv([v.tagsEl(v.tag([v.simpleTag("N".repeat(300), fill(0x41, 4000)), v.simpleTag("PROMPT", GRAPH)])), v.cluster()]));
+    const skipped = await readVideoTags(longName);
+    assert.deepEqual(skipped.tags, { PROMPT: GRAPH });
+    assert.ok(longName.ranges.every(([start, end]) => end - start < 300));
+    function el(id, ...parts) {
+        return v.el(id, ...parts);
+    }
+});
+
+test("containers fail closed", async () => {
+    const { ID } = v;
+    const invalid = { status: "unreadable", reason: "invalid-container" };
+    const rows = [
+        ["box size below its header", v.concat(v.ftyp(), v.box("free", v.opts({ size: 4 })), v.mdat(), v.moov(v.mdtaUdta({ prompt: GRAPH })))],
+        ["child box past its parent", v.concat(v.ftyp(), v.mdat(), v.moov(v.udta(v.meta(v.hdlr("mdta"), v.keys(["prompt"]), v.ilst(v.item(1, v.data(GRAPH), v.opts({ size: 5000 })))))))],
+        ["top-level box past the end of the file", v.concat(v.ftyp(), v.mdat(), v.box("moov", v.opts({ size: 4000 })))],
+        ["64-bit size above MAX_SAFE_INTEGER", v.concat(v.ftyp(), v.box("mdat", v.opts({ largesize: 2 ** 60 })), v.moov(v.mdtaUdta({ prompt: GRAPH })))],
+        ["64-bit size below its header", v.concat(v.ftyp(), v.box("mdat", v.opts({ largesize: 12 })), v.moov(v.mdtaUdta({ prompt: GRAPH })))],
+        ["keys count larger than the entries", v.mp4(v.udta(v.meta(v.hdlr("mdta"), v.keys(["prompt"], { count: 3 }), v.ilst(v.item(1, v.data(GRAPH))))))],
+        ["keys entry overrunning the box", v.mp4(v.udta(v.meta(v.hdlr("mdta"), v.box("keys", v.u32(0), v.u32(1), v.u32(90), v.latin1("mdta"), v.utf8("prompt")), v.ilst(v.item(1, v.data(GRAPH))))))],
+        ["no moov box", v.concat(v.ftyp(), v.mdat(), v.free(32))],
+        ["not a container at all", v.utf8("#EXTM3U\n#EXT-X-VERSION:3\n".padEnd(64, "x"))],
+        ["bad VINT: zero ID byte", v.concat(v.ebmlHeader(), v.idBytes(ID.Segment), v.vint(20, { width: 8 }), [0x00, 0x81, 0x00], new Uint8Array(17))],
+        ["bad VINT: zero size byte", v.concat(v.ebmlHeader(), v.idBytes(ID.Segment), v.vint(20, { width: 8 }), v.idBytes(ID.Info), [0x00], new Uint8Array(17))],
+        ["element past its parent", v.mkv([v.el(ID.Tags, v.tag([v.simpleTag("PROMPT", GRAPH)]), v.opts({ size: 5000 }))])],
+        ["element past the end of the file", v.concat(v.ebmlHeader(), v.idBytes(ID.Segment), v.vint(5000, { width: 8 }), v.globalTags({ PROMPT: GRAPH }))],
+        ["unknown DocType", v.perKeyMkv({ PROMPT: GRAPH }, { docType: "avi" })],
+        ["EBML header without a DocType", v.concat(v.el(ID.EBML, v.el(ID.EBMLVersion, [1])), v.segment([v.globalTags({ PROMPT: GRAPH })]))],
+        ["Matroska bytes after a stray size", v.concat(v.ebmlHeader(), v.idBytes(ID.Segment), v.vint(3, { width: 8 }), [0xec, 0x82, 0x00])],
+    ];
+    for (const [name, bytes] of rows) {
+        assert.deepEqual(await read(bytes), invalid, name);
+    }
+    // Trailing bytes shorter than a box header are ignored, as ffmpeg does.
+    assert.deepEqual((await read(v.concat(v.perKeyMp4({ prompt: GRAPH }), [1, 2, 3]))).tags, { prompt: GRAPH });
+    // 10,001 boxes stop the read.
+    const many = v.concat(v.ftyp(), ...Array.from({ length: 10001 }, () => v.free()), v.mdat(), v.moov(v.mdtaUdta({ prompt: GRAPH })));
+    assert.deepEqual(await read(many), { status: "limited", reason: "chunk-limit" });
+    const manyElements = v.mkv([...Array.from({ length: 10001 }, () => v.voidEl(1)), v.globalTags({ PROMPT: GRAPH })]);
+    assert.deepEqual(await read(manyElements), { status: "limited", reason: "chunk-limit" });
+    // Exactly 10,000 elements still read (ftyp, mdat, moov, udta, meta, hdlr, keys, its entry, ilst, the item and its data atom).
+    const exact = v.concat(v.ftyp(), ...Array.from({ length: 10000 - 11 }, () => v.free()), v.mdat(), v.moov(v.mdtaUdta({ prompt: GRAPH })));
+    assert.deepEqual((await read(exact)).tags, { prompt: GRAPH });
+    // A vanished file is a read failure, not a container error.
+    assert.deepEqual(await readVideoTags(v.failingBlob(4096), { prefix: v.ftyp() }), { status: "unreadable", reason: "read-failure" });
+    assert.deepEqual(await readVideoTags({}), { status: "unreadable", reason: "read-failure" });
+});
+
+// ffprobe's string_validation=replace (av_utf8_decode): one U+FFFD per failing
+// decode step. A bad continuation or a truncated sequence fails one byte at a time;
+// an overlong form, a surrogate, a code above U+10FFFF, U+FFFE/U+FFFF and 5/6-byte
+// forms fail as one whole sequence. Each row was read through ffprobe 9.0.2.
+const FFPROBE_REPLACEMENTS = [
+    ["one Latin-1 byte", [0x63, 0x61, 0x66, 0xe9], "caf�"],
+    ["truncated 4-byte sequence", [0x61, 0xf0, 0x9f, 0x22, 0x62], 'a��"b'],
+    ["overlong 2-byte form", [0x61, 0xc0, 0xaf, 0x62], "a�b"],
+    ["C1 overlong lead", [0x61, 0xc1, 0xbf, 0x62], "a�b"],
+    ["overlong 3-byte form", [0x61, 0xe0, 0x80, 0x80, 0x62], "a�b"],
+    ["overlong 4-byte form", [0x61, 0xf0, 0x80, 0x80, 0x80, 0x62], "a�b"],
+    ["UTF-16 surrogate", [0x61, 0xed, 0xa0, 0x80, 0x62], "a�b"],
+    ["U+FFFE", [0x61, 0xef, 0xbf, 0xbe, 0x62], "a�b"],
+    ["U+FFFF", [0x61, 0xef, 0xbf, 0xbf, 0x62], "a�b"],
+    ["U+FDD0 is kept", [0x61, 0xef, 0xb7, 0x90, 0x62], "a﷐b"],
+    ["5-byte form", [0x61, 0xf8, 0x88, 0x80, 0x80, 0x80, 0x62], "a�b"],
+    ["6-byte form", [0x61, 0xfc, 0x84, 0x80, 0x80, 0x80, 0x80, 0x62], "a�b"],
+    ["truncated 5-byte form", [0x61, 0xf8, 0x88, 0x62], "a��b"],
+    ["above U+10FFFF", [0x61, 0xf4, 0x90, 0x80, 0x80, 0x62], "a�b"],
+    ["0xFE", [0x61, 0xfe, 0x62], "a�b"],
+    ["0xFF", [0x61, 0xff, 0x62], "a�b"],
+    ["0xFE 0xFF", [0xfe, 0xff], "��"],
+    ["lone continuation byte", [0x61, 0x80, 0x62], "a�b"],
+    ["truncated at the end", [0x61, 0xe2, 0x82], "a��"],
+    ["lead byte before a whole sequence", [0x61, 0xc3, 0xc3, 0xa9, 0x62], "a�éb"],
+    ["valid emoji", [0x61, 0xf0, 0x9f, 0x98, 0x80, 0x62], "a\u{1f600}b"],
+    ["control characters", [0x61, 0x01, 0x1f, 0x62], "a\u0001\u001fb"],
+];
+
+test("invalid UTF-8 reads as ffprobe replaces it; names that do not decode are not carriers", async () => {
+    for (const [name, bytes, expected] of FFPROBE_REPLACEMENTS) {
+        assert.deepEqual((await read(v.perKeyMp4({ prompt: bytes }))).tags, { prompt: expected }, name);
+        assert.deepEqual((await read(v.perKeyMkv({ PROMPT: bytes }))).tags, { PROMPT: expected }, `${name} (Matroska)`);
+    }
+    // A long value with one bad byte keeps every other byte.
+    const long = Buffer.concat([Buffer.from('{"a":"'), Buffer.alloc(100000, 0x78), Buffer.from([0xe9]), Buffer.from('"}')]);
+    assert.equal((await read(v.perKeyMp4({ workflow: long }))).tags.workflow, `{"a":"${"x".repeat(100000)}�"}`);
+    assert.deepEqual((await read(v.perKeyMp4({ prompt: [0x7b, 0xff, 0xfe, 0x7d] }))).tags, { prompt: "{��}" });
+    // A keys name with a bad byte is reported as "prom�pt", which is no carrier.
+    const badKey = v.mp4(v.udta(v.meta(v.hdlr("mdta"), v.keys([Uint8Array.from([0x70, 0x72, 0x6f, 0x6d, 0xff, 0x70, 0x74]), "prompt"]), v.ilst(v.item(1, v.data("junk")), v.item(2, v.data(GRAPH))))));
+    assert.deepEqual(await read(badKey), complete({ prompt: GRAPH }, { container: "iso-bmff", format: "isom" }));
+    const onlyBadKey = v.mp4(v.udta(v.meta(v.hdlr("mdta"), v.keys([Uint8Array.from([0xc3, 0x28])]), v.ilst(v.item(1, v.data(GRAPH))))));
+    assert.deepEqual((await read(onlyBadKey)).tags, {});
+    assert.deepEqual((await read(v.mkv([v.tagsEl(v.tag([v.simpleTag([0xff], GRAPH), v.simpleTag("WORKFLOW", WORKFLOW)]))]))).tags, { WORKFLOW: WORKFLOW });
+    assert.deepEqual((await read(v.mkv([v.tagsEl(v.tag([v.simpleTag("PROMPT", [0xff, 0xfe])]))]))).tags, { PROMPT: "��" });
+});
+
+test("ilst items with index 0 or past the keys are skipped; the first keys box wins", async () => {
+    const wf = '{"nodes":[]}';
+    const stray = v.mp4(v.udta(v.meta(v.hdlr("mdta"), v.keys(["prompt", "workflow"]), v.ilst(v.item(1, v.data(GRAPH)), v.item(5, v.data("junk")), v.item(2, v.data(wf))))));
+    assert.deepEqual((await read(stray)).tags, { prompt: GRAPH, workflow: wf });
+    const zero = v.mp4(v.udta(v.meta(v.hdlr("mdta"), v.keys(["prompt", "workflow"]), v.ilst(v.item(0, v.data("junk")), v.item(1, v.data(GRAPH)), v.item(2, v.data(wf))))));
+    assert.deepEqual((await read(zero)).tags, { prompt: GRAPH, workflow: wf });
+    assert.deepEqual((await read(v.mp4(v.udta(v.meta(v.hdlr("mdta"), v.keys(["prompt"]), v.ilst(v.item(2, v.data(GRAPH)))))))).tags, {});
+    // ffmpeg reads one keys box: a second one (even malformed) addresses nothing.
+    const twoKeys = v.mp4(v.udta(v.meta(v.hdlr("mdta"), v.keys(["prompt"]), v.keys(["workflow"]), v.ilst(v.item(1, v.data(GRAPH))))));
+    assert.deepEqual((await read(twoKeys)).tags, { prompt: GRAPH });
+    const brokenSecond = v.mp4(v.udta(v.meta(v.hdlr("mdta"), v.keys(["prompt"]), v.keys(["workflow"], { count: 9 }), v.ilst(v.item(1, v.data(GRAPH))))));
+    assert.deepEqual((await read(brokenSecond)).tags, { prompt: GRAPH });
+    const emptyFirst = v.mp4(v.udta(v.meta(v.hdlr("mdta"), v.keys([]), v.keys(["prompt"]), v.ilst(v.item(1, v.data(GRAPH))))));
+    assert.deepEqual((await read(emptyFirst)).tags, {});
+    // A keys box too short for its count field is not read, so the next one is.
+    const shortFirst = v.mp4(v.udta(v.meta(v.hdlr("mdta"), v.fullBox("keys"), v.keys(["prompt"]), v.ilst(v.item(1, v.data(GRAPH))))));
+    assert.deepEqual((await read(shortFirst)).tags, { prompt: GRAPH });
+});
+
+test("one tag per ASCII-folded name and Mac Roman text, as ffmpeg's tag dictionary holds them", async () => {
+    const mdta = (names, items, legacy = []) => v.mp4(v.udta(v.meta(v.hdlr("mdta"), v.keys(names), v.ilst(...items.map(([index, value, type = 1]) => v.item(index, v.data(value, { type }))), ...legacy))));
+    // ISO-BMFF joins case variants with ";" under the spelling set last (checked through ffprobe).
+    assert.deepEqual((await read(mdta(["workflow", "Workflow"], [[1, "A"], [2, "B"]]))).tags, { Workflow: "A;B" });
+    // A legacy comment atom and a later mdta COMMENT key become one COMMENT tag: no exact-case envelope.
+    const merged = await read(mdta(["COMMENT"], [[1, "y"]], [v.legacyItem("©cmt", v.data("x"))]));
+    assert.deepEqual(merged.tags, { comment: "y;x" });
+    const mergedUpper = await read(v.mp4(v.udta(v.meta(v.hdlr("mdta"), v.keys(["COMMENT"]), v.ilst(v.legacyItem("©cmt", v.data("x")), v.item(1, v.data("y")))))));
+    assert.deepEqual(mergedUpper.tags, { COMMENT: "x;y" });
+    assert.deepEqual(selectVideoCarrier(mergedUpper.tags).records, {});
+    // Matroska keeps the last spelling and value; a string-less SimpleTag deletes any spelling.
+    assert.deepEqual((await read(v.mkv([v.tagsEl(v.tag([v.simpleTag("PROMPT", GRAPH), v.simpleTag("Prompt", "B")]))]))).tags, { Prompt: "B" });
+    assert.deepEqual((await read(v.mkv([v.tagsEl(v.tag([v.simpleTag("PROMPT", GRAPH), v.simpleTag("Prompt", undefined)]))]))).tags, {});
+    // ffprobe's format_tags filter folds ASCII only: a Kelvin-sign "workflow" is never reported.
+    assert.deepEqual((await read(mdta(["prompt", "worKflow"], [[1, GRAPH], [2, WORKFLOW]]))).tags, { prompt: GRAPH });
+    // data types 0 and 3 are Mac Roman to ffmpeg (types 1 and 4 UTF-8, type 2 not text).
+    const cafe = [0x63, 0x61, 0x66, 0x8e];
+    for (const [type, expected] of [[0, "café"], [3, "café"], [1, "caf�"], [4, "caf�"]]) {
+        assert.deepEqual((await read(mdta(["prompt"], [[1, cafe, type]]))).tags, { prompt: expected }, `type ${type}`);
+    }
+    assert.deepEqual((await read(mdta(["prompt"], [[1, cafe, 2]]))).tags, {});
+    const high = Uint8Array.from({ length: 128 }, (_, index) => 0x80 + index);
+    const roomy = Buffer.concat([high, Buffer.alloc(400, 0x7c)]);
+    assert.equal((await read(mdta(["prompt"], [[1, roomy, 0]]))).tags.prompt, new TextDecoder("macintosh").decode(roomy));
+    // ffmpeg writes the UTF-8 into twice the input size: a long non-ASCII value is cut,
+    // here inside U+00DA, which ffprobe then replaces (its own output for these bytes).
+    const tight = Buffer.concat([high, Buffer.from([0x7c, 0x01, 0x7f])]);
+    assert.equal((await read(mdta(["prompt"], [[1, tight, 0]]))).tags.prompt, new TextDecoder("macintosh").decode(high).slice(0, 114) + "�");
+});
+
+test("budgets: roles share 2 MiB, an oversized workflow is left unread, duplicates and the total are bounded", async () => {
+    const limited = { status: "limited", reason: "metadata-limit" };
+    // Role tags over 2 MiB together stop the read (the server fails the whole read).
+    assert.deepEqual(await read(v.perKeyMp4({ prompt: fill(0x41, ROLE_BUDGET + 1) })), limited);
+    assert.deepEqual(await read(v.perKeyMkv({ PROMPT: fill(0x41, ROLE_BUDGET - 10), PARAMETERS: fill(0x42, 11) })), limited);
+    const atRole = await read(v.perKeyMp4({ prompt: fill(0x41, ROLE_BUDGET - 10), comfyassets_prompt_pairs: fill(0x42, 10) }));
+    assert.equal(atRole.status, "complete");
+    assert.equal(atRole.tags.prompt.length, ROLE_BUDGET - 10);
+    // A per-key workflow over 4 MiB is not sliced; the result says the workflow is limited.
+    const big = v.recordingBlob(v.perKeyMp4({ prompt: GRAPH, workflow: fill(0x7b, WORKFLOW_BUDGET + 1) }));
+    const over = await readVideoTags(big);
+    // It stays in the tag map as a marker, so the carrier choice still sees a per-key workflow.
+    assert.deepEqual(over, { status: "complete", container: "iso-bmff", format: "isom", tags: { prompt: GRAPH, workflow: m.OVERSIZED_WORKFLOW }, limited: ["workflow"] });
+    assert.ok(big.ranges.every(([start, end]) => end - start < WORKFLOW_BUDGET), "the oversized workflow was never sliced");
+    const bigMkv = await read(v.perKeyMkv({ WORKFLOW: fill(0x7b, WORKFLOW_BUDGET + 1), PROMPT: GRAPH }));
+    assert.deepEqual([bigMkv.limited, Object.keys(bigMkv.tags)], [["workflow"], ["WORKFLOW", "PROMPT"]]);
+    const atWorkflow = await read(v.perKeyMp4({ workflow: fill(0x7b, WORKFLOW_BUDGET) }));
+    assert.deepEqual([atWorkflow.limited, atWorkflow.tags.workflow.length], [[], WORKFLOW_BUDGET]);
+    // Fold-duplicates draw from the duplicate budgets (2x each class), then stop.
+    // (ffmpeg keeps one tag per ASCII-folded name: Matroska keeps the last spelling and value.)
+    const twice = await read(v.perKeyMkv({ PROMPT: fill(0x41, ROLE_BUDGET), prompt: fill(0x42, ROLE_BUDGET) }));
+    assert.deepEqual(Object.keys(twice.tags), ["prompt"]);
+    assert.equal(twice.tags.prompt[0], "B");
+    const thrice = await read(v.perKeyMkv({ PROMPT: fill(0x41, ROLE_BUDGET), prompt: fill(0x41, ROLE_BUDGET), Prompt: fill(0x41, ROLE_BUDGET + 1) }));
+    assert.deepEqual(thrice, limited);
+    const workflowTwice = await read(v.perKeyMkv({ WORKFLOW: fill(0x7b, WORKFLOW_BUDGET), workflow: fill(0x7b, WORKFLOW_BUDGET) }));
+    assert.equal(workflowTwice.status, "complete");
+    // Envelopes count against the total only; the total is the server's probe bound.
+    const envelope = await read(v.mp4(v.legacyUdta({ comment: fill(0x41, 5 * MiB) })));
+    assert.equal(envelope.tags.comment.length, 5 * MiB);
+    const tooMuch = await read(v.mp4(v.legacyUdta({ comment: fill(0x41, TAG_READ_BYTES - 10), description: fill(0x42, 11) })));
+    assert.deepEqual(tooMuch, limited);
+    assert.equal(m.VIDEO_TAG_READ_BYTES, TAG_READ_BYTES);
+});
+
+test("an oversized per-key workflow keeps its place in the carrier choice, as in the server's first probe", async () => {
+    const big = fill(0x7b, WORKFLOW_BUDGET + 1);
+    const envelope = '{"workflow": {"nodes": [1]}}';
+    const beside = (comment) => v.mp4(v.udta(v.meta(v.hdlr("mdta"), v.keys(["workflow"]), v.ilst(v.item(1, v.data(big)), v.legacyItem("©cmt", v.data(comment))))));
+    const read1 = await read(beside(envelope));
+    assert.deepEqual([read1.tags, read1.limited], [{ workflow: m.OVERSIZED_WORKFLOW, comment: envelope }, ["workflow"]]);
+    // The server picks the per-key carrier (no role on either side) and marks the workflow limited.
+    assert.deepEqual(selectVideoCarrier(read1.tags, { limited: read1.limited }), { status: "complete", records: {}, record_formats: { workflow: "limited" }, carrierText: null, limited: ["workflow"] });
+    // A role envelope still wins and clears it.
+    const roleEnvelope = '{"prompt": "{}", "workflow": {"a": 1}}';
+    const read2 = await read(beside(roleEnvelope));
+    assert.deepEqual(selectVideoCarrier(read2.tags, { limited: read2.limited }), { status: "complete", records: { prompt: "{}", workflow: '{"a": 1}' }, record_formats: { prompt: "original-text", workflow: "reserialized-json" }, carrierText: roleEnvelope, limited: [] });
+    // ISO-BMFF joins a repeat onto the oversized value, which stays over budget and unread.
+    const repeat = (extra) => v.mp4(v.udta(v.meta(v.hdlr("mdta"), v.keys(["workflow", "prompt"]), v.ilst(v.item(1, v.data(big)), v.item(1, v.data(extra)), v.item(2, v.data(GRAPH))))));
+    const repeatedBlob = v.recordingBlob(repeat(WORKFLOW));
+    const repeated = await readVideoTags(repeatedBlob);
+    assert.deepEqual([repeated.tags, repeated.limited], [{ workflow: m.OVERSIZED_WORKFLOW, prompt: GRAPH }, ["workflow"]]);
+    assert.ok(repeatedBlob.ranges.every(([start, end]) => end - start < 1024), "neither workflow part was sliced");
+    // When the first probe would overflow its bound, the server re-probes without the
+    // per-key workflow: then it takes no part in the choice ...
+    const dropped = await read(repeat(fill(0x41, TAG_READ_BYTES - WORKFLOW_BUDGET)));
+    assert.deepEqual([dropped.tags, dropped.limited], [{ prompt: GRAPH }, ["workflow"]]);
+    // ... and the re-probe's own bound (twice the role budget plus 64 KiB) applies to the rest.
+    const wide = '{"workflow": {"n": "' + "x".repeat(TAG_READ_BYTES - WORKFLOW_BUDGET) + '"}}';
+    assert.deepEqual(await read(beside(wide)), { status: "limited", reason: "metadata-limit" });
+    assert.equal(m.VIDEO_REPROBE_BYTES, 2 * ROLE_BUDGET + 64 * 1024);
+    // Matroska keeps the last value of a name, and a string-less SimpleTag deletes it.
+    const replaced = await read(v.mkv([v.tagsEl(v.tag([v.simpleTag("WORKFLOW", big), v.simpleTag("WORKFLOW", WORKFLOW)]))]));
+    assert.deepEqual([replaced.tags, replaced.limited], [{ WORKFLOW: WORKFLOW }, []]);
+    const deleted = await read(v.mkv([v.tagsEl(v.tag([v.simpleTag("WORKFLOW", big), v.simpleTag("WORKFLOW", undefined), v.simpleTag("PROMPT", GRAPH)]))]));
+    assert.deepEqual([deleted.tags, deleted.limited], [{ PROMPT: GRAPH }, []]);
+});
+
+test("avi and asf are recognised but not read", async () => {
+    const avi = v.concat(v.latin1("RIFF"), v.u32(1000), v.latin1("AVI LIST"), new Uint8Array(16));
+    const asf = v.concat([0x30, 0x26, 0xb2, 0x75, 0x8e, 0x66, 0xcf, 0x11, 0xa6, 0xd9, 0x00, 0xaa, 0x00, 0x62, 0xce, 0x6c], new Uint8Array(16));
+    for (const bytes of [avi, asf]) {
+        const blob = v.recordingBlob(v.concat(bytes, new Uint8Array(4096)));
+        assert.deepEqual(await readVideoTags(blob, { prefix: bytes }), { status: "unsupported", reason: "video-container-unsupported" });
+        assert.deepEqual(blob.ranges, []);
+    }
+    // The classifier's kind can be handed in instead of a prefix.
+    const blob = v.recordingBlob(new Uint8Array(64));
+    assert.deepEqual(await readVideoTags(blob, { kind: "avi" }), { status: "unsupported", reason: "video-container-unsupported" });
+    assert.deepEqual(blob.ranges, []);
+    // Without a prefix the reader slices the first 32 bytes itself.
+    const mp4 = v.recordingBlob(v.perKeyMp4({ prompt: GRAPH }));
+    assert.deepEqual((await readVideoTags(mp4)).tags, { prompt: GRAPH });
+    assert.deepEqual(mp4.ranges[0], [0, 32]);
+});
+
+test("abort during a moov walk settles once", async () => {
+    const unhandled = [];
+    const onUnhandled = (reason) => unhandled.push(reason);
+    process.on("unhandledRejection", onUnhandled);
+    try {
+        const bytes = v.concat(v.ftyp(), ...Array.from({ length: 50 }, () => v.free()), v.mdat(), v.moov(v.mdtaUdta({ prompt: GRAPH, workflow: WORKFLOW })));
+        const controller = new AbortController();
+        let slices = 0;
+        const blob = {
+            size: bytes.length,
+            slice(start, end) {
+                slices += 1;
+                if (slices === 3) controller.abort();
+                return { arrayBuffer: () => new Promise((resolve) => setTimeout(() => resolve(bytes.slice(start, end).buffer), 2)) };
+            },
+        };
+        await assert.rejects(readVideoTags(blob, { signal: controller.signal }), { name: "AbortError" });
+        const already = new AbortController();
+        already.abort();
+        await assert.rejects(readVideoTags(v.blobOf(bytes), { signal: already.signal }), { name: "AbortError" });
+        const slow = v.slowBlob(bytes, 1);
+        const during = new AbortController();
+        const pending = readVideoTags(slow, { signal: during.signal, yieldControl: async () => during.abort() });
+        await assert.rejects(pending, { name: "AbortError" });
+        await new Promise((resolve) => setTimeout(resolve, 10));
+        assert.deepEqual(unhandled, []);
+    } finally {
+        process.off("unhandledRejection", onUnhandled);
+    }
+});
+
+test("yieldControl is awaited while walking many boxes", async () => {
+    let yields = 0;
+    const many = v.concat(v.ftyp(), ...Array.from({ length: 3000 }, () => v.free()), v.mdat(), v.moov(v.mdtaUdta({ prompt: GRAPH })));
+    const result = await readVideoTags(v.slowBlob(many, 0), { yieldControl: async () => { yields += 1; } });
+    assert.equal(result.status, "complete");
+    assert.ok(yields > 0, "the reader handed control back");
+});
+
+test("carrier selection: per-key tags, then the comment envelope, then description, never merged", () => {
+    const ENVELOPE_WORKFLOW = '{"last_node_id": 9, "nodes": [{"id": 3, "type": "KSampler", "widgets_values": [7, "fixed", 20, 8.0, "euler", "normal", 1.0]}, {"id": 6, "type": "CLIPTextEncode", "widgets_values": ["a red fox in snow, detailed fur"]}, {"id": 7, "type": "CLIPTextEncode", "widgets_values": ["blurry, lowres"]}], "links": [], "version": 0.4}';
+    const original = (keys) => Object.fromEntries(keys.map((key) => [key, "original-text"]));
+    const text = (tags, keys) => Object.fromEntries(keys.map((key) => [key, tags[key] ?? tags[key.toUpperCase()]]));
+    // The recorded fixtures.
+    for (const name of ["core-faststart.mp4", "core-moov-end.mp4", "core.mov", "core.m4v", "core.mkv", "core-tags.webm"]) {
+        const tags = RECORDED[name];
+        const keys = Object.keys(tags).map((key) => key.toLowerCase()).filter((key) => PER_KEY.has(key));
+        assert.deepEqual(selectVideoCarrier(tags), { status: "complete", records: text(tags, keys), record_formats: original(keys), carrierText: null, limited: [] }, name);
+    }
+    const legacy = RECORDED["legacy-comment.mp4"];
+    const comment = pythonJsonLoads(legacy.comment);
+    assert.deepEqual(selectVideoCarrier(legacy), {
+        status: "complete",
+        records: { prompt: comment.get("prompt"), workflow: ENVELOPE_WORKFLOW },
+        record_formats: { prompt: "original-text", workflow: "reserialized-json" },
+        carrierText: legacy.comment,
+        limited: [],
+    });
+    assert.ok(comment.get("prompt").startsWith('{"3":{"class_type":"KSampler"'));
+    // Matroska upper-cases COMMENT, and envelope names are exact-case (ruling M1): nothing is read.
+    assert.deepEqual(selectVideoCarrier(RECORDED["legacy-comment.webm"]), { status: "complete", records: {}, record_formats: {}, carrierText: null, limited: [] });
+    // Synthetic maps.
+    const envelope = JSON.stringify({ prompt: '{"9":{}}', workflow: { v: 1 } });
+    const cases = [
+        ["per-key wins over comment", { prompt: GRAPH, workflow: WORKFLOW, comment: envelope }, { records: { prompt: GRAPH, workflow: WORKFLOW }, record_formats: original(["prompt", "workflow"]), carrierText: null }],
+        ["Matroska names fold", { PROMPT: GRAPH, Workflow: WORKFLOW, ENCODER: "Lavf" }, { records: { prompt: GRAPH, workflow: WORKFLOW }, record_formats: original(["prompt", "workflow"]), carrierText: null }],
+        ["identical fold-duplicates are one record", { PROMPT: GRAPH, prompt: GRAPH }, { records: { prompt: GRAPH }, record_formats: original(["prompt"]), carrierText: null }],
+        ["differing fold-duplicates fail", { PROMPT: GRAPH, prompt: '{"9":{}}' }, { status: "unreadable", reason: "duplicate-metadata" }],
+        ["differing workflow duplicates fail even with a prompt", { Workflow: WORKFLOW, WORKFLOW: "{}", prompt: GRAPH }, { status: "unreadable", reason: "duplicate-metadata" }],
+        ["envelope names never fold", { comment: envelope, COMMENT: "a free-text note" }, { records: { prompt: '{"9":{}}', workflow: '{"v": 1}' }, record_formats: { prompt: "original-text", workflow: "reserialized-json" }, carrierText: envelope }],
+        ["non-JSON comment falls through to description", { comment: "not json", description: envelope }, { records: { prompt: '{"9":{}}', workflow: '{"v": 1}' }, record_formats: { prompt: "original-text", workflow: "reserialized-json" }, carrierText: envelope }],
+        ["a JSON array comment is not an envelope", { comment: '["prompt"]', description: '{"workflow": {}}' }, { records: { workflow: "{}" }, record_formats: { workflow: "reserialized-json" }, carrierText: '{"workflow": {}}' }],
+        ["a role envelope beats a workflow-only per-key carrier", { workflow: WORKFLOW, comment: envelope }, { records: { prompt: '{"9":{}}', workflow: '{"v": 1}' }, record_formats: { prompt: "original-text", workflow: "reserialized-json" }, carrierText: envelope }],
+        ["role description beats workflow-only comment", { comment: '{"workflow": {"a": 1}}', description: '{"parameters": "cat"}' }, { records: { parameters: "cat" }, record_formats: { parameters: "original-text" }, carrierText: '{"parameters": "cat"}' }],
+        ["workflow-only carriers: the first wins, never merged", { comment: '{"workflow": {"a": 1}}', description: '{"workflow": {"b": 2}}' }, { records: { workflow: '{"a": 1}' }, record_formats: { workflow: "reserialized-json" }, carrierText: '{"workflow": {"a": 1}}' }],
+        ["description only", { description: envelope }, { records: { prompt: '{"9":{}}', workflow: '{"v": 1}' }, record_formats: { prompt: "original-text", workflow: "reserialized-json" }, carrierText: envelope }],
+        ["no known key", { encoder: "Lavf", major_brand: "isom" }, { records: {}, record_formats: {}, carrierText: null }],
+        ["envelope with only unknown keys is no carrier", { comment: '{"title": "x"}' }, { records: {}, record_formats: {}, carrierText: null }],
+        ["null role marker is refused", { comment: '{"prompt": null, "workflow": {}}' }, { status: "unreadable", reason: "invalid-metadata" }],
+        ["numeric role is refused", { description: '{"parameters": 5}' }, { status: "unreadable", reason: "invalid-metadata" }],
+        ["a non-text non-object workflow is invalid but the roles stay", { comment: '{"prompt": "{}", "workflow": null}' }, { records: { prompt: "{}" }, record_formats: { prompt: "original-text", workflow: "invalid" }, carrierText: '{"prompt": "{}", "workflow": null}' }],
+        ["an object prompt is reserialized", { comment: '{"prompt": {"1": {"class_type": "A"}}}' }, { records: { prompt: '{"1": {"class_type": "A"}}' }, record_formats: { prompt: "reserialized-json" }, carrierText: '{"prompt": {"1": {"class_type": "A"}}}' }],
+        ["a non-text per-key role marker is refused", { prompt: 5, workflow: WORKFLOW }, { status: "unreadable", reason: "invalid-metadata" }],
+        ["a non-text per-key workflow is invalid", { prompt: GRAPH, workflow: 5 }, { records: { prompt: GRAPH }, record_formats: { prompt: "original-text", workflow: "invalid" }, carrierText: null }],
+        ["a workflow-only envelope with an invalid workflow is still the carrier", { comment: '{"workflow": 5}' }, { records: {}, record_formats: { workflow: "invalid" }, carrierText: '{"workflow": 5}' }],
+        ["not a map", null, { records: {}, record_formats: {}, carrierText: null }],
+    ];
+    for (const [name, tags, expected] of cases) {
+        const want = expected.status ? expected : { status: "complete", limited: [], ...expected };
+        assert.deepEqual(selectVideoCarrier(tags), want, name);
+    }
+    // The server's budget loop: a 4 MiB workflow is limited, roles over 2 MiB fail, exact sizes pass.
+    const bigWorkflow = "{" + "0".repeat(WORKFLOW_BUDGET);
+    assert.deepEqual(selectVideoCarrier({ prompt: GRAPH, workflow: bigWorkflow }), { status: "complete", records: { prompt: GRAPH }, record_formats: { prompt: "original-text", workflow: "limited" }, carrierText: null, limited: ["workflow"] });
+    assert.deepEqual(selectVideoCarrier({ workflow: bigWorkflow.slice(1) }).records, { workflow: bigWorkflow.slice(1) });
+    assert.deepEqual(selectVideoCarrier({ comment: JSON.stringify({ prompt: "x", workflow: { k: "0".repeat(WORKFLOW_BUDGET) } }) }).record_formats, { prompt: "original-text", workflow: "limited" });
+    assert.deepEqual(selectVideoCarrier({ prompt: "a".repeat(ROLE_BUDGET), parameters: "b" }), { status: "limited", reason: "metadata-limit" });
+    assert.deepEqual(selectVideoCarrier({ prompt: "é".repeat(ROLE_BUDGET / 2) }).records.prompt.length, ROLE_BUDGET / 2);
+    assert.deepEqual(selectVideoCarrier({ prompt: "é".repeat(ROLE_BUDGET / 2) + "a" }), { status: "limited", reason: "metadata-limit" });
+    assert.deepEqual(selectVideoCarrier({ prompt: "\ud800".repeat(ROLE_BUDGET / 3) + "aaa" }), { status: "limited", reason: "metadata-limit" });
+    assert.deepEqual(selectVideoCarrier({ prompt: "\u{1f600}".repeat(ROLE_BUDGET / 4) }).records.prompt.length, ROLE_BUDGET / 2);
+    assert.deepEqual(selectVideoCarrier({ prompt: "\u{1f600}".repeat(ROLE_BUDGET / 4) + "a" }), { status: "limited", reason: "metadata-limit" });
+    // The reader's limited workflow (the server's role-only re-probe): kept with the per-key
+    // carrier or no carrier, cleared when an envelope wins.
+    assert.deepEqual(selectVideoCarrier({ prompt: GRAPH }, { limited: ["workflow"] }).record_formats, { prompt: "original-text", workflow: "limited" });
+    assert.deepEqual(selectVideoCarrier({ encoder: "x" }, { limited: ["workflow"] }), { status: "complete", records: {}, record_formats: { workflow: "limited" }, carrierText: null, limited: ["workflow"] });
+    assert.deepEqual(selectVideoCarrier({ comment: envelope }, { limited: ["workflow"] }).limited, []);
+    assert.deepEqual(selectVideoCarrier({ comment: envelope }, { limited: ["workflow"] }).record_formats, { prompt: "original-text", workflow: "reserialized-json" });
+});
+
+test("pythonJsonLoads and pythonJsonDumps spell values as Python's json module does", () => {
+    const rows = [
+        ["1.0", "1.0"], ["8.0", "8.0"], ["0.1", "0.1"], ["1e16", "1e+16"], ["1e15", "1000000000000000.0"],
+        ["123456789012345678", "123456789012345678"], ["0.0001", "0.0001"], ["0.00001", "1e-05"], ["1e-7", "1e-07"],
+        ["1.5e300", "1.5e+300"], ["1e400", "Infinity"], ["-0", "0"], ["-0.0", "-0.0"], ["1E2", "100.0"], ["2.5", "2.5"],
+        ["100.0", "100.0"], ["1e21", "1e+21"], ["123456.789", "123456.789"], ["0.000123", "0.000123"], ["-1e-05", "-1e-05"],
+        ["9007199254740993", "9007199254740993"], ["1.7976931348623157e308", "1.7976931348623157e+308"], ["5e-324", "5e-324"],
+        ["NaN", "NaN"], ["-Infinity", "-Infinity"], ["Infinity", "Infinity"], ["[1.0e+2]", "[100.0]"], ["-1e400", "-Infinity"],
+        ["[]", "[]"], ["{}", "{}"], ["[true, false, null]", "[true, false, null]"], ['{"k":[1,{"z":1}]}', '{"k": [1, {"z": 1}]}'],
+        [' \t\n\r{"a" : 1 } ', '{"a": 1}'],
+        ['{"a":"é\\u0001\u007f\\"\\\\\\n\\t\\b\\f\\r\\/ 😀 \\ud800"}', '{"a": "\\u00e9\\u0001\\u007f\\"\\\\\\n\\t\\b\\f\\r/ \\ud83d\\ude00 \\ud800"}'],
+        ['{"a":"\\ud83d\\ude00"}', '{"a": "\\ud83d\\ude00"}'],
+        ['{"10":1,"3":2,"b":3,"a":4}', '{"10": 1, "3": 2, "b": 3, "a": 4}'],
+        ['{"a":1,"b":2,"a":3}', '{"a": 3, "b": 2}'],
+    ];
+    for (const [input, output] of rows) {
+        assert.equal(pythonJsonDumps(pythonJsonLoads(input)), output, input);
+    }
+    const bad = ['{"a":1} x', '{"a":"\u0001"}', '{"a":"\\x"}', "01", "1.", ".5", "+1", "[1,]", '{"a":1,}', "{a:1}", "'a'", "", "   ", "[", "nul", "-", "1e", '"\\u12"', "[1 2]", "tru", "Infinit"];
+    for (const input of bad) {
+        assert.throws(() => pythonJsonLoads(input), { name: "PythonJsonError" }, JSON.stringify(input));
+    }
+    assert.throws(() => pythonJsonLoads("[".repeat(2000) + "]".repeat(2000)), { name: "PythonJsonError", recursion: true });
+    // CPython's int() refuses more than 4300 digits (sign excluded); floats have no such limit.
+    const digits = (count) => "1" + "0".repeat(count - 1);
+    assert.equal(pythonJsonDumps(pythonJsonLoads(digits(4300))), digits(4300));
+    assert.equal(pythonJsonDumps(pythonJsonLoads("-" + digits(4300))), "-" + digits(4300));
+    for (const input of [digits(4301), "-" + digits(4301), `[${digits(5000)}]`]) {
+        assert.throws(() => pythonJsonLoads(input), (error) => error.name === "PythonJsonError" && error.recursion === false, input.slice(0, 8));
+    }
+    assert.equal(pythonJsonLoads(digits(4301) + ".0").isFloat, true);
+    // So an envelope holding such an integer is skipped, as the server's _carriers does.
+    const hugeInt = `{"workflow": {"a": ${digits(4400)}}}`;
+    assert.deepEqual(selectVideoCarrier({ comment: hugeInt, description: '{"workflow": {"b": 2}}' }).carrierText, '{"workflow": {"b": 2}}');
+    assert.throws(() => pythonJsonDumps(undefined), { name: "PythonJsonError" });
+    assert.throws(() => pythonJsonLoads(5), { name: "PythonJsonError" });
+    assert.equal(pythonJsonDumps(pythonJsonLoads("[".repeat(500) + "]".repeat(500))), "[".repeat(500) + "]".repeat(500));
+    const parsed = pythonJsonLoads('{"s":"x","n":1,"f":1.5,"b":true,"z":null,"l":[1]}');
+    assert.ok(parsed instanceof Map);
+    assert.deepEqual([...parsed.keys()], ["s", "n", "f", "b", "z", "l"]);
+    assert.equal(parsed.get("s"), "x");
+    assert.equal(parsed.get("b"), true);
+    assert.equal(parsed.get("z"), null);
+    assert.ok(Array.isArray(parsed.get("l")));
+});

tests/js/workspace-toolbar.test.js

@@ -765,6 +765,20 @@ test('the Filters button shows the active refinement count', async () => {
         assert.equal(filtersButton.textContent, 'Filters');
         for (const bad of [-1, 1.5, '1', null, Number.NaN]) assert.throws(() => toolbar.setActiveFilters(bad), TypeError);
         assert.equal(filtersButton.textContent, 'Filters');
+        // An unchanged count leaves the label's text node alone: every listing reload reports the count, and WebKit
+        // drops a click whose mousedown hit a text node that was replaced before mouseup.
+        let writes = 0;
+        const proto = Object.getPrototypeOf(filtersButton);
+        const descriptor = Object.getOwnPropertyDescriptor(proto, 'textContent');
+        Object.defineProperty(filtersButton, 'textContent', {
+            get() { return descriptor.get.call(this); },
+            set(text) { writes += 1; descriptor.set.call(this, text); },
+            configurable: true,
+        });
+        toolbar.setActiveFilters(0);
+        assert.equal(writes, 0);
+        toolbar.setActiveFilters(3);
+        assert.deepEqual([writes, filtersButton.textContent], [1, 'Filters · 3']);
         assert.throws(() => mountToolbar(new dom.Element('div'), { scope, categories: [], sorts: [['newest', 'Recently added']],
             activeFilters: -1, onChange() {}, onFilters() {}, onSaveView() {} }), TypeError);
     } finally {

tests/js/workspace-vocabulary-state.test.js

@@ -507,10 +507,49 @@ test('category preview text states the real impact and whether it renames or con
         '“Landscapes” is used by 2 other saved prompts. Apply renames it to “Scenery”.');
 });
 
+test('accepts image_count and names images in the delete text', async () => {
+    const {createVocabularyStream, createVocabularyCommand, readTagImpact, deleteTagText} = await helpers;
+    const rows = [{...tag(1, 'mist', 2), image_count: 3}, tag(2, 'dusk', 0)];
+    const stream = createVocabularyStream({api: queued([tagPage(rows)]).api, kind: 'tags'});
+    assert.equal(await stream.loadMore(), 2);
+    assert.deepEqual(stream.state.items.map(row => row.image_count), [3, undefined]);
+    for (const bad of [{...tag(1), image_count: -1}, {...tag(1), image_count: '3'}, {...tag(1), image_count: 1.5}]) {
+        const rejected = createVocabularyStream({api: queued([tagPage([bad])]).api, kind: 'tags'});
+        await rejected.loadMore();
+        assert.ok(rejected.state.error instanceof TypeError, JSON.stringify(bad));
+        assert.deepEqual(rejected.state.items, []);
+    }
+    const withImages = {...impact, images: 2};
+    const api = {get: async () => ({success: true, library_epoch: EPOCH, impact: withImages})};
+    assert.deepEqual((await readTagImpact(api, {id: 4, name: 'mist'})).impact, withImages);
+    await assert.rejects(readTagImpact({get: async () => ({success: true, library_epoch: EPOCH, impact: {...impact, images: -2}})},
+        {id: 4, name: 'mist'}), TypeError);
+    const command = createVocabularyCommand({api: {
+        patch: async () => renameReceipt({tag: {...tag(1, 'dusk', 2), image_count: 1}, images_changed: 1}),
+        post: async () => ({success: true, library_epoch: EPOCH, deleted: true, components_changed: 0, pairs_changed: 0, images_changed: 2}),
+    }});
+    await command.renameTag({id: 1, name: 'dawn', prompt_count: 2}, 'dusk', EPOCH);
+    assert.equal(command.state.phase, 'done');
+    assert.equal(command.state.receipt.images_changed, 1);
+    command.reset();
+    await command.deleteTag({id: 1, name: 'dusk'}, EPOCH);
+    assert.equal(command.state.phase, 'done');
+    const malformed = createVocabularyCommand({api: {post: async () => ({success: true, library_epoch: EPOCH, deleted: true,
+        components_changed: 0, pairs_changed: 0, images_changed: 'two'})}});
+    await malformed.deleteTag({id: 1, name: 'dusk'}, EPOCH);
+    assert.equal(malformed.state.phase, 'unknown-outcome');
+    assert.equal(deleteTagText('mist', withImages),
+        'Deleting “mist” removes the name and its links from 3 positive prompts, 3 negative variants (1 archived), 1 other saved prompt and 2 images. The prompts, negative variants, images and history themselves are kept.');
+    assert.equal(deleteTagText('mist', {...zero, images: 1}),
+        'Deleting “mist” removes the name and its links from 1 image. The prompts, negative variants, images and history themselves are kept.');
+    assert.equal(deleteTagText('unused', {...zero, images: 0}),
+        '“unused” is not linked to any saved prompt or image. Deleting it removes only the name.');
+});
+
 test('delete text names the tag, its real links and what stays unchanged', async () => {
     const {deleteTagText} = await helpers;
     assert.equal(deleteTagText('mist', impact),
         'Deleting “mist” removes the name and its links from 3 positive prompts, 3 negative variants (1 archived) and 1 other saved prompt. The prompts, negative variants, images and history themselves are kept.');
     assert.equal(deleteTagText('unused', zero),
-        '“unused” is not linked to any saved prompt. Deleting it removes only the name.');
+        '“unused” is not linked to any saved prompt or image. Deleting it removes only the name.');
 });

tests/test_annotation_schema.py

@@ -0,0 +1,602 @@
+"""Image annotation schema group (#249): install, validation and restore safety."""
+
+import hashlib
+import re
+import sqlite3
+import tempfile
+import unittest
+from contextlib import closing
+from pathlib import Path
+from unittest.mock import patch
+
+try:
+    from tests.open_handles import assert_closed
+except ImportError:  # discovered with tests/ as the top-level directory
+    from open_handles import assert_closed
+
+from database.models import PromptModel
+
+try:  # guarded so a missing module fails assertions, not test loading
+    from database.annotation_schema import (
+        annotation_schema_error,
+        migrate_annotation_state,
+    )
+except ImportError:
+    annotation_schema_error = None
+    migrate_annotation_state = None
+
+METADATA = "media_annotation_metadata"
+FILES = "media_annotation_files"
+BYTES = "media_annotation_bytes"
+IMAGE_TAGS = "media_image_tags"
+TABLES = (METADATA, FILES, BYTES, IMAGE_TAGS)
+# The exact DDL the plan pins; the module must produce precisely these objects.
+DDL = {
+    METADATA: (
+        f"CREATE TABLE {METADATA}(singleton_id INTEGER PRIMARY KEY "
+        "CHECK(singleton_id=1), schema_version INTEGER NOT NULL "
+        "CHECK(schema_version=1), image_index_revision INTEGER NOT NULL DEFAULT 0 "
+        "CHECK(image_index_revision>=0))"
+    ),
+    FILES: (
+        f"CREATE TABLE {FILES}(id TEXT PRIMARY KEY CHECK(length(id)=36), "
+        "locator TEXT NOT NULL UNIQUE, created_at TEXT NOT NULL)"
+    ),
+    BYTES: (
+        f"CREATE TABLE {BYTES}(id TEXT PRIMARY KEY CHECK(length(id)=36), "
+        f"file_id TEXT NOT NULL REFERENCES {FILES}(id) ON DELETE CASCADE, "
+        "sha256 TEXT NOT NULL CHECK(length(sha256)=64 AND sha256 NOT GLOB "
+        "'*[^0-9a-f]*'), size INTEGER NOT NULL CHECK(size>=0), "
+        "mtime_ns INTEGER NOT NULL, verified_at TEXT NOT NULL, "
+        "first_seen_at TEXT NOT NULL, tag_revision INTEGER NOT NULL DEFAULT 0 "
+        "CHECK(tag_revision>=0), UNIQUE(file_id, sha256))"
+    ),
+    IMAGE_TAGS: (
+        f"CREATE TABLE {IMAGE_TAGS}(byte_id TEXT NOT NULL REFERENCES {BYTES}(id) "
+        "ON DELETE CASCADE, tag_id INTEGER NOT NULL REFERENCES tags(id) "
+        "ON DELETE CASCADE, position INTEGER NOT NULL CHECK(position>=0), "
+        "PRIMARY KEY(byte_id, tag_id))"
+    ),
+}
+INDEXES = {
+    "idx_media_annotation_bytes_fingerprint": (
+        f"CREATE INDEX idx_media_annotation_bytes_fingerprint ON {BYTES}"
+        "(file_id, size, mtime_ns)"
+    ),
+    "idx_media_image_tags_tag": (
+        f"CREATE INDEX idx_media_image_tags_tag ON {IMAGE_TAGS}(tag_id)"
+    ),
+    "idx_media_image_tags_position": (
+        f"CREATE UNIQUE INDEX idx_media_image_tags_position ON {IMAGE_TAGS}"
+        "(byte_id, position)"
+    ),
+}
+AUTOINDEXES = {FILES: 2, BYTES: 2, IMAGE_TAGS: 1}
+LEGACY_TABLES = (
+    "prompts",
+    "generated_images",
+    "tags",
+    "prompt_tags",
+    "prompt_pair_tags",
+    "image_favorites",
+)
+FILE_ID = "0f8e4b2a-1c3d-4e5f-8a9b-0c1d2e3f4a5b"
+BYTE_ID = "6a1b2c3d-4e5f-4a6b-8c9d-0e1f2a3b4c5d"
+OTHER_ID = "7b2c3d4e-5f6a-4b7c-9d0e-1f2a3b4c5d6e"
+DIGEST = hashlib.sha256(b"img").hexdigest()
+NOW = "2026-10-22T00:00:00.000000+00:00"
+MAX_LOCATOR = 4096
+
+
+def _compact(sql):
+    return re.sub(r"\s+", "", sql or "").lower()
+
+
+def expected_catalog():
+    objects = {("table", name, _compact(sql)) for name, sql in DDL.items()}
+    objects |= {("index", name, _compact(sql)) for name, sql in INDEXES.items()}
+    objects |= {
+        ("index", f"sqlite_autoindex_{table}_{n}", "")
+        for table, count in AUTOINDEXES.items()
+        for n in range(1, count + 1)
+    }
+    return objects
+
+
+def catalog(conn):
+    names = tuple(TABLES) + tuple(INDEXES)
+    marks = ",".join("?" * len(names))
+    table_marks = ",".join("?" * len(TABLES))
+    return {
+        (kind, name, _compact(sql))
+        for kind, name, sql in conn.execute(
+            "SELECT type,name,sql FROM sqlite_master "
+            f"WHERE name IN ({marks}) OR tbl_name IN ({table_marks})",  # nosec B608
+            names + tuple(TABLES),
+        )
+    }
+
+
+def insert_group_rows(conn, tag_name="img-only", locator="/out/img.png"):
+    """One file, one byte record and one image tag carrying *tag_name*."""
+    tag_id = conn.execute("INSERT INTO tags(name) VALUES(?)", (tag_name,)).lastrowid
+    conn.execute(f"INSERT INTO {FILES} VALUES(?,?,?)", (FILE_ID, locator, NOW))
+    conn.execute(
+        f"INSERT INTO {BYTES} VALUES(?,?,?,?,?,?,?,?)",
+        (BYTE_ID, FILE_ID, DIGEST, 3, 1_700_000_000_000_000_000, NOW, NOW, 1),
+    )
+    conn.execute(f"INSERT INTO {IMAGE_TAGS} VALUES(?,?,0)", (BYTE_ID, tag_id))
+    return tag_id
+
+
+def rows(conn, sql, *params):
+    """Plain tuples whatever row factory *conn* uses."""
+    return [tuple(row) for row in conn.execute(sql, params)]
+
+
+def image_tags(conn):
+    return rows(
+        conn,
+        f"SELECT t.name, it.position FROM {IMAGE_TAGS} it "
+        "JOIN tags t ON t.id = it.tag_id ORDER BY it.position",
+    )
+
+
+def metadata(conn):
+    return rows(conn, f"SELECT schema_version,image_index_revision FROM {METADATA}")
+
+
+def table_rows(conn, tables):
+    found = {}
+    for table in tables:
+        present = conn.execute(
+            "SELECT 1 FROM sqlite_master WHERE type='table' AND name=?", (table,)
+        ).fetchone()
+        if present:
+            found[table] = rows(conn, f"SELECT * FROM {table}")  # nosec B608
+    return found
+
+
+def populate_legacy(conn):
+    conn.execute(
+        "INSERT INTO prompts(id,text,hash,category,notes,run_count) VALUES"
+        "(7,'positive','p','keep','notes',12)"
+    )
+    conn.execute("INSERT INTO prompts(id,text,hash) VALUES(9,'negative','n')")
+    conn.execute(
+        "INSERT INTO generated_images(id,prompt_id,image_path,filename,file_path) "
+        "VALUES(4,7,'/tmp/image.png','image.png','/tmp/image.png')"
+    )
+    conn.execute("INSERT INTO tags(id,name) VALUES(3,'keep-tag')")
+    conn.execute("INSERT INTO prompt_tags VALUES(7,3)")
+    conn.execute(
+        "INSERT INTO prompt_pairs(id,positive_prompt_id,negative_prompt_id) "
+        "VALUES(6,7,9)"
+    )
+    conn.execute("INSERT INTO prompt_pair_tags VALUES(6,3)")
+    conn.execute("INSERT INTO image_favorites VALUES('root','a.png',?)", (NOW,))
+
+
+class AnnotationSchemaTests(unittest.TestCase):
+    def setUp(self):
+        self.temp = tempfile.TemporaryDirectory()
+        self.addCleanup(self.temp.cleanup)
+        self.addCleanup(assert_closed, self, self.temp.name)
+        self.addCleanup(PromptModel.reset_schema_cache)
+        self.models = []
+        self.addCleanup(self.close_models)
+        self.path = str(Path(self.temp.name) / "live.db")
+
+    def close_models(self):
+        for model in self.models:
+            model.close_all()
+
+    def model(self, path=None):
+        model = PromptModel(path or self.path)
+        self.models.append(model)
+        return model
+
+    def require_module(self):
+        self.assertIsNotNone(
+            annotation_schema_error, "database.annotation_schema is not implemented"
+        )
+
+    def drop_group(self, path):
+        """Remove the group from a closed database file (an old backup)."""
+        with closing(sqlite3.connect(path)) as raw, raw:
+            raw.execute("PRAGMA foreign_keys=OFF")
+            for table in reversed(TABLES):
+                raw.execute(f"DROP TABLE {table}")
+
+    def test_fresh_install_and_reopen_is_idempotent(self):
+        self.require_module()
+        model = self.model()
+        with model.get_connection() as conn:
+            self.assertEqual(catalog(conn), expected_catalog())
+            self.assertIsNone(annotation_schema_error(conn, allow_missing=False))
+            self.assertEqual(metadata(conn), [(1, 0)])
+            insert_group_rows(conn)
+        self.assertEqual(
+            PromptModel.verify_database_file(self.path, require_library_state=True),
+            (True, "ok"),
+        )
+        model.close_all()
+        for _ in range(2):
+            PromptModel.reset_schema_cache()
+            model = self.model()
+            with model.get_connection() as conn:
+                self.assertEqual(catalog(conn), expected_catalog())
+                self.assertIsNone(annotation_schema_error(conn, allow_missing=False))
+                self.assertEqual(metadata(conn), [(1, 0)])
+                self.assertEqual(image_tags(conn), [("img-only", 0)])
+                # Running the migration again on a current database changes nothing.
+                migrate_annotation_state(conn)
+                self.assertEqual(metadata(conn), [(1, 0)])
+                self.assertEqual(image_tags(conn), [("img-only", 0)])
+            model.close_all()
+        # A tag delete and a file delete cascade into the junction and bytes.
+        PromptModel.reset_schema_cache()
+        model = self.model()
+        with model.get_connection() as conn:
+            conn.execute("DELETE FROM tags WHERE name='img-only'")
+            self.assertEqual(image_tags(conn), [])
+            conn.execute(f"DELETE FROM {FILES}")
+            self.assertEqual(rows(conn, f"SELECT * FROM {BYTES}"), [])
+            self.assertIsNone(annotation_schema_error(conn, allow_missing=False))
+
+    def test_old_database_without_group_migrates_without_touching_legacy_rows(self):
+        self.require_module()
+        model = self.model()
+        with model.get_connection() as conn:
+            populate_legacy(conn)
+        model.close_all()
+        self.drop_group(self.path)
+        with closing(sqlite3.connect(self.path)) as raw:
+            self.assertIsNone(annotation_schema_error(raw))
+            self.assertEqual(
+                annotation_schema_error(raw, allow_missing=False),
+                "Missing image annotation schema",
+            )
+            legacy_rows_before = table_rows(raw, LEGACY_TABLES)
+        self.assertTrue(PromptModel.verify_database_file(self.path)[0])
+        self.assertFalse(
+            PromptModel.verify_database_file(self.path, require_library_state=True)[0]
+        )
+        PromptModel.reset_schema_cache()
+
+        model = self.model()
+
+        with model.get_connection() as conn:
+            self.assertEqual(catalog(conn), expected_catalog())
+            self.assertIsNone(annotation_schema_error(conn, allow_missing=False))
+            self.assertEqual(metadata(conn), [(1, 0)])
+            legacy_rows_after = table_rows(conn, LEGACY_TABLES)
+            # No backfill: existing prompt tags never become image tags.
+            counts = [
+                conn.execute(f"SELECT COUNT(*) FROM {table}").fetchone()[0]
+                for table in (FILES, BYTES, IMAGE_TAGS)
+            ]
+        self.assertEqual(counts, [0, 0, 0])
+        self.assertEqual(legacy_rows_after, legacy_rows_before)
+        self.assertEqual(
+            PromptModel.verify_database_file(self.path, require_library_state=True),
+            (True, "ok"),
+        )
+
+    def tampered(self, name, statements, *, checks_off=False):
+        """A closed current database whose group was altered after the fact."""
+        path = str(Path(self.temp.name) / f"{name.replace(' ', '-')}.db")
+        model = self.model(path)
+        with model.get_connection() as conn:
+            populate_legacy(conn)
+            insert_group_rows(conn)
+        model.close_all()
+        with closing(sqlite3.connect(path)) as raw, raw:
+            raw.execute("PRAGMA foreign_keys=OFF")
+            if checks_off:
+                raw.execute("PRAGMA ignore_check_constraints=ON")
+            for statement in statements:
+                if isinstance(statement, tuple):
+                    raw.execute(*statement)
+                else:
+                    raw.execute(statement)
+        return path
+
+    def assert_rejected_unrepaired(self, path):
+        with closing(sqlite3.connect(path)) as raw:
+            schema_before = raw.execute(
+                "SELECT type,name,tbl_name,sql FROM sqlite_master ORDER BY name"
+            ).fetchall()
+            rows_before = table_rows(raw, TABLES)
+            for allow_missing in (False, True):
+                error = annotation_schema_error(raw, allow_missing=allow_missing)
+                self.assertIsNotNone(error)
+                self.assertIn("image annotation", error.lower())
+        self.assertFalse(PromptModel.verify_database_file(path)[0])
+        self.assertFalse(
+            PromptModel.verify_database_file(path, require_library_state=True)[0]
+        )
+        PromptModel.reset_schema_cache()
+        with self.assertRaises(sqlite3.DatabaseError):
+            self.model(path)
+        PromptModel.reset_schema_cache()
+        with closing(sqlite3.connect(path)) as raw:
+            self.assertEqual(
+                raw.execute(
+                    "SELECT type,name,tbl_name,sql FROM sqlite_master ORDER BY name"
+                ).fetchall(),
+                schema_before,
+            )
+            self.assertEqual(table_rows(raw, TABLES), rows_before)
+
+    def test_partial_or_malformed_group_is_rejected_without_repair(self):
+        self.require_module()
+        cases = {
+            "missing metadata": (f"DROP TABLE {METADATA}",),
+            "missing files": (f"DROP TABLE {FILES}",),
+            "missing bytes": (f"DROP TABLE {BYTES}",),
+            "missing image tags": (f"DROP TABLE {IMAGE_TAGS}",),
+            "reserved view": (
+                f"DROP TABLE {FILES}",
+                f"CREATE VIEW {FILES} AS SELECT 'x' AS id",
+            ),
+            "extra trigger": (
+                f"CREATE TRIGGER media_annotation_trigger AFTER INSERT ON {FILES} "
+                "BEGIN SELECT 1; END",
+            ),
+            "extra index": (
+                f"CREATE INDEX idx_media_annotation_extra ON {FILES}(created_at)",
+            ),
+            "extra column": (
+                f"DROP TABLE {IMAGE_TAGS}",
+                DDL[IMAGE_TAGS].replace(
+                    "position INTEGER", "note TEXT, position INTEGER"
+                ),
+            ),
+            "wrong primary key": (
+                f"DROP TABLE {FILES}",
+                DDL[FILES].replace("id TEXT PRIMARY KEY", "id TEXT"),
+            ),
+            "wrong foreign key": (
+                f"DROP TABLE {BYTES}",
+                DDL[BYTES].replace("ON DELETE CASCADE", "ON DELETE RESTRICT"),
+                INDEXES["idx_media_annotation_bytes_fingerprint"],
+            ),
+            "wrong check": (
+                f"DROP TABLE {METADATA}",
+                DDL[METADATA].replace("schema_version=1", "schema_version=2"),
+                f"INSERT INTO {METADATA} VALUES(1,2,0)",
+            ),
+            "missing unique": (
+                f"DROP TABLE {FILES}",
+                DDL[FILES].replace("NOT NULL UNIQUE", "NOT NULL"),
+            ),
+            "missing index": ("DROP INDEX idx_media_image_tags_tag",),
+            "non-unique position index": (
+                "DROP INDEX idx_media_image_tags_position",
+                INDEXES["idx_media_image_tags_position"].replace("UNIQUE ", ""),
+            ),
+            "wrong index columns": (
+                "DROP INDEX idx_media_annotation_bytes_fingerprint",
+                INDEXES["idx_media_annotation_bytes_fingerprint"].replace(
+                    ", mtime_ns", ""
+                ),
+            ),
+            "orphan tag": (f"INSERT INTO {IMAGE_TAGS} VALUES('{BYTE_ID}',404,1)",),
+            "orphan byte": (
+                (
+                    f"INSERT INTO {BYTES} VALUES(?,?,?,1,1,?,?,0)",
+                    (
+                        OTHER_ID,
+                        "9" * 8 + "-0000-4000-8000-" + "0" * 12,
+                        DIGEST,
+                        NOW,
+                        NOW,
+                    ),
+                ),
+            ),
+            "bad file uuid": (
+                (f"UPDATE {FILES} SET id=?", ("x" * 36,)),
+                (f"UPDATE {BYTES} SET file_id=?", ("x" * 36,)),
+            ),
+            "uppercase byte uuid": (
+                f"UPDATE {BYTES} SET id=upper(id)",
+                f"UPDATE {IMAGE_TAGS} SET byte_id=upper(byte_id)",
+            ),
+            "null file id": ((f"INSERT INTO {FILES} VALUES(NULL,?,?)", ("/n", NOW)),),
+            "blob ids": (
+                f"UPDATE {FILES} SET id=CAST(id AS BLOB)",
+                f"UPDATE {BYTES} SET file_id=CAST(file_id AS BLOB)",
+            ),
+            "no metadata row": (f"DELETE FROM {METADATA}",),
+            "long created_at": ((f"UPDATE {FILES} SET created_at=?", ("x" * 65,)),),
+            "blob verified_at": (f"UPDATE {BYTES} SET verified_at=X'00'",),
+            "text mtime": (f"UPDATE {BYTES} SET mtime_ns='soon'",),
+            "duplicate position": (
+                "DROP INDEX idx_media_image_tags_position",
+                "INSERT INTO tags(id,name) VALUES(77,'dup')",
+                f"INSERT INTO {IMAGE_TAGS} VALUES('{BYTE_ID}',77,0)",
+            ),
+            "long locator": (
+                (f"UPDATE {FILES} SET locator=?", ("x" * (MAX_LOCATOR + 1),)),
+            ),
+            "multibyte long locator": (
+                (f"UPDATE {FILES} SET locator=?", ("é" * (MAX_LOCATOR + 1),)),
+            ),
+            "nul locator bypass": (
+                (f"UPDATE {FILES} SET locator=?", ("x\0" + "y" * MAX_LOCATOR,)),
+            ),
+        }
+        for label, statements in cases.items():
+            with self.subTest(label):
+                self.assert_rejected_unrepaired(self.tampered(label, statements))
+        unchecked = {
+            "two metadata rows": (f"INSERT INTO {METADATA} VALUES(2,1,0)",),
+            "wrong schema version": (f"UPDATE {METADATA} SET schema_version=2",),
+            "negative index revision": (
+                f"UPDATE {METADATA} SET image_index_revision=-1",
+            ),
+            "bad digest": (f"UPDATE {BYTES} SET sha256=upper(sha256)",),
+            "short digest": (f"UPDATE {BYTES} SET sha256=substr(sha256,1,63)",),
+            "negative size": (f"UPDATE {BYTES} SET size=-1",),
+            "negative tag revision": (f"UPDATE {BYTES} SET tag_revision=-1",),
+            "negative position": (f"UPDATE {IMAGE_TAGS} SET position=-1",),
+        }
+        for label, statements in unchecked.items():
+            with self.subTest(label):
+                path = self.tampered(label, statements, checks_off=True)
+                self.assert_rejected_unrepaired(path)
+        accepted = {
+            "locator at bound": ("x" * MAX_LOCATOR,),
+            "multibyte locator at bound": ("é" * MAX_LOCATOR,),
+        }
+        for label, (locator,) in accepted.items():
+            with self.subTest(label):
+                path = self.tampered(
+                    label, ((f"UPDATE {FILES} SET locator=?", (locator,)),)
+                )
+                with closing(sqlite3.connect(path)) as raw:
+                    self.assertIsNone(annotation_schema_error(raw, allow_missing=False))
+                self.assertEqual(
+                    PromptModel.verify_database_file(path, require_library_state=True),
+                    (True, "ok"),
+                )
+
+    def test_install_failure_rolls_back_and_retries(self):
+        self.require_module()
+        with patch(
+            "database.annotation_schema.annotation_schema_error",
+            side_effect=[None, "injected final validation failure"],
+        ):
+            with self.assertRaisesRegex(sqlite3.DatabaseError, "injected"):
+                self.model()
+        with closing(sqlite3.connect(self.path)) as raw:
+            self.assertEqual(catalog(raw), set())
+            self.assertIsNotNone(
+                raw.execute(
+                    "SELECT 1 FROM sqlite_master WHERE name='saved_views'"
+                ).fetchone()
+            )
+        PromptModel.reset_schema_cache()
+        model = self.model()
+        self.assertEqual(
+            PromptModel.verify_database_file(model.db_path, require_library_state=True),
+            (True, "ok"),
+        )
+        with model.get_connection() as conn:
+            self.assertEqual(metadata(conn), [(1, 0)])
+
+    def test_wal_backup_restore_round_trips_group(self):
+        self.require_module()
+        source_path = str(Path(self.temp.name) / "source.db")
+        source = self.model(source_path)
+        with source.get_connection() as conn:
+            conn.execute("PRAGMA wal_autocheckpoint=0")
+            conn.execute("INSERT INTO prompts(id,text,hash) VALUES(7,'p','p')")
+            insert_group_rows(conn)
+            conn.commit()
+            source_files = rows(conn, f"SELECT * FROM {FILES}")
+            source_bytes = rows(conn, f"SELECT * FROM {BYTES}")
+        self.assertGreater(Path(source_path + "-wal").stat().st_size, 0)
+        live = self.model()
+        with live.get_connection() as conn:
+            conn.execute("INSERT INTO prompts(id,text,hash) VALUES(1,'live','live')")
+            insert_group_rows(conn, tag_name="live-only", locator="/out/live.png")
+        source_before = Path(source_path).read_bytes()
+        self.assertEqual(
+            PromptModel.verify_database_file(source_path, require_library_state=True),
+            (True, "ok"),
+        )
+        # The rows still sit in the WAL; a backup file must carry them too.
+        backup_file = str(Path(self.temp.name) / "backup.db")
+        self.assertTrue(source.backup_database(backup_file))
+        with closing(sqlite3.connect(backup_file)) as raw:
+            self.assertEqual(image_tags(raw), [("img-only", 0)])
+            self.assertIsNone(annotation_schema_error(raw, allow_missing=False))
+
+        safety = live.restore_from_file(source_path)
+
+        self.assertTrue(Path(safety).is_file())
+        self.assertEqual(Path(source_path).read_bytes(), source_before)
+        with live.get_connection() as conn:
+            restored_image_tags = image_tags(conn)
+            self.assertEqual(rows(conn, f"SELECT * FROM {FILES}"), source_files)
+            self.assertEqual(rows(conn, f"SELECT * FROM {BYTES}"), source_bytes)
+            self.assertEqual(metadata(conn), [(1, 0)])
+            self.assertIsNone(annotation_schema_error(conn, allow_missing=False))
+            self.assertEqual(rows(conn, "SELECT id FROM prompts"), [(7,)])
+        self.assertEqual(restored_image_tags, [("img-only", 0)])
+        with closing(sqlite3.connect(safety)) as raw:
+            self.assertEqual(image_tags(raw), [("live-only", 0)])
+            self.assertEqual(raw.execute("SELECT id FROM prompts").fetchall(), [(1,)])
+        # The same rows restore from the backup file as well.
+        PromptModel.reset_schema_cache()
+        live.restore_from_file(backup_file)
+        with live.get_connection() as conn:
+            self.assertEqual(image_tags(conn), [("img-only", 0)])
+        PromptModel.close_all_instances()
+        assert_closed(self, self.temp.name)
+
+    def test_staged_failure_preserves_source_live_safety_and_handles(self):
+        self.require_module()
+        live = self.model()
+        with live.get_connection() as conn:
+            conn.execute("INSERT INTO prompts(id,text,hash) VALUES(1,'live','live')")
+            insert_group_rows(conn, tag_name="live-only")
+        # A source with a malformed group is refused before any change.
+        bad = self.tampered(
+            "restore bad",
+            (f"UPDATE {BYTES} SET sha256=upper(sha256)",),
+            checks_off=True,
+        )
+        # A staging failure while installing the group must be refused too.
+        old = str(Path(self.temp.name) / "old.db")
+        self.model(old).close_all()
+        self.drop_group(old)
+        for source, patched in ((bad, False), (old, True)):
+            with self.subTest(source=Path(source).name):
+                before = Path(source).read_bytes()
+                safety = sorted(Path(self.temp.name).glob("live.db.backup_*"))
+                if patched:
+                    context = patch(
+                        "database.models.migrate_annotation_state",
+                        side_effect=sqlite3.OperationalError("staging failed"),
+                    )
+                    expected = sqlite3.OperationalError
+                else:
+                    context = patch.object(PromptModel, "backup_database")
+                    expected = ValueError
+                with context as hook, self.assertRaises(expected):
+                    live.restore_from_file(source)
+                if not patched:
+                    hook.assert_not_called()
+                self.assertEqual(Path(source).read_bytes(), before)
+                self.assertEqual(
+                    sorted(Path(self.temp.name).glob("live.db.backup_*")), safety
+                )
+                assert_closed(self, source)
+                with live.get_connection() as conn:
+                    self.assertEqual(image_tags(conn), [("live-only", 0)])
+                    self.assertEqual(metadata(conn), [(1, 0)])
+        self.assertFalse(
+            PromptModel.verify_database_file(old, require_library_state=True)[0]
+        )
+        self.assertTrue(PromptModel.verify_database_file(old)[0])
+
+    def test_verify_requires_group_only_with_require_library_state(self):
+        self.require_module()
+        self.model().close_all()
+        self.assertEqual(
+            PromptModel.verify_database_file(self.path, require_library_state=True),
+            (True, "ok"),
+        )
+        self.drop_group(self.path)
+        self.assertEqual(PromptModel.verify_database_file(self.path), (True, "ok"))
+        self.assertEqual(
+            PromptModel.verify_database_file(self.path, require_library_state=True),
+            (False, "Missing image annotation schema"),
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_annotations.py

@@ -0,0 +1,598 @@
+"""Per-image tag sets (#249 A2): byte-exact storage in the shared vocabulary."""
+
+import hashlib
+import os
+import re
+import sqlite3
+import tempfile
+import unittest
+from unittest.mock import patch
+
+try:
+    from tests.open_handles import assert_closed
+except ImportError:  # discovered with tests/ as the top-level directory
+    from open_handles import assert_closed
+
+from database.annotation_schema import annotation_schema_error
+from database.library_commands import LibraryCommandError
+from database.operations import PromptDatabase
+from utils.hashing import generate_prompt_hash
+
+try:  # guarded so a missing module fails assertions, not test loading
+    from database import annotations as annotations_module
+    from database.annotations import (
+        MAX_IMAGE_TAGS,
+        AnnotationCommandError,
+        normalize_image_tags,
+    )
+except ImportError:
+    annotations_module = None
+    MAX_IMAGE_TAGS = None
+    AnnotationCommandError = None
+    normalize_image_tags = None
+
+UUID = re.compile(r"[0-9a-f]{8}(?:-[0-9a-f]{4}){3}-[0-9a-f]{12}")
+DIGEST_A = hashlib.sha256(b"bytes a").hexdigest()
+DIGEST_B = hashlib.sha256(b"bytes b").hexdigest()
+SIZE = 7
+MTIME = 1_700_000_000_123_456_789
+ANNOTATION_TABLES = (
+    "media_annotation_files",
+    "media_annotation_bytes",
+    "media_image_tags",
+    "media_annotation_metadata",
+)
+LEGACY_TABLES = (
+    "tags",
+    "prompt_tags",
+    "prompt_pair_tags",
+    "generated_images",
+    "image_prompt_pairs",
+)
+
+
+class AnnotationTestCase(unittest.TestCase):
+    def setUp(self):
+        self.assertIsNotNone(
+            annotations_module, "database.annotations is not implemented"
+        )
+        directory = tempfile.TemporaryDirectory()
+        self.addCleanup(directory.cleanup)
+        self.addCleanup(assert_closed, self, directory.name)
+        self.root = directory.name
+        self.db = PromptDatabase(os.path.join(self.root, "annotations.db"))
+        self.addCleanup(self.db.close_all)
+        self.epoch = self.db.get_execution_epoch()
+
+    def locator(self, name: str) -> str:
+        return os.path.normcase(os.path.join(os.path.realpath(self.root), name))
+
+    def replace(self, loc, digest=DIGEST_A, tags=(), *, revision=0, **over):
+        options = {
+            "locator": loc,
+            "sha256": digest,
+            "size": SIZE,
+            "mtime_ns": MTIME,
+            "expected_epoch": self.epoch,
+            "expected_tag_revision": revision,
+            "tags": list(tags) if isinstance(tags, tuple) else tags,
+        }
+        options.update(over)
+        return self.db.replace_image_tags(**options)
+
+    def add(self, loc, digest=DIGEST_A, tags=(), **over):
+        options = {
+            "locator": loc,
+            "sha256": digest,
+            "size": SIZE,
+            "mtime_ns": MTIME,
+            "expected_epoch": self.epoch,
+            "tags": list(tags) if isinstance(tags, tuple) else tags,
+        }
+        options.update(over)
+        return self.db.add_image_tags(**options)
+
+    def read(self, loc, digest=DIGEST_A):
+        return self.db.read_image_tags(loc, digest)
+
+    def rows(self, *tables):
+        """Every row of each table; ``verified_at`` is a fingerprint, not content."""
+        conn = self.db.model.get_connection()
+        result = []
+        for table in tables:
+            columns = [
+                row[1]
+                for row in conn.execute(f"PRAGMA table_info({table})")
+                if row[1] != "verified_at"
+            ]
+            result.append(
+                [
+                    tuple(row)
+                    for row in conn.execute(
+                        f"SELECT {','.join(columns)} FROM {table} ORDER BY 1,2"
+                    )
+                ]
+            )
+        return result
+
+    def counts(self):
+        conn = self.db.model.get_connection()
+        return {
+            table: conn.execute(f"SELECT COUNT(*) FROM {table}").fetchone()[0]
+            for table in ANNOTATION_TABLES
+        }
+
+    def index_revision(self):
+        return (
+            self.db.model.get_connection()
+            .execute("SELECT image_index_revision FROM media_annotation_metadata")
+            .fetchone()[0]
+        )
+
+    def schema_clean(self):
+        self.assertIsNone(
+            annotation_schema_error(self.db.model.get_connection(), allow_missing=False)
+        )
+
+
+class NormalizeImageTagsTests(unittest.TestCase):
+    def setUp(self):
+        self.assertIsNotNone(
+            normalize_image_tags, "database.annotations is not implemented"
+        )
+
+    def test_case_insensitive_dedupe_keeps_first_spelling_and_position(self):
+        self.assertEqual(
+            normalize_image_tags([" Forest ", "mist", "forest", "MIST", "Rain"]),
+            ["Forest", "mist", "Rain"],
+        )
+        self.assertEqual(normalize_image_tags([]), [])
+        self.assertEqual(MAX_IMAGE_TAGS, 100)
+
+    def test_limit_and_invalid_names_rejected(self):
+        for bad in (
+            None,
+            "forest",
+            {"a": 1},
+            [1],
+            [None],
+            ["a,b"],
+            [""],
+            ["   "],
+            ["x" * 101],
+            [f"t{i}" for i in range(MAX_IMAGE_TAGS + 1)],
+        ):
+            with self.subTest(bad=bad), self.assertRaises(ValueError):
+                normalize_image_tags(bad)
+        # Duplicates fold before the cap is applied.
+        many = [f"t{i}" for i in range(MAX_IMAGE_TAGS)]
+        self.assertEqual(normalize_image_tags(many + ["T0", "T99"]), many)
+
+
+class ReplaceImageTagsTests(AnnotationTestCase):
+    def test_first_save_creates_owner_and_byte_atomically_including_empty_set(self):
+        loc = self.locator("a.png")
+        receipt = self.replace(loc, tags=["Forest", "mist", "forest"])
+        self.assertEqual(receipt["tags"], ["Forest", "mist"])
+        self.assertEqual(receipt["tag_revision"], 1)
+        self.assertEqual(
+            receipt,
+            {
+                "library_epoch": self.epoch,
+                "sha256": DIGEST_A,
+                "tags": ["Forest", "mist"],
+                "tag_revision": 1,
+                "replayed": False,
+            },
+        )
+        conn = self.db.model.get_connection()
+        files = conn.execute("SELECT * FROM media_annotation_files").fetchall()
+        self.assertEqual(len(files), 1)
+        self.assertRegex(files[0]["id"], UUID)
+        self.assertEqual(files[0]["locator"], loc)
+        byte = conn.execute("SELECT * FROM media_annotation_bytes").fetchone()
+        self.assertRegex(byte["id"], UUID)
+        self.assertEqual(
+            (byte["file_id"], byte["sha256"], byte["size"], byte["mtime_ns"]),
+            (files[0]["id"], DIGEST_A, SIZE, MTIME),
+        )
+        self.assertEqual(byte["tag_revision"], 1)
+        self.assertEqual(byte["verified_at"], byte["first_seen_at"])
+        self.assertEqual(
+            [
+                tuple(row)
+                for row in conn.execute(
+                    "SELECT t.name, mit.position FROM media_image_tags mit "
+                    "JOIN tags t ON t.id = mit.tag_id ORDER BY mit.position"
+                )
+            ],
+            [("Forest", 0), ("mist", 1)],
+        )
+        self.assertEqual(self.index_revision(), 1)
+        # An explicit empty set is a real save: owner and byte record exist.
+        empty = self.replace(self.locator("b.png"), DIGEST_B, [])
+        self.assertEqual((empty["tags"], empty["tag_revision"]), ([], 1))
+        self.assertEqual(self.counts()["media_annotation_files"], 2)
+        self.assertEqual(self.counts()["media_annotation_bytes"], 2)
+        self.assertEqual(
+            self.db.read_image_tags(self.locator("b.png"), DIGEST_B),
+            {"tags": [], "tag_revision": 1, "prior_count": 0},
+        )
+        self.assertEqual(self.index_revision(), 2)
+        self.schema_clean()
+
+    def test_save_changes_only_target_byte_set(self):
+        a, b, c = self.locator("a.png"), self.locator("b.png"), self.locator("c.png")
+        prompt = self.db.save_prompt(
+            "Shared forest",
+            tags=["shared"],
+            prompt_hash=generate_prompt_hash("Shared forest"),
+        )
+        pair = self.db.save_prompt_pair("Shared forest", "blurry", tags=["pairtag"])
+        image_a = self.db.link_image_to_prompt(prompt, a)
+        image_b = self.db.link_image_to_prompt(prompt, b)
+        self.db.link_image_to_pair(pair, image_a, {"seed": 1})
+        self.db.link_image_to_pair(pair, image_b, {"seed": 2})
+        self.replace(b, DIGEST_A, ["sibling"])
+        self.replace(c, DIGEST_A, ["copy"])  # equal bytes at another path
+        before = self.rows(*LEGACY_TABLES)
+        sibling_before = self.db.read_image_tags(b, DIGEST_A)
+        copy_before = self.db.read_image_tags(c, DIGEST_A)
+
+        receipt = self.replace(a, DIGEST_A, ["only a"])
+
+        self.assertEqual((receipt["tags"], receipt["tag_revision"]), (["only a"], 1))
+        after = self.rows(*LEGACY_TABLES)
+        # The registry gained the new name; nothing else moved.
+        self.assertEqual(after[1:], before[1:])
+        self.assertEqual([row for row in after[0] if row[1] != "only a"], before[0])
+        self.assertEqual(self.db.read_image_tags(b, DIGEST_A), sibling_before)
+        self.assertEqual(self.db.read_image_tags(c, DIGEST_A), copy_before)
+        self.assertEqual(self.db.get_prompt_by_id(prompt)["tags"], ["shared"])
+        self.assertEqual(self.db.get_prompt_pair(pair)["tags"], ["pairtag"])
+
+    def test_changed_bytes_at_same_locator_start_empty_and_keep_prior(self):
+        loc = self.locator("a.png")
+        self.replace(loc, DIGEST_A, ["old"])
+        self.assertEqual(
+            self.db.read_image_tags(loc, DIGEST_B),
+            {"tags": [], "tag_revision": 0, "prior_count": 1},
+        )
+        receipt = self.replace(loc, DIGEST_B, ["new"], mtime_ns=MTIME + 5)
+        self.assertEqual((receipt["tags"], receipt["tag_revision"]), (["new"], 1))
+        self.assertEqual(
+            self.db.read_image_tags(loc, DIGEST_A),
+            {"tags": ["old"], "tag_revision": 1, "prior_count": 1},
+        )
+        self.assertEqual(
+            self.db.read_image_tags(loc, DIGEST_B),
+            {"tags": ["new"], "tag_revision": 1, "prior_count": 1},
+        )
+        self.assertEqual(self.counts()["media_annotation_files"], 1)
+        self.assertEqual(self.counts()["media_annotation_bytes"], 2)
+        self.assertEqual(
+            self.db.read_image_tags(self.locator("unseen.png"), DIGEST_A),
+            {"tags": [], "tag_revision": 0, "prior_count": 0},
+        )
+
+    def test_revision_conflict_returns_current_set(self):
+        loc = self.locator("a.png")
+        self.replace(loc, tags=["a"])
+        before = self.rows(*ANNOTATION_TABLES, "tags")
+        with self.assertRaises(AnnotationCommandError) as ctx:
+            self.replace(loc, tags=["b"], revision=0)
+        self.assertEqual(
+            (ctx.exception.code, ctx.exception.current["tag_revision"]),
+            ("tag-conflict", 1),
+        )
+        self.assertIsInstance(ctx.exception, LibraryCommandError)
+        self.assertEqual(ctx.exception.epoch, self.epoch)
+        self.assertEqual(ctx.exception.current["tags"], ["a"])
+        self.assertEqual(ctx.exception.current["sha256"], DIGEST_A)
+        with self.assertRaises(AnnotationCommandError) as ahead:
+            self.replace(loc, tags=["a"], revision=5)
+        self.assertEqual(ahead.exception.code, "tag-conflict")
+        self.assertEqual(self.rows(*ANNOTATION_TABLES, "tags"), before)
+        self.assertFalse(self.db.model.get_connection().in_transaction)
+
+    def test_identical_replay_after_commit_is_success_replayed(self):
+        loc = self.locator("a.png")
+        first = self.replace(loc, tags=["a", "b"])
+        self.assertEqual((first["tag_revision"], first["replayed"]), (1, False))
+        before = self.rows(*ANNOTATION_TABLES)
+        # The lost-response retry: same set, the revision it was sent against.
+        retry = self.replace(loc, tags=["a", "b"], revision=0)
+        self.assertEqual(retry, {**first, "replayed": True})
+        # Saving the current set at the current revision writes nothing either.
+        again = self.replace(loc, tags=["a", "b"], revision=1)
+        self.assertEqual(again, {**first, "replayed": True})
+        self.assertEqual(self.rows(*ANNOTATION_TABLES), before)
+        # Order is part of the set.
+        reordered = self.replace(loc, tags=["b", "a"], revision=1)
+        self.assertEqual(
+            (reordered["tags"], reordered["tag_revision"], reordered["replayed"]),
+            (["b", "a"], 2, False),
+        )
+        self.assertEqual(self.index_revision(), 2)
+
+    def test_replay_with_new_fingerprint_moves_only_the_index_revision(self):
+        loc = self.locator("a.png")
+        first = self.replace(loc, tags=["a"])
+        index = self.index_revision()
+        same = self.replace(loc, tags=["a"], revision=1)
+        self.assertEqual(same, {**first, "replayed": True})
+        self.assertEqual(self.index_revision(), index)  # identical: nothing moved
+        touched = self.replace(loc, tags=["a"], revision=1, mtime_ns=MTIME + 5)
+        self.assertEqual(touched, {**first, "replayed": True})
+        self.assertEqual(self.index_revision(), index + 1)  # currency changed
+        self.assertEqual(
+            self.db.current_image_tags({loc: (SIZE, MTIME + 5)}), {loc: ["a"]}
+        )
+        self.assertEqual(self.db.current_image_tags({loc: (SIZE, MTIME)}), {})
+        self.assertEqual(self.read(loc)["tag_revision"], 1)
+
+    def test_epoch_checked_inside_transaction(self):
+        loc = self.locator("a.png")
+        seen = []
+        original = PromptDatabase._library_epoch
+
+        def spy(db, conn, expected_epoch):
+            seen.append(conn.in_transaction)
+            return original(db, conn, expected_epoch)
+
+        with patch.object(PromptDatabase, "_library_epoch", spy):
+            self.replace(loc, tags=["a"])
+            self.add(loc, tags=["b"])
+        self.assertEqual(seen, [True, True])
+        stale = "11111111-2222-4333-8444-555555555555"
+        for call in (
+            lambda: self.replace(
+                self.locator("x.png"), tags=["a"], expected_epoch=stale
+            ),
+            lambda: self.add(self.locator("x.png"), tags=["a"], expected_epoch=stale),
+        ):
+            with self.assertRaises(LibraryCommandError) as ctx:
+                call()
+            self.assertEqual(ctx.exception.code, "storage-changed")
+            self.assertEqual(ctx.exception.epoch, self.epoch)
+        self.assertEqual(self.counts()["media_annotation_files"], 1)
+        for bad_epoch in (None, 7, "not-a-uuid"):
+            with self.assertRaises(ValueError):
+                self.replace(loc, tags=["a"], expected_epoch=bad_epoch)
+
+    def test_limit_and_invalid_names_rejected_without_partial_write(self):
+        loc = self.locator("a.png")
+        before = self.rows(*ANNOTATION_TABLES, "tags")
+        for tags in (
+            ["ok", "a,b"],
+            ["ok", ""],
+            ["ok", 3],
+            "ok",
+            [f"t{i}" for i in range(101)],
+        ):
+            with self.subTest(tags=tags), self.assertRaises(ValueError):
+                self.replace(loc, tags=tags)
+            with self.subTest(tags=tags, command="add"), self.assertRaises(ValueError):
+                self.add(loc, tags=tags)
+        for field, value in (
+            ("locator", ""),
+            ("locator", None),
+            ("locator", "x" * 4097),
+            ("sha256", DIGEST_A.upper()),
+            ("sha256", DIGEST_A[:-1]),
+            ("sha256", None),
+            ("size", -1),
+            ("size", True),
+            ("size", "7"),
+            ("mtime_ns", 1.5),
+            ("mtime_ns", None),
+            ("expected_tag_revision", -1),
+            ("expected_tag_revision", False),
+        ):
+            with self.subTest(field=field, value=value), self.assertRaises(ValueError):
+                self.replace(loc, tags=["ok"], **{field: value})
+        self.assertEqual(self.rows(*ANNOTATION_TABLES, "tags"), before)
+        self.assertEqual(self.index_revision(), 0)
+
+    def test_writers_supply_uuid_ids_and_refuse_to_write_without_one(self):
+        loc = self.locator("a.png")
+        before = self.rows(*ANNOTATION_TABLES)
+        with patch.object(annotations_module, "uuid4", return_value=""):
+            with self.assertRaises(ValueError):
+                self.replace(loc, tags=["a"])
+        self.assertEqual(self.rows(*ANNOTATION_TABLES), before)
+        self.replace(loc, tags=["a"])
+        self.add(self.locator("b.png"), DIGEST_B, ["b"])
+        conn = self.db.model.get_connection()
+        for table in ("media_annotation_files", "media_annotation_bytes"):
+            for (value,) in conn.execute(f"SELECT id FROM {table}"):
+                self.assertRegex(value, UUID)
+        self.schema_clean()
+
+    def test_never_resolves_owner_through_prompt_lookup(self):
+        loc = self.locator("a.png")
+        prompt = self.db.save_prompt("Linked", tags=["prompt-tag"])
+        self.db.link_image_to_prompt(prompt, loc)
+        with patch.object(
+            PromptDatabase,
+            "get_prompt_id_for_image",
+            side_effect=AssertionError("prompt lookup is not an owner"),
+        ):
+            self.replace(loc, tags=["img"])
+            self.add(loc, tags=["more"])
+            self.assertEqual(
+                self.db.read_image_tags(loc, DIGEST_A)["tags"], ["img", "more"]
+            )
+            self.assertEqual(
+                self.db.current_image_tags({loc: (SIZE, MTIME)}),
+                {loc: ["img", "more"]},
+            )
+            self.assertTrue(
+                self.db.refresh_image_fingerprint(loc, DIGEST_A, SIZE, MTIME + 1)
+            )
+            list(self.db.iter_image_tag_records())
+        self.assertEqual(self.db.get_prompt_by_id(prompt)["tags"], ["prompt-tag"])
+
+
+class AddImageTagsTests(AnnotationTestCase):
+    def test_add_merges_without_removing_and_caps_at_100(self):
+        loc = self.locator("a.png")
+        self.replace(loc, tags=["Forest", "mist"])
+        receipt = self.add(loc, tags=["forest", "Rain", "rain"])
+        self.assertEqual(
+            receipt,
+            {
+                "library_epoch": self.epoch,
+                "sha256": DIGEST_A,
+                "tags": ["Forest", "mist", "Rain"],
+                "tag_revision": 2,
+                "replayed": False,
+                "added": 1,
+            },
+        )
+        nothing = self.add(loc, tags=["MIST"])
+        self.assertEqual(
+            (nothing["tags"], nothing["tag_revision"], nothing["replayed"]),
+            (["Forest", "mist", "Rain"], 2, True),
+        )
+        self.assertEqual(nothing["added"], 0)
+        before = self.rows(*ANNOTATION_TABLES, "tags")
+        with self.assertRaises(AnnotationCommandError) as ctx:
+            self.add(loc, tags=[f"t{i}" for i in range(98)])
+        self.assertEqual(ctx.exception.code, "tag-limit")
+        self.assertEqual(ctx.exception.current["tag_revision"], 2)
+        self.assertEqual(self.rows(*ANNOTATION_TABLES, "tags"), before)
+        exact = self.add(loc, tags=[f"t{i}" for i in range(97)])
+        self.assertEqual((len(exact["tags"]), exact["added"]), (100, 97))
+        # An unseen file: add creates the owner and byte record like a save.
+        fresh = self.add(self.locator("b.png"), DIGEST_B, ["one"])
+        self.assertEqual(
+            (fresh["tags"], fresh["tag_revision"], fresh["added"]), (["one"], 1, 1)
+        )
+        self.schema_clean()
+
+
+class CurrentImageTagsTests(AnnotationTestCase):
+    def test_current_tags_require_matching_fingerprint(self):
+        loc = self.locator("a.png")
+        other = self.locator("b.png")
+        self.replace(loc, tags=["a"])
+        self.replace(other, DIGEST_B, [], mtime_ns=MTIME + 9)
+        self.assertEqual(
+            self.db.current_image_tags({loc: (SIZE, MTIME), other: (SIZE, MTIME + 9)}),
+            {loc: ["a"], other: []},
+        )
+        self.assertEqual(self.db.current_image_tags({loc: (SIZE, MTIME + 1)}), {})
+        self.assertEqual(self.db.current_image_tags({loc: (SIZE + 1, MTIME)}), {})
+        self.assertEqual(self.db.current_image_tags({}), {})
+        self.assertEqual(
+            self.db.current_image_tags({self.locator("none.png"): (SIZE, MTIME)}), {}
+        )
+        # Refreshing the verified fingerprint makes the record current again.
+        self.assertTrue(
+            self.db.refresh_image_fingerprint(loc, DIGEST_A, SIZE, MTIME + 1)
+        )
+        self.assertFalse(self.db.refresh_image_fingerprint(loc, DIGEST_B, SIZE, MTIME))
+        self.assertFalse(
+            self.db.refresh_image_fingerprint(
+                self.locator("x.png"), DIGEST_A, SIZE, MTIME
+            )
+        )
+        self.assertEqual(
+            self.db.current_image_tags({loc: (SIZE, MTIME + 1)}), {loc: ["a"]}
+        )
+        self.assertEqual(self.db.current_image_tags({loc: (SIZE, MTIME)}), {})
+        self.assertEqual(
+            self.db.read_image_tags(loc, DIGEST_A),
+            {"tags": ["a"], "tag_revision": 1, "prior_count": 0},
+        )
+        for bad in (None, [], {loc: (SIZE,)}, {loc: (SIZE, 1.5)}, {3: (SIZE, MTIME)}):
+            with self.subTest(bad=bad), self.assertRaises(ValueError):
+                self.db.current_image_tags(bad)
+
+    def test_newest_verification_wins_on_equal_fingerprints(self):
+        loc = self.locator("a.png")
+        stamps = iter(f"2026-10-22T00:00:0{i}.000000+00:00" for i in range(10))
+        with patch.object(annotations_module, "utc_now_iso", side_effect=stamps):
+            self.replace(loc, DIGEST_A, ["old bytes"])
+            self.replace(loc, DIGEST_B, ["new bytes"])  # same size and mtime
+        self.assertEqual(
+            self.db.current_image_tags({loc: (SIZE, MTIME)}), {loc: ["new bytes"]}
+        )
+        with patch.object(
+            annotations_module,
+            "utc_now_iso",
+            return_value="2026-10-22T00:01:00.000000+00:00",
+        ):
+            self.assertTrue(
+                self.db.refresh_image_fingerprint(loc, DIGEST_A, SIZE, MTIME)
+            )
+        self.assertEqual(
+            self.db.current_image_tags({loc: (SIZE, MTIME)}), {loc: ["old bytes"]}
+        )
+
+    def test_lookup_is_chunked_and_records_stream_in_position_order(self):
+        locators = [self.locator(f"{i:04}.png") for i in range(503)]
+        self.replace(locators[0], tags=["z", "y"])
+        self.replace(locators[502], DIGEST_B, ["x"], mtime_ns=MTIME + 2)
+        self.replace(locators[250], DIGEST_B, [], mtime_ns=MTIME + 1)
+        statements = []
+        conn = self.db.model.get_connection()
+        conn.set_trace_callback(statements.append)
+        try:
+            current = self.db.current_image_tags(
+                {
+                    loc: (SIZE, MTIME + (2 if i == 502 else 0))
+                    for i, loc in enumerate(locators)
+                }
+            )
+        finally:
+            conn.set_trace_callback(None)
+        self.assertEqual(current, {locators[0]: ["z", "y"], locators[502]: ["x"]})
+        lookups = [s for s in statements if "media_annotation_files" in s]
+        self.assertEqual(len(lookups), 2)
+        self.assertEqual(
+            list(self.db.iter_image_tag_records()),
+            [
+                (locators[0], SIZE, MTIME, ["z", "y"]),
+                (locators[250], SIZE, MTIME + 1, []),
+                (locators[502], SIZE, MTIME + 2, ["x"]),
+            ],
+        )
+
+
+class AnnotationIsolationTests(AnnotationTestCase):
+    def test_cascade_from_vocabulary_delete_keeps_byte_records(self):
+        loc = self.locator("a.png")
+        self.replace(loc, tags=["gone", "kept"])
+        self.db.delete_tag_all_prompts("gone")
+        self.assertEqual(
+            self.db.read_image_tags(loc, DIGEST_A),
+            {
+                "tags": ["kept"],
+                "tag_revision": 2,
+                "prior_count": 0,
+            },  # moved by the delete
+        )
+        self.schema_clean()
+
+    def test_junction_failure_leaves_nothing_behind(self):
+        loc = self.locator("a.png")
+        self.replace(loc, tags=["a"])
+        conn = self.db.model.get_connection()
+        conn.execute(
+            "CREATE TRIGGER reject_image_tag BEFORE INSERT ON media_image_tags "
+            "BEGIN SELECT RAISE(ABORT,'image tag failure'); END"
+        )
+        conn.commit()
+        before = self.rows(*ANNOTATION_TABLES, "tags")
+        with self.assertRaises(sqlite3.IntegrityError):
+            self.replace(loc, tags=["b"], revision=1)
+        with self.assertRaises(sqlite3.IntegrityError):
+            self.add(self.locator("new.png"), DIGEST_B, ["c"])
+        self.assertFalse(conn.in_transaction)
+        self.assertEqual(self.rows(*ANNOTATION_TABLES, "tags"), before)
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_api_admin.py

@@ -2476,7 +2476,8 @@ def _write_mp4(path, comment):
 
 
 class TestScanReadsVideos(AdminAPITestCase):
-    """Videos saved by ComfyUI carry the graph in their comment tag."""
+    """Videos carry the graph in per-key container tags (current savers) or in
+    the old VideoHelperSuite comment envelope; the scan reads one carrier."""
 
     def setUp(self):
         super().setUp()
@@ -2506,6 +2507,56 @@ class TestScanReadsVideos(AdminAPITestCase):
         images = self.api.db.get_prompt_images(prompts[0]["id"])
         self.assertTrue(images[0]["image_path"].endswith("clip.mp4"))
 
+    async def test_scan_links_per_key_tag_video_with_default_profile(self):
+        """A current saver's per-key tags scan exactly like the same graph as PNG."""
+        from utils import video_metadata
+
+        _write_png(self.output_dir / "same.png", prompt_text="a cat")
+        graph = {
+            "1": {"class_type": "CLIPTextEncode", "inputs": {"text": "a cat"}},
+            "2": {"class_type": "CLIPTextEncode", "inputs": {"text": "bad"}},
+            "3": {
+                "class_type": "KSampler",
+                "inputs": {"positive": ["1", 0], "negative": ["2", 0]},
+            },
+        }
+        (self.output_dir / "clip.mkv").write_bytes(b"\x1aE\xdf\xa3 fixture")
+        probed = {
+            "format": {
+                "tags": {
+                    "PROMPT": json.dumps(graph),
+                    "WORKFLOW": '{"nodes":[]}',
+                    "COMMENT": json.dumps({"prompt": "must not win"}),
+                }
+            }
+        }
+        GalleryConfig.MONITORING_DIRECTORIES = [self._root(self.output_dir)]
+
+        with (
+            patch.object(video_metadata.shutil, "which", return_value="/bin/ffprobe"),
+            patch.object(
+                video_metadata.subprocess,
+                "run",
+                return_value=MagicMock(returncode=0, stdout=json.dumps(probed)),
+            ) as run,
+            self.assertNoLogs(level="ERROR"),
+        ):
+            body = await (
+                await self.client.request("POST", "/prompt_manager/scan")
+            ).text()
+
+        self.assertEqual(run.call_args.args[0][-1], str(self.output_dir / "clip.mkv"))
+        self.assertIn('"found": 2', body)
+        prompts = self.api.db.get_recent_prompts(limit=10)["prompts"]
+        self.assertEqual(sorted(p["text"] for p in prompts), ["a cat", "bad"])
+        positive = next(p for p in prompts if p["text"] == "a cat")
+        images = self.api.db.get_prompt_images(positive["id"])
+        self.assertEqual(
+            sorted(os.path.basename(image["image_path"]) for image in images),
+            ["clip.mkv", "same.png"],
+        )
+        self.assertNotIn("must not win", json.dumps(prompts))
+
     async def test_scan_without_ffprobe_skips_videos_quietly(self):
         from unittest.mock import patch
         from utils import video_metadata

tests/test_api_annotation_suggest.py

@@ -0,0 +1,556 @@
+"""Scored top-N tag suggestions for a captured image (#249 B2).
+
+``POST /prompt_manager/image-annotations/suggest`` runs the selected legacy
+engine on the private capture (never the current file), behind the one
+process-wide model gate shared with the legacy Auto-tag routes. It never
+downloads a model and never writes. No real model is loaded: the engine is a
+fake service.
+"""
+
+import asyncio
+import json
+import math
+import os
+import threading
+import unittest
+import uuid
+from unittest.mock import MagicMock, patch
+
+try:
+    from test_api_annotations import AnnotationAPITestCase
+    from test_api_images import make_png
+except ImportError:
+    from tests.test_api_annotations import AnnotationAPITestCase
+    from tests.test_api_images import make_png
+
+try:  # guarded so a missing module fails assertions, not test loading
+    from py import model_gate
+except ImportError:
+    model_gate = None
+
+from py import config as config_module
+
+AutotagConfig = getattr(config_module, "AutotagConfig", None)
+
+SUGGEST = "/prompt_manager/image-annotations/suggest"
+WAIT = 30.0
+WD14 = {"general_threshold": 0.35, "character_threshold": 0.85}
+LLM = {"instructions": "Generate tags."}
+UNAVAILABLE = "Download it from the legacy Auto-tag page."
+
+SCORED = [("sandstone", 0.98), ("desert", 0.96)] + [
+    (f"tag_{i:03d}", round(0.9 - i * 0.005, 4)) for i in range(150)
+]
+
+
+class FakeTagger:
+    """Stand-in for ``AutoTagService`` with the suggestion entry point."""
+
+    default_prompt = "server default prompt"
+    custom_prompt = "legacy prompt"
+    wd14_general_threshold = 0.35
+    wd14_character_threshold = 0.85
+
+    def __init__(self):
+        self.models_config = {
+            "wd14-swinv2": {},
+            "wd14-vit": {},
+            "gguf": {},
+            "hf": {},
+        }
+        self.downloaded = {name: True for name in self.models_config}
+        self.loaded = None
+        self.load_calls = []
+        self.unload_calls = 0
+        self.calls = []
+        self.read_paths = []
+        self.read_bytes = []
+        self.download_model = MagicMock(side_effect=AssertionError("downloaded"))
+        self.load_error = None
+        self.block = None  # (entered, proceed) events
+
+    def get_models_status(self):
+        return {
+            name: {"downloaded": self.downloaded[name], "model_path": None}
+            for name in self.models_config
+        }
+
+    def is_model_loaded(self):
+        return self.loaded is not None
+
+    def get_loaded_model_type(self):
+        return self.loaded
+
+    def load_model(self, model_type, use_gpu=True):
+        self.load_calls.append((model_type, use_gpu))
+        if self.load_error is not None:
+            raise self.load_error
+        self.loaded = model_type
+        return True
+
+    def unload_model(self):
+        self.unload_calls += 1
+        self.loaded = None
+
+    def generate_tags(self, image_path, **kwargs):
+        return ["legacy"]
+
+    def generate_tag_suggestions(
+        self,
+        image_path,
+        *,
+        limit,
+        prompt=None,
+        general_threshold=None,
+        character_threshold=None,
+    ):
+        self.read_paths.append(image_path)
+        with open(image_path, "rb") as handle:
+            self.read_bytes.append(handle.read())
+        self.calls.append(
+            {
+                "limit": limit,
+                "prompt": prompt,
+                "general_threshold": general_threshold,
+                "character_threshold": character_threshold,
+                "model": self.loaded,
+            }
+        )
+        if self.block is not None:
+            entered, proceed = self.block
+            entered.set()
+            proceed.wait(WAIT)
+        if self.loaded.startswith("wd14"):
+            return list(SCORED[:limit])
+        return [(name, None) for name, _ in SCORED[:limit]]
+
+
+class SuggestTestCase(AnnotationAPITestCase):
+    async def get_application(self):
+        app = await super().get_application()
+        self.assertIsNotNone(model_gate, "py.model_gate is not implemented")
+        self.assertIsNotNone(AutotagConfig, "AutotagConfig is not implemented")
+        self.tagger = FakeTagger()
+        patcher = patch("py.autotag.get_autotag_service", return_value=self.tagger)
+        patcher.start()
+        self.addCleanup(patcher.stop)
+        AutotagConfig.SUGGESTION_LIMIT = 10
+        self.addCleanup(setattr, AutotagConfig, "SUGGESTION_LIMIT", 10)
+        return app
+
+    async def tearDownAsync(self):
+        leaked = model_gate.holder() if model_gate is not None else None
+        if leaked is not None:
+            model_gate.release(leaked)
+        await super().tearDownAsync()
+        self.assertIsNone(leaked, "the suggest route left the model gate held")
+
+    async def observed(self, path="sub/a.png", size=(8, 6)):
+        make_png(self.output_dir / path, size=size)
+        return await self.observe_ok(path)
+
+    def suggest_body(self, observed, model_type="wd14-vit", options=None, **changes):
+        if options is None:
+            options = dict(WD14 if model_type.startswith("wd14") else LLM)
+        return {
+            "target": dict(observed["target"]),
+            "capture_token": observed["capture"]["token"],
+            "expected_epoch": observed["library_epoch"],
+            "model_type": model_type,
+            "options": options,
+            **changes,
+        }
+
+    async def suggest(self, body):
+        response = await self.client.post(SUGGEST, json=body)
+        return response.status, await self.json_of(response)
+
+    async def suggest_ok(self, body):
+        status, data = await self.suggest(body)
+        self.assertEqual(status, 200, data)
+        return data
+
+    def capture_path(self, observed):
+        return self.api._captures.get(observed["capture"]["token"]).path
+
+    async def wait_until_free(self):
+        for _ in range(3000):
+            if model_gate.holder() is None:
+                return
+            await asyncio.sleep(0.01)
+        self.fail(f"model gate still held by {model_gate.holder()}")
+
+
+class TestSuggestions(SuggestTestCase):
+    async def test_wd14_suggestions_are_scored_top_n_from_the_capture(self):
+        observed = await self.observed()
+
+        body = await self.suggest_ok(self.suggest_body(observed))
+
+        self.assertEqual(
+            body["suggestions"][:2],
+            [
+                {"name": "sandstone", "confidence": 0.98},
+                {"name": "desert", "confidence": 0.96},
+            ],
+        )
+        self.assertEqual(len(body["suggestions"]), 10)
+        self.assertEqual(body["model_type"], "wd14-vit")
+        self.assertEqual(
+            body["options"],
+            {
+                "general_threshold": 0.35,
+                "character_threshold": 0.85,
+                "use_gpu": True,
+                "suggestion_limit": 10,
+            },
+        )
+        self.assertEqual(self.tagger.read_paths, [self.capture_path(observed)])
+        self.assertEqual(
+            self.tagger.calls[0],
+            {
+                "limit": 10,
+                "prompt": None,
+                "general_threshold": 0.35,
+                "character_threshold": 0.85,
+                "model": "wd14-vit",
+            },
+        )
+        self.assertEqual(self.tagger.load_calls, [("wd14-vit", True)])
+        self.tagger.download_model.assert_not_called()
+        self.assertIsNone(model_gate.holder())
+        self.assert_no_paths(body)
+
+    async def test_suggestions_keep_the_requested_limit(self):
+        observed = await self.observed()
+        for limit in (1, 10, 100):
+            with self.subTest(limit=limit):
+                options = {**WD14, "suggestion_limit": limit}
+                body = await self.suggest_ok(
+                    self.suggest_body(observed, options=options)
+                )
+                self.assertEqual(len(body["suggestions"]), limit)
+                self.assertEqual(body["options"]["suggestion_limit"], limit)
+        # Omitted: the persisted autotag.defaults.suggestion_limit.
+        AutotagConfig.SUGGESTION_LIMIT = 3
+        body = await self.suggest_ok(self.suggest_body(observed))
+        self.assertEqual(len(body["suggestions"]), 3)
+        self.assertEqual(body["options"]["suggestion_limit"], 3)
+
+    async def test_llm_suggestions_send_instructions_and_null_confidence(self):
+        observed = await self.observed()
+        self.tagger.custom_prompt = "legacy prompt"
+        for model_type in ("gguf", "hf"):
+            with self.subTest(model_type=model_type):
+                options = {"instructions": f"tags for {model_type}", "use_gpu": False}
+                body = await self.suggest_ok(
+                    self.suggest_body(observed, model_type, options)
+                )
+                self.assertEqual(body["model_type"], model_type)
+                self.assertEqual(body["options"], {**options, "suggestion_limit": 10})
+                self.assertEqual(
+                    body["suggestions"][:2],
+                    [
+                        {"name": "sandstone", "confidence": None},
+                        {"name": "desert", "confidence": None},
+                    ],
+                )
+                call = self.tagger.calls[-1]
+                self.assertEqual(call["prompt"], f"tags for {model_type}")
+                self.assertIsNone(call["general_threshold"])
+                self.assertEqual(self.tagger.load_calls[-1], (model_type, False))
+        self.assertEqual(self.tagger.custom_prompt, "legacy prompt")
+
+    async def test_loaded_model_is_reused_and_another_is_swapped_in(self):
+        observed = await self.observed()
+        self.tagger.loaded = "wd14-vit"
+        await self.suggest_ok(self.suggest_body(observed))
+        self.assertEqual(self.tagger.load_calls, [])
+        await self.suggest_ok(self.suggest_body(observed, "wd14-swinv2"))
+        self.assertEqual(self.tagger.load_calls, [("wd14-swinv2", True)])
+        self.assertEqual(self.tagger.calls[-1]["model"], "wd14-swinv2")
+
+    async def test_suggest_uses_capture_bytes_not_current_file(self):
+        observed = await self.observed()
+        original = (self.output_dir / "sub" / "a.png").read_bytes()
+        make_png(self.output_dir / "sub" / "a.png", size=(9, 9), color=(0, 0, 255))
+        self.assertNotEqual((self.output_dir / "sub" / "a.png").read_bytes(), original)
+
+        await self.suggest_ok(self.suggest_body(observed))
+
+        capture_path = self.capture_path(observed)
+        self.assertEqual(self.tagger.read_paths, [capture_path])
+        self.assertEqual(self.tagger.read_bytes, [original])
+        self.assertNotEqual(
+            os.path.normcase(capture_path),
+            os.path.normcase(str(self.output_dir / "sub" / "a.png")),
+        )
+
+    async def test_suggest_writes_nothing(self):
+        observed = await self.observed()
+        before = self.db_snapshot()
+        rows = self.bytes_rows()
+
+        await self.suggest_ok(self.suggest_body(observed))
+        await self.suggest_ok(self.suggest_body(observed, "gguf"))
+
+        self.assertEqual(self.db_snapshot(), before)
+        self.assertEqual(self.bytes_rows(), rows)
+        # The capture stays usable for Save after Generate.
+        self.assertIsNotNone(self.api._captures.get(observed["capture"]["token"]))
+
+
+class TestModelAvailability(SuggestTestCase):
+    async def test_suggest_never_downloads_and_reports_model_unavailable(self):
+        observed = await self.observed()
+        self.tagger.downloaded["wd14-vit"] = False
+
+        status, body = await self.suggest(self.suggest_body(observed))
+
+        self.assertEqual((status, body["code"]), (409, "model-unavailable"))
+        self.assertIn(UNAVAILABLE, body["error"])
+        self.assertIn("WD14 ViT", body["error"])
+        self.tagger.download_model.assert_not_called()
+        self.assertEqual(self.tagger.load_calls, [])
+        self.assertEqual(self.tagger.calls, [])
+        self.assertIsNone(model_gate.holder())
+
+    async def test_load_failure_is_a_generic_error_and_frees_everything(self):
+        observed = await self.observed()
+        self.tagger.load_error = RuntimeError(f"cannot load {self.output_dir}")
+
+        status, body = await self.suggest(self.suggest_body(observed))
+
+        self.assertEqual((status, body["code"]), (500, "suggest-failed"))
+        self.assert_no_paths(body)
+        self.assertIsNone(model_gate.holder())
+        entry = self.api._captures._entries[observed["capture"]["token"]]
+        self.assertEqual(entry.uses, 0)
+
+    async def test_gate_work_that_never_starts_releases_the_capture_pin(self):
+        observed = await self.observed()
+        loop = asyncio.get_running_loop()
+        real = loop.run_in_executor
+
+        def refuse_gate_work(executor, func, *args):
+            if getattr(func, "__qualname__", "").startswith("ModelGate.run"):
+                raise RuntimeError("cannot schedule new futures after shutdown")
+            return real(executor, func, *args)
+
+        with patch.object(loop, "run_in_executor", refuse_gate_work):
+            status, body = await self.suggest(self.suggest_body(observed))
+
+        self.assertEqual((status, body["code"]), (500, "suggest-failed"))
+        self.assertIsNone(model_gate.holder())
+        entry = self.api._captures._entries[observed["capture"]["token"]]
+        self.assertEqual(entry.uses, 0)
+        self.assertEqual(self.tagger.calls, [])
+
+
+class TestModelGate(SuggestTestCase):
+    async def _blocked_suggest(self, observed, model_type="wd14-vit"):
+        entered, proceed = threading.Event(), threading.Event()
+        self.tagger.block = (entered, proceed)
+        task = asyncio.ensure_future(
+            self.client.post(SUGGEST, json=self.suggest_body(observed, model_type))
+        )
+        await asyncio.get_running_loop().run_in_executor(None, entered.wait, WAIT)
+        self.assertTrue(entered.is_set())
+        return task, proceed
+
+    async def test_suggest_reports_busy_while_legacy_batch_runs(self):
+        observed = await self.observed()
+        prompt_id = self._save_prompt("batch prompt")
+        self.api.db.link_image_to_prompt(prompt_id, str(self.output_dir / "sub/a.png"))
+        started, proceed = threading.Event(), threading.Event()
+
+        def slow_generate(image_path, **kwargs):
+            started.set()
+            proceed.wait(WAIT)
+            return ["batch"]
+
+        self.tagger.generate_tags = slow_generate
+        batch = asyncio.ensure_future(
+            self.client.post("/prompt_manager/autotag/start?model_type=wd14-vit")
+        )
+        try:
+            await asyncio.get_running_loop().run_in_executor(None, started.wait, WAIT)
+            status, body = await self.suggest(self.suggest_body(observed, "gguf"))
+            self.assertEqual((status, body["code"]), (409, "model-busy"))
+            self.assertEqual(self.tagger.calls, [])
+            self.assertEqual(self.tagger.loaded, "wd14-vit")  # not swapped
+        finally:
+            proceed.set()
+            response = await batch
+            await response.read()
+        await self.wait_until_free()
+        await self.suggest_ok(self.suggest_body(observed))
+
+    async def test_legacy_unload_returns_model_busy_while_native_generate_runs(self):
+        observed = await self.observed()
+        task, proceed = await self._blocked_suggest(observed)
+        try:
+            response = await self.client.post("/prompt_manager/autotag/unload")
+            self.assertEqual(response.status, 409)
+            self.assertEqual((await response.json())["code"], "model-busy")
+            response = await self.client.post(
+                "/prompt_manager/autotag/single",
+                json={"root_index": 0, "path": "sub/a.png", "model_type": "gguf"},
+            )
+            self.assertEqual(response.status, 409)
+            status, body = await self.suggest(self.suggest_body(observed, "gguf"))
+            self.assertEqual((status, body["code"]), (409, "model-busy"))
+            self.assertEqual(self.tagger.unload_calls, 0)
+            self.assertEqual(self.tagger.loaded, "wd14-vit")
+        finally:
+            proceed.set()
+            response = await task
+        self.assertEqual(response.status, 200)
+        await self.wait_until_free()
+        self.assertEqual([call["model"] for call in self.tagger.calls], ["wd14-vit"])
+
+    async def test_gate_held_after_client_disconnect_until_executor_finishes(self):
+        observed = await self.observed()
+        token = observed["capture"]["token"]
+        task, proceed = await self._blocked_suggest(observed)
+        task.cancel()
+        with self.assertRaises(asyncio.CancelledError):
+            await task
+        try:
+            # The client is gone; the inference still runs, so the gate and
+            # the capture pin stay until the worker thread exits.
+            self.assertIsNotNone(model_gate.holder())
+            response = await self.client.post("/prompt_manager/autotag/unload")
+            self.assertEqual(response.status, 409)
+            self.assertEqual(self.api._captures._entries[token].uses, 1)
+        finally:
+            proceed.set()
+        await self.wait_until_free()
+        for _ in range(3000):
+            if self.api._captures._entries[token].uses == 0:
+                break
+            await asyncio.sleep(0.01)
+        self.assertEqual(self.api._captures._entries[token].uses, 0)
+        response = await self.client.post("/prompt_manager/autotag/unload")
+        self.assertEqual(response.status, 200)
+
+
+class TestRequestValidation(SuggestTestCase):
+    async def test_strict_option_validation(self):
+        observed = await self.observed()
+        bad_options = {
+            "wd14-vit": [
+                {**WD14, "general_threshold": True},
+                {**WD14, "general_threshold": math.nan},
+                {**WD14, "general_threshold": math.inf},
+                {**WD14, "general_threshold": "0.3"},
+                {**WD14, "general_threshold": 1.5},
+                {**WD14, "character_threshold": -0.1},
+                {"general_threshold": 0.35},
+                {**WD14, "instructions": "x"},
+                {**WD14, "unknown": 1},
+                {**WD14, "use_gpu": "yes"},
+                {**WD14, "suggestion_limit": 0},
+                {**WD14, "suggestion_limit": 101},
+                {**WD14, "suggestion_limit": 2.5},
+                {**WD14, "suggestion_limit": "10"},
+                {**WD14, "suggestion_limit": True},
+            ],
+            "gguf": [
+                {**LLM, "general_threshold": 0.35},
+                {**LLM, "character_threshold": 0.85},
+                {"instructions": "x" * 4001},
+                {"instructions": "\ud800"},  # not encodable as UTF-8
+                {"instructions": "   "},
+                {"instructions": 5},
+                {},
+                {**LLM, "suggestion_limit": 0},
+            ],
+        }
+        for model_type, cases in bad_options.items():
+            for options in cases:
+                with self.subTest(model_type=model_type, options=repr(options)[:80]):
+                    status, body = await self.suggest(
+                        self.suggest_body(observed, model_type, options)
+                    )
+                    self.assertEqual(status, 400, body)
+        for body in (
+            {**self.suggest_body(observed), "extra": 1},
+            {k: v for k, v in self.suggest_body(observed).items() if k != "options"},
+            self.suggest_body(observed, options=[]),
+            self.suggest_body(observed, model_type="ollama"),
+            self.suggest_body(observed, model_type=7, options=dict(WD14)),
+            self.suggest_body(observed, capture_token="../x"),
+            self.suggest_body(observed, expected_epoch="nope"),
+            self.suggest_body(observed, target={"root_index": 0}),
+        ):
+            with self.subTest(body=json.dumps(body)[:120]):
+                status, _ = await self.suggest(body)
+                self.assertEqual(status, 400)
+        # The edges that are allowed.
+        await self.suggest_ok(
+            self.suggest_body(
+                observed,
+                options={
+                    "general_threshold": 0,
+                    "character_threshold": 1,
+                    "suggestion_limit": 100,
+                },
+            )
+        )
+        await self.suggest_ok(
+            self.suggest_body(observed, "hf", options={"instructions": "x" * 4000})
+        )
+        self.assertEqual(len(self.tagger.calls), 2)
+        self.tagger.download_model.assert_not_called()
+
+    async def test_capture_token_must_match_target_and_epoch(self):
+        observed = await self.observed()
+        make_png(self.output_dir / "sub" / "b.png")
+        other = await self.observe_ok("sub/b.png")
+
+        status, body = await self.suggest(
+            self.suggest_body(observed, capture_token="x" * 43)
+        )
+        self.assertEqual((status, body["code"]), (409, "capture-expired"))
+        status, body = await self.suggest(
+            self.suggest_body(observed, capture_token=other["capture"]["token"])
+        )
+        self.assertEqual((status, body["code"]), (409, "capture-expired"))
+        status, body = await self.suggest(
+            self.suggest_body(observed, expected_epoch=str(uuid.uuid4()))
+        )
+        self.assertEqual((status, body["code"]), (409, "storage-changed"))
+        self.assertEqual(body["library_epoch"], self.epoch)
+        status, body = await self.suggest(
+            self.suggest_body(
+                observed, target={**observed["target"], "root_scope": "0" * 64}
+            )
+        )
+        self.assertEqual((status, body["code"]), (409, "source-changed"))
+        self.api._captures.release(observed["capture"]["token"])
+        status, body = await self.suggest(self.suggest_body(observed))
+        self.assertEqual((status, body["code"]), (409, "capture-expired"))
+        self.assertEqual(self.tagger.calls, [])
+        for token in (observed, other):
+            entry = self.api._captures._entries.get(token["capture"]["token"])
+            self.assertTrue(entry is None or entry.uses == 0)
+
+    async def test_video_and_oversized_dimensions_unsupported_media(self):
+        (self.output_dir / "clip.mp4").write_bytes(b"\x00\x00\x00\x18ftypmp42" * 4)
+        video = await self.observe_ok("clip.mp4")
+        self.assertFalse(video["capture"]["generate"]["supported"])
+        big = await self.observed("big.png", size=(4097, 1))
+        self.assertFalse(big["capture"]["generate"]["supported"])
+
+        for observed in (video, big):
+            status, body = await self.suggest(self.suggest_body(observed))
+            self.assertEqual((status, body["code"]), (415, "unsupported-media"))
+            self.assertEqual(body["error"], observed["capture"]["generate"]["reason"])
+        self.assertEqual(self.tagger.calls, [])
+        self.assertEqual(self.tagger.load_calls, [])
+        self.assertIsNone(model_gate.holder())
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_api_autotag.py

@@ -11,7 +11,6 @@ import shutil
 import sys
 import tempfile
 import threading
-import types
 import unittest
 from unittest.mock import MagicMock
 from pathlib import Path
@@ -23,12 +22,12 @@ sys.modules.setdefault("server", _mock_server)  # py.config imports it at module
 
 sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
 
-if "folder_paths" not in sys.modules:
-    _stub = types.ModuleType("folder_paths")
-    _stub.models_dir = tempfile.gettempdir()
-    _stub.base_path = tempfile.gettempdir()
-    _stub.get_output_directory = tempfile.gettempdir
-    sys.modules["folder_paths"] = _stub
+try:
+    from host_stubs import install_folder_paths_stub
+except ImportError:  # imported as tests.test_api_autotag
+    from tests.host_stubs import install_folder_paths_stub
+
+install_folder_paths_stub()
 
 from aiohttp import web
 from aiohttp.test_utils import AioHTTPTestCase
@@ -37,6 +36,11 @@ from database.operations import PromptDatabase
 from py.api import PromptManagerAPI
 from utils.hashing import generate_prompt_hash
 
+try:  # guarded so a missing module fails assertions, not test loading
+    from py import model_gate
+except ImportError:
+    model_gate = None
+
 
 class FakeAutotagService:
     """Stand-in for AutoTagService that records calls and returns fixed tags."""
@@ -128,6 +132,11 @@ class AutotagAPITestCase(AioHTTPTestCase):
     async def tearDownAsync(self):
         await super().tearDownAsync()  # closes the aiohttp test client
         self._service_patch.stop()
+        if model_gate is not None:
+            leaked = model_gate.holder()
+            if leaked is not None:  # never leak the process-wide gate
+                model_gate.release(leaked)
+            self.assertIsNone(leaked, "a route left the model gate held")
         self.api.db.close_all()
         for path in (
             self._temp_db.name,
@@ -947,6 +956,159 @@ class TestUnloadAutotag(AutotagAPITestCase):
         self.assertEqual(resp.status, 500)
 
 
+class TestLegacyRoutesShareTheModelGate(AutotagAPITestCase):
+    """While another model operation holds the gate, every legacy model route
+    answers ``409 model-busy`` and touches no model (the one allowed legacy
+    change)."""
+
+    BUSY = {"success": False, "code": "model-busy"}
+
+    def setUp(self):
+        super().setUp()
+        self.assertIsNotNone(model_gate, "py.model_gate is not implemented")
+
+    async def _busy(self, resp):
+        self.assertEqual(resp.status, 409)
+        data = await resp.json()
+        self.assertEqual({k: data[k] for k in self.BUSY}, self.BUSY)
+        self.assertIn("busy", data["error"].lower())
+        return data
+
+    async def test_legacy_single_download_start_return_model_busy_while_gate_held(
+        self,
+    ):
+        image = self._make_image(self.output_dir)
+        self.service._loaded_type = "wd14-vit"
+        self.assertTrue(model_gate.try_acquire("test-holder"))
+        try:
+            await self._busy(
+                await self._post_json(
+                    "/prompt_manager/autotag/single",
+                    {"path": str(image), "model_type": "gguf"},
+                )
+            )
+            await self._busy(
+                await self.client.request(
+                    "POST", "/prompt_manager/autotag/download/gguf"
+                )
+            )
+            await self._busy(
+                await self.client.request(
+                    "POST", "/prompt_manager/autotag/start?model_type=gguf"
+                )
+            )
+            await self._busy(
+                await self.client.request("POST", "/prompt_manager/autotag/unload")
+            )
+            self.assertEqual(model_gate.holder(), "test-holder")
+        finally:
+            model_gate.release("test-holder")
+
+        self.assertEqual(self.service.load_calls, [])
+        self.assertEqual(self.service.generate_calls, [])
+        self.assertEqual(self.service.download_calls, [])
+        self.assertEqual(self.service.unload_calls, 0)
+        self.assertEqual(self.service.get_loaded_model_type(), "wd14-vit")
+
+        # A refused start released the batch lock: the next start runs.
+        resp = await self.client.request(
+            "POST", "/prompt_manager/autotag/start?model_type=gguf"
+        )
+        self.assertEqual(resp.status, 200)
+        self.assertEqual(self._sse_events(await resp.text())[-1]["type"], "complete")
+
+    async def test_legacy_routes_hold_the_gate_only_while_they_run(self):
+        image = self._make_image(self.output_dir)
+        seen = []
+
+        def generate(image_path, **kwargs):
+            seen.append(model_gate.holder())
+            return ["x"]
+
+        def download(model_type, progress_callback=None):
+            seen.append(model_gate.holder())
+            return True
+
+        def load(model_type, use_gpu=True):
+            seen.append(model_gate.holder())
+            self.service._loaded_type = model_type
+            return True
+
+        self.service.generate_tags = generate
+        self.service.download_model = download
+        self.service.load_model = load
+
+        resp = await self._post_json(
+            "/prompt_manager/autotag/single", {"path": str(image)}
+        )
+        self.assertEqual(resp.status, 200)
+        self.assertIsNone(model_gate.holder())
+        resp = await self.client.request(
+            "POST", "/prompt_manager/autotag/download/gguf"
+        )
+        self.assertEqual(self._sse_events(await resp.text())[-1]["type"], "complete")
+        self.assertIsNone(model_gate.holder())
+        resp = await self.client.request("POST", "/prompt_manager/autotag/unload")
+        self.assertEqual(resp.status, 200)
+        self.assertIsNone(model_gate.holder())
+
+        self.assertEqual(len(seen), 3)
+        self.assertTrue(all(owner and owner.startswith("legacy-") for owner in seen))
+
+    async def test_validation_errors_win_over_the_gate(self):
+        self.assertTrue(model_gate.try_acquire("test-holder"))
+        try:
+            resp = await self._post_json("/prompt_manager/autotag/single", {})
+            self.assertEqual(resp.status, 400)
+            resp = await self.client.request(
+                "POST", "/prompt_manager/autotag/start?general_threshold=x"
+            )
+            self.assertEqual(resp.status, 400)
+        finally:
+            model_gate.release("test-holder")
+
+    async def test_batch_holds_the_gate_until_it_completes(self):
+        prompt = self._save_prompt("p")
+        self._link_image(prompt, self._make_image(self.output_dir, "a.png"))
+        started = threading.Event()
+        proceed = threading.Event()
+
+        def slow_generate(image_path, **kwargs):
+            started.set()
+            proceed.wait(30)
+            return ["slow"]
+
+        self.service.generate_tags = slow_generate
+        run = asyncio.ensure_future(
+            self.client.request("POST", "/prompt_manager/autotag/start")
+        )
+        try:
+            await asyncio.get_running_loop().run_in_executor(None, started.wait, 30)
+            self.assertTrue(model_gate.holder().startswith("legacy-batch:"))
+            await self._busy(
+                await self.client.request("POST", "/prompt_manager/autotag/unload")
+            )
+        finally:
+            proceed.set()
+            resp = await run
+            events = self._sse_events(await resp.text())
+        self.assertEqual(events[-1]["type"], "complete")
+        self.assertIsNone(model_gate.holder())
+
+    async def test_models_response_carries_the_suggestion_limit(self):
+        from py.config import AutotagConfig
+
+        resp = await self.client.get("/prompt_manager/autotag/models")
+        data = await resp.json()
+        self.assertEqual(data["suggestion_limit"], 10)
+        AutotagConfig.SUGGESTION_LIMIT = 25
+        try:
+            resp = await self.client.get("/prompt_manager/autotag/models")
+            self.assertEqual((await resp.json())["suggestion_limit"], 25)
+        finally:
+            AutotagConfig.SUGGESTION_LIMIT = 10
+
+
 class TestScanOutputDir(AutotagAPITestCase):
 
     async def test_lists_images_with_thumbnails(self):

tests/test_api_core.py

@@ -448,6 +448,20 @@ class TestUiRoutes(CoreRoutesTestCase):
         resp = await self.client.request("GET", "/prompt_manager/test")
         self.assertTrue((await resp.json())["success"])
 
+    async def test_image_annotation_routes_are_registered(self):
+        table = {
+            (route.method, route.resource.canonical)
+            for route in self.app.router.routes()
+            if route.resource is not None
+        }
+        for entry in (
+            ("POST", "/prompt_manager/image-annotations/observe"),
+            ("GET", "/prompt_manager/image-annotations/capture/{token}"),
+            ("DELETE", "/prompt_manager/image-annotations/capture/{token}"),
+            ("PUT", "/prompt_manager/image-annotations/tags"),
+        ):
+            self.assertIn(entry, table)
+
 
 class TestStaticRoutes(CoreRoutesTestCase):
 

tests/test_api_images.py

@@ -18,7 +18,12 @@ from unittest.mock import AsyncMock, MagicMock, patch
 sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
 
 # ComfyUI-only modules are stubbed so the API can be imported standalone.
-sys.modules.setdefault("folder_paths", MagicMock())
+try:
+    from host_stubs import install_folder_paths_stub
+except ImportError:  # imported as tests.test_api_images
+    from tests.host_stubs import install_folder_paths_stub
+
+install_folder_paths_stub()
 sys.modules.setdefault("server", MagicMock())
 
 from aiohttp import web  # noqa: E402

tests/test_api_prompt_inspector.py

@@ -1,7 +1,11 @@
 """Physical-file prompt inspection is bounded, conservative and read-only."""
 
+import io
 import json
-from unittest.mock import patch
+import shutil
+import subprocess
+import unittest
+from unittest.mock import MagicMock, patch
 
 from PIL import Image, PngImagePlugin
 
@@ -15,6 +19,9 @@ except ImportError:
 from utils.pair_execution import EXECUTION_KEY, record_pair_execution
 from database.pair_ingestion import ingest_resolved_pair
 from utils.prompt_pair_resolver import resolve_prompt_pair
+from utils import video_metadata
+
+HAVE_FFMPEG = bool(shutil.which("ffmpeg") and shutil.which("ffprobe"))
 
 
 def graph(positive="portrait", negative="blurry"):
@@ -28,6 +35,54 @@ def graph(positive="portrait", negative="blurry"):
     }
 
 
+def write_video(path, metadata_tags=True, **tags):
+    """Tiny black clip carrying one container tag per keyword (needs ffmpeg).
+
+    MP4 keeps arbitrary keys only with ``use_metadata_tags`` (mdta atoms), the
+    movflag the core SaveVideo node sets; without it ffmpeg writes the legacy
+    ilst atoms that old VideoHelperSuite files carry. Matroska and WebM
+    upper-case the key names.
+    """
+    command = [
+        "ffmpeg",
+        "-v",
+        "quiet",
+        "-f",
+        "lavfi",
+        "-i",
+        "color=c=black:s=16x16:d=0.2",
+    ]
+    if metadata_tags and path.suffix.lower() in (".mp4", ".mov", ".m4v"):
+        command += ["-movflags", "use_metadata_tags"]
+    for key, value in tags.items():
+        command += ["-metadata", f"{key}={value}"]
+    subprocess.run(command + ["-y", str(path)], check=True, timeout=60)
+    return path
+
+
+class ProbeOutput(io.BytesIO):
+    """ffprobe stdout that records the bounded read size asked of it."""
+
+    requested_size = None
+
+    def read(self, size=-1):
+        self.requested_size = size
+        return super().read(size)
+
+
+def probe_process(payload, poll=0, returncode=0):
+    """A fake ffprobe whose stdout is ``payload`` bytes."""
+    process = MagicMock()
+    process.stdout = ProbeOutput(payload)
+    process.poll.return_value = poll
+    process.returncode = returncode
+    return process
+
+
+def probe_payload(tags):
+    return json.dumps({"format": {"tags": tags}}).encode()
+
+
 def png(path, **metadata):
     info = PngImagePlugin.PngInfo()
     for key, value in metadata.items():
@@ -312,7 +367,7 @@ class TestPromptInspector(ImageAPITestCase):
     async def test_corrupt_truncated_duplicate_and_expanding_metadata_fail_closed(self):
         import struct
         import zlib
-        from utils.prompt_inspector import MAX_METADATA_BYTES
+        from utils.prompt_inspector import VIDEO_REPROBE_BYTES
 
         path = self.output_dir / "image.png"
         base = png(path, prompt=graph()).read_bytes()
@@ -325,8 +380,8 @@ class TestPromptInspector(ImageAPITestCase):
         cases.append(bytes(broken))
         for kind, payload in (
             (b"tEXt", b"prompt\0{}"),
-            (b"zTXt", b"prompt\0\0" + zlib.compress(b"x" * (MAX_METADATA_BYTES + 1))),
-            (b"tEXt", b"prompt\0" + b"x" * (MAX_METADATA_BYTES + 1)),
+            (b"zTXt", b"prompt\0\0" + zlib.compress(b"x" * (VIDEO_REPROBE_BYTES + 1))),
+            (b"tEXt", b"prompt\0" + b"x" * (VIDEO_REPROBE_BYTES + 1)),
             (b"tEXt", b"bad-key-without-separator"),
             (b"zTXt", b"prompt\0\1bad-compression"),
             (b"iTXt", b"prompt\0\2\0\0\0text"),
@@ -385,9 +440,7 @@ class TestPromptInspector(ImageAPITestCase):
         self.assertEqual(result["prompts"]["positive"]["status"], "absent")
 
     async def test_video_inspection_uses_bounded_container_evidence(self):
-        import io
-        from unittest.mock import MagicMock
-        from utils.prompt_inspector import MAX_METADATA_BYTES
+        from utils.prompt_inspector import VIDEO_PROBE_BYTES, VIDEO_REPROBE_BYTES
 
         (self.output_dir / "clip.mp4").write_bytes(b"video fixture")
         metadata = {"prompt": graph("Video portrait", "")}
@@ -412,14 +465,25 @@ class TestPromptInspector(ImageAPITestCase):
         process.wait.assert_called()
         self.assertNotIn("shell", spawn.call_args.kwargs)
         self.assertIn(str(self.output_dir / "clip.mp4"), spawn.call_args.args[0])
-        process.stdout = io.BytesIO(b"x" * (MAX_METADATA_BYTES + 1))
+        # An envelope over the probe bound cannot be split: the role-only
+        # re-probe overflows too and the read fails instead of guessing.
+        processes = [
+            probe_process(b"x" * (VIDEO_PROBE_BYTES + 1)),
+            probe_process(b"x" * (VIDEO_REPROBE_BYTES + 1)),
+        ]
         with (
             patch("utils.prompt_inspector.ffprobe_path", return_value="/tool/ffprobe"),
-            patch("utils.prompt_inspector.subprocess.Popen", return_value=process),
+            patch(
+                "utils.prompt_inspector.subprocess.Popen", side_effect=processes
+            ) as spawn,
         ):
             _, result = await self.inspect("clip.mp4")
+        self.assertEqual(spawn.call_count, 2)
         self.assertEqual(result["prompts"]["positive"]["status"], "unresolved")
-        self.assertTrue(process.stdout.closed)
+        self.assertEqual(result["metadata"]["status"], "unreadable")
+        for process in processes:
+            self.assertTrue(process.stdout.closed)
+            process.wait.assert_called()
 
     async def test_video_missing_probe_and_timeout_preserve_uncertainty(self):
         import io
@@ -895,11 +959,9 @@ class TestPromptInspector(ImageAPITestCase):
         self.assertNotIn("\ufffd".encode(), body)
         self.assertLessEqual(len(body), 8 * 1024 * 1024)
 
-    async def video_inspect(self, comment=None, description=None):
-        import io
+    async def video_inspect(self, comment=None, description=None, **per_key):
         import threading
-        from unittest.mock import MagicMock
-        from utils.prompt_inspector import MAX_METADATA_BYTES
+        from utils.prompt_inspector import VIDEO_PROBE_BYTES
 
         path = self.output_dir / "clip.mp4"
         path.write_bytes(b"fixture")
@@ -908,17 +970,8 @@ class TestPromptInspector(ImageAPITestCase):
             for key, value in (("comment", comment), ("description", description))
             if value is not None
         }
-        payload = json.dumps({"format": {"tags": tags}}).encode()
-
-        class ProbeOutput(io.BytesIO):
-            def read(self, size=-1):
-                self.requested_size = size
-                return super().read(size)
-
-        process = MagicMock()
-        process.stdout = ProbeOutput(payload)
-        process.poll.return_value = 0
-        process.returncode = 0
+        tags.update(per_key)
+        process = probe_process(probe_payload(tags))
         timers = []
         timer_type = threading.Timer
 
@@ -937,11 +990,156 @@ class TestPromptInspector(ImageAPITestCase):
             response, result = await self.inspect("clip.mp4")
         self.assertEqual(spawn.call_count, 1)
         self.assertTrue(process.stdout.closed)
-        self.assertEqual(process.stdout.requested_size, MAX_METADATA_BYTES + 1)
+        self.assertEqual(process.stdout.requested_size, VIDEO_PROBE_BYTES + 1)
         self.assertEqual(len(timers), 1)
         self.assertTrue(timers[0].finished.is_set())
+        self.assertEqual(
+            spawn.call_args.args[0][spawn.call_args.args[0].index("-show_entries") + 1],
+            "format_tags=prompt,workflow,parameters,"
+            "comfyassets_prompt_pairs,comment,description",
+        )
         return response, result
 
+    async def test_video_separate_prompt_workflow_tags_are_read(self):
+        source = graph("per-key portrait", "per-key negative")
+        workflow = '{"nodes":[{"id":"per-key"}],"cfg":6.5000000000000000001}'
+        for tags in (
+            {"prompt": json.dumps(source), "workflow": workflow},
+            {"PROMPT": json.dumps(source), "WORKFLOW": workflow},
+        ):
+            _, result = await self.video_inspect(
+                json.dumps({"prompt": graph("comment must not win")}), **tags
+            )
+            self.assertEqual(
+                result["prompts"]["positive"], self.side("per-key portrait")
+            )
+            self.assertEqual(
+                result["prompts"]["negative"], self.side("per-key negative")
+            )
+            self.assertEqual(
+                result["metadata"]["record_formats"],
+                {"prompt": "original-text", "workflow": "original-text"},
+            )
+            self.assertEqual(result["metadata"]["raw"]["workflow"], workflow)
+            self.assertEqual(
+                result["metadata"]["workflow_graph"]["value"]["cfg"],
+                "6.5000000000000000001",
+            )
+            self.assertNotIn("must not win", json.dumps(result))
+        _, result = await self.video_inspect(
+            **{"PROMPT": json.dumps(source), "prompt": json.dumps(graph("other"))}
+        )
+        self.assertEqual(result["prompts"]["positive"]["status"], "unresolved")
+        self.assertEqual(result["metadata"]["status"], "unreadable")
+
+    @unittest.skipUnless(HAVE_FFMPEG, "ffmpeg/ffprobe not installed")
+    async def test_video_per_key_and_comment_tags_read_from_real_containers(self):
+        video_metadata.ffprobe_path.cache_clear()
+        self.addCleanup(video_metadata.ffprobe_path.cache_clear)
+        source = json.dumps(graph("real portrait", ""))
+        workflow = '{"nodes":[{"id":"real"}]}'
+        for name in ("per_key.mp4", "per_key.mkv", "per_key.webm"):
+            write_video(self.output_dir / name, prompt=source, workflow=workflow)
+            response, result = await self.inspect(name)
+            self.assertEqual(response.status, 200, name)
+            self.assertEqual(
+                result["prompts"]["positive"], self.side("real portrait"), name
+            )
+            self.assertEqual(result["metadata"]["raw"]["workflow"], workflow, name)
+            self.assertEqual(
+                result["metadata"]["record_formats"],
+                {"prompt": "original-text", "workflow": "original-text"},
+                name,
+            )
+        # Old VideoHelperSuite envelope in the legacy ilst comment atom, unchanged.
+        write_video(
+            self.output_dir / "vhs.mp4",
+            metadata_tags=False,
+            comment=json.dumps({"prompt": json.loads(source), "workflow": {"v": 1}}),
+        )
+        _, result = await self.inspect("vhs.mp4")
+        self.assertEqual(result["prompts"]["positive"], self.side("real portrait"))
+        self.assertEqual(
+            result["metadata"]["record_formats"],
+            {"prompt": "reserialized-json", "workflow": "reserialized-json"},
+        )
+        self.assertEqual(result["metadata"]["workflow_graph"]["value"], {"v": 1})
+
+    async def test_video_oversize_workflow_reprobes_role_only_and_marks_limited(self):
+        from utils.prompt_inspector import VIDEO_PROBE_BYTES, VIDEO_REPROBE_BYTES
+
+        path = self.output_dir / "clip.mp4"
+        path.write_bytes(b"fixture")
+        source = json.dumps(graph("bounded portrait", ""))
+        processes = [
+            probe_process(b"x" * (VIDEO_PROBE_BYTES + 1)),
+            probe_process(probe_payload({"PROMPT": source})),
+        ]
+        with (
+            patch("utils.prompt_inspector.ffprobe_path", return_value="/tool/ffprobe"),
+            patch(
+                "utils.prompt_inspector.subprocess.Popen", side_effect=processes
+            ) as spawn,
+        ):
+            response, result = await self.inspect("clip.mp4")
+        self.assertEqual(response.status, 200)
+        self.assertEqual(spawn.call_count, 2)
+        first, second = (call.args[0] for call in spawn.call_args_list)
+        self.assertEqual(
+            first[first.index("-show_entries") + 1],
+            "format_tags=prompt,workflow,parameters,"
+            "comfyassets_prompt_pairs,comment,description",
+        )
+        self.assertEqual(
+            second[second.index("-show_entries") + 1],
+            "format_tags=prompt,parameters,comfyassets_prompt_pairs,"
+            "comment,description",
+        )
+        self.assertEqual(processes[0].stdout.requested_size, VIDEO_PROBE_BYTES + 1)
+        self.assertEqual(processes[1].stdout.requested_size, VIDEO_REPROBE_BYTES + 1)
+        for process in processes:
+            self.assertTrue(process.stdout.closed)
+            process.wait.assert_called()
+        self.assertEqual(result["prompts"]["positive"], self.side("bounded portrait"))
+        self.assertEqual(result["metadata"]["status"], "available")
+        self.assertEqual(
+            result["metadata"]["record_formats"],
+            {"prompt": "original-text", "workflow": "limited"},
+        )
+        self.assertNotIn("workflow", result["metadata"]["raw"])
+        self.assertEqual(result["metadata"]["workflow_graph"]["status"], "limited")
+        # The dropped per-key workflow belonged to an unselected carrier when
+        # an envelope wins the re-probe: that envelope is whole, not limited.
+        envelope = json.dumps({"prompt": json.loads(source), "workflow": {"v": 2}})
+        processes = [
+            probe_process(b"x" * (VIDEO_PROBE_BYTES + 1)),
+            probe_process(probe_payload({"comment": envelope})),
+        ]
+        with (
+            patch("utils.prompt_inspector.ffprobe_path", return_value="/tool/ffprobe"),
+            patch("utils.prompt_inspector.subprocess.Popen", side_effect=processes),
+        ):
+            _, result = await self.inspect("clip.mp4")
+        self.assertEqual(result["prompts"]["positive"], self.side("bounded portrait"))
+        self.assertEqual(result["metadata"]["workflow_graph"]["value"], {"v": 2})
+        self.assertEqual(
+            result["metadata"]["record_formats"]["workflow"], "reserialized-json"
+        )
+        # A workflow-only per-key video over budget is limited, not absent.
+        processes = [
+            probe_process(b"x" * (VIDEO_PROBE_BYTES + 1)),
+            probe_process(probe_payload({})),
+        ]
+        with (
+            patch("utils.prompt_inspector.ffprobe_path", return_value="/tool/ffprobe"),
+            patch("utils.prompt_inspector.subprocess.Popen", side_effect=processes),
+        ):
+            _, result = await self.inspect("clip.mp4")
+        self.assertEqual(result["metadata"]["status"], "available")
+        self.assertEqual(result["metadata"]["record_formats"], {"workflow": "limited"})
+        self.assertEqual(result["metadata"]["workflow_graph"]["status"], "limited")
+        self.assertEqual(result["prompts"]["positive"]["status"], "absent")
+
     async def test_video_role_description_wins_over_workflow_only_comment(self):
         evidence = {}
         source = graph("chosen", "")
@@ -1444,3 +1642,86 @@ class TestPromptInspector(ImageAPITestCase):
                     lookup.assert_called_once_with(
                         str(path), False, report_ambiguity=True
                     )
+
+
+class TestInspectorImageTags(ImageAPITestCase):
+    """B3: the inspector envelope's ``image_tags`` facet (spec amendment S1)."""
+
+    inspect = TestPromptInspector.inspect
+
+    def tag(self, path, tags):
+        import hashlib
+        import os
+
+        stat = path.stat()
+        return self.api.db.replace_image_tags(
+            locator=os.path.normcase(os.path.realpath(path)),
+            sha256=hashlib.sha256(path.read_bytes()).hexdigest(),
+            size=stat.st_size,
+            mtime_ns=stat.st_mtime_ns,
+            expected_epoch=self.api.db.get_execution_epoch(),
+            expected_tag_revision=0,
+            tags=tags,
+        )
+
+    async def test_inspector_reports_current_none_and_unavailable_image_tags(self):
+        import os
+
+        tagged = png(self.output_dir / "sub" / "tagged.png", prompt=graph())
+        changed = png(self.output_dir / "changed.png", prompt=graph())
+        png(self.output_dir / "untagged.png", prompt=graph())
+        empty = png(self.output_dir / "empty.png", prompt=graph())
+        self.tag(tagged, ["dunes", "warm_light"])
+        self.tag(changed, ["dunes"])
+        self.tag(empty, [])
+        stat = changed.stat()
+        os.utime(changed, ns=(stat.st_atime_ns, stat.st_mtime_ns + 1_000_000))
+
+        response, envelope = await self.inspect("sub/tagged.png")
+        self.assertEqual(response.status, 200, envelope)
+        self.assertEqual(
+            envelope.get("image_tags"),
+            {"state": "current", "tags": ["dunes", "warm_light"]},
+        )
+        # The tags never change the Library facts of the prompt link.
+        self.assertEqual(envelope["library"]["tags"], [])
+        _, changed_body = await self.inspect("changed.png")
+        self.assertEqual(changed_body.get("image_tags"), {"state": "none", "tags": []})
+        _, untagged = await self.inspect("untagged.png")
+        self.assertEqual(untagged.get("image_tags"), {"state": "none", "tags": []})
+        _, saved_empty = await self.inspect("empty.png")
+        self.assertEqual(
+            saved_empty.get("image_tags"), {"state": "current", "tags": []}
+        )
+        with patch.object(
+            self.api.db, "current_image_tags", side_effect=RuntimeError("boom")
+        ):
+            with self.assertLogs("prompt_manager", level="WARNING") as logged:
+                response, failed = await self.inspect("sub/tagged.png")
+        self.assertTrue(
+            any(r.exc_info and "image tags" in r.getMessage() for r in logged.records),
+            logged.output,
+        )
+        self.assertEqual(response.status, 200, failed)
+        self.assertEqual(failed.get("image_tags"), {"state": "unavailable", "tags": []})
+        self.assertEqual(failed["prompts"], envelope["prompts"])
+        self.assertNotIn(str(self.comfy_dir), json.dumps(envelope))
+
+    async def test_inspector_image_tags_use_the_inspection_stable_stat(self):
+        tagged = png(self.output_dir / "tagged.png", prompt=graph())
+        self.tag(tagged, ["dunes"])
+        stat = tagged.stat()
+        seen = []
+        real = self.api.db.current_image_tags
+
+        def spy(stats):
+            seen.append(dict(stats))
+            return real(stats)
+
+        with patch.object(self.api.db, "current_image_tags", side_effect=spy):
+            _, body = await self.inspect("tagged.png")
+        self.assertEqual(
+            body.get("image_tags"), {"state": "current", "tags": ["dunes"]}
+        )
+        self.assertEqual(list(seen[0].values()), [(stat.st_size, stat.st_mtime_ns)])
+        self.assertEqual(len(seen), 1)