diff --git a/.github/workflows/tier0.yml b/.github/workflows/tier0.yml
index 75dead2..f380824 100644
--- a/.github/workflows/tier0.yml
+++ b/.github/workflows/tier0.yml
@@ -23,8 +23,8 @@ jobs:
       - name: Install Tier 0 deps
         run: |
           python -m pip install --upgrade pip
-          # Pins mirror pyproject (Phase 1 baseline). Tier 0 only needs
-          # torch + numpy + pytest; everything else is Mac-only.
+          # Tier 0 only needs torch + numpy + pytest; everything else
+          # is Mac-only.
           python -m pip install \
             "torch==2.0.1" "numpy<1.25" \
             "pytest>=8" "pytest-xdist"
diff --git a/.github/workflows/tier2.yml b/.github/workflows/tier2.yml
index 0950dc7..98943f4 100644
--- a/.github/workflows/tier2.yml
+++ b/.github/workflows/tier2.yml
@@ -16,14 +16,14 @@ jobs:
       github.event_name == 'workflow_dispatch' ||
       (github.event_name == 'pull_request' &&
        contains(github.event.pull_request.labels.*.name, 'run-m2'))
-    # Self-hosted Mac registered by the maintainer. See docs/ci-m2.md
-    # for runner setup and the COMFY_DIR / checkpoint prerequisites.
+    # Self-hosted Apple Silicon runner. Prerequisites: COMFY_DIR pointing at
+    # a runner-owned ComfyUI clone, plus a cached SD1.5 checkpoint.
     runs-on: [self-hosted, macOS, ARM64, coreml]
     timeout-minutes: 90
     steps:
       - uses: actions/checkout@v4
 
-      # Hybrid ComfyUI strategy (see docs/ci-m2.md "ComfyUI version under test"):
+      # Hybrid ComfyUI strategy:
       #   - schedule (nightly)  -> latest origin/master + ComfyUI's own
       #     requirements.txt (constrained). Canary for upstream API breakage.
       #   - PR label / dispatch -> the pinned requires-comfyui SHA + the frozen
@@ -46,7 +46,7 @@ jobs:
         # git reset --hard and rewrites custom_nodes). Cloned on first run.
         run: |
           set -euo pipefail
-          if [ -z "${COMFY_DIR:-}" ]; then echo "COMFY_DIR unset (see docs/ci-m2.md)"; exit 1; fi
+          if [ -z "${COMFY_DIR:-}" ]; then echo "COMFY_DIR unset"; exit 1; fi
           # Init-in-place rather than `git clone`: COMFY_DIR may already hold the
           # cached checkpoint (models/checkpoints) or converted .mlmodelc, and
           # `git clone` refuses a non-empty target. init + fetch + `checkout -f`
@@ -95,22 +95,6 @@ jobs:
             uv sync --group comfy
           fi
 
-      - name: Convert UNet variants if missing
-        # Idempotent: converter.convert / compile_model skip when the .mlmodelc
-        # already exists, so this is a no-op once the runner cache is warm.
-        run: |
-          set -euo pipefail
-          for V in none 8 6 4; do
-            if [ "$V" = "none" ]; then SUF=""; else SUF="_q$V"; fi
-            OUT="$COMFY_DIR/models/unet/v1-5-pruned-emaonly_1x512x512_se${SUF}_unet.mlmodelc"
-            if [ -d "$OUT" ]; then echo "skip nbits=$V (exists)"; continue; fi
-            echo "converting nbits=$V"
-            # --no-sync: keep the deps we just installed (latest mode's
-            # ComfyUI requirements / pinned mode's --group comfy); a bare
-            # `uv run` would re-sync to the lock and drop them.
-            QUANT_NBITS="$V" uv run --no-sync python bench/scripts/convert_sd15.py
-          done
-
       - name: Start ComfyUI server (background)
         run: |
           cd "$COMFY_DIR"
@@ -126,29 +110,11 @@ jobs:
           echo "comfy failed to start"; tail -100 /tmp/comfyui-ci.log; exit 1
 
       - name: Run Tier 2 (m2 marker)
+        # The golden-image workflow drives the Core ML Converter node, so the
+        # UNet is converted on demand on the first run and reused from the
+        # runner-local cache afterwards.
         run: uv run --no-sync pytest -m m2 tests/ -v
 
-      - name: Run bench harness (baseline UNet)
-        run: |
-          uv run --no-sync python bench/run.py \
-            --model "$COMFY_DIR/models/unet/v1-5-pruned-emaonly_1x512x512_se_unet.mlmodelc" \
-            --compute-units CPU_AND_NE CPU_AND_GPU \
-            --repeats 30 --assumed-steps 20
-
-      - name: Run quantization tradeoff matrix (Phase 6)
-        # Benches none/8/6/4 on ANE and records size / ms-step / PSNR-vs-none
-        # into bench/results/quant_matrix_<sha>.{json,md}.
-        run: uv run --no-sync python bench/scripts/quant_matrix.py
-
-      - name: Upload bench results
-        uses: actions/upload-artifact@v4
-        if: always()
-        with:
-          name: bench-results
-          path: |
-            bench/results/*.json
-            bench/results/*.md
-
       - name: Stop ComfyUI server
         if: always()
         run: pkill -f "main.py.*8188" || true
diff --git a/.gitignore b/.gitignore
index b578bb9..b883889 100644
--- a/.gitignore
+++ b/.gitignore
@@ -8,5 +8,4 @@ comfy_env/
 coremlsuite-venv/
 experiment_results/
 test_results/
-bench/scripts/*.log
 tests/m2/_latest_generated.png
diff --git a/Makefile b/Makefile
deleted file mode 100644
index c1d51d8..0000000
--- a/Makefile
+++ /dev/null
@@ -1,88 +0,0 @@
-# ComfyUI-CoreMLSuite — tiered test/bench dispatcher (Phase 4).
-#
-# Tiers (see MODERNIZATION_SPEC.md):
-#   Tier 0 (unit):  framework-free pure-logic tests, run anywhere in seconds.
-#   Tier 1 (smoke): macOS-ARM, no ANE, no full model — converts a synthetic
-#                   micro-UNet to catch coremltools / ml-stable-diffusion
-#                   API breakage in minutes.
-#   Tier 2 (m2):    real ANE on Apple Silicon; integration + bench.
-#
-# COMFY_DIR defaults to the canonical custom-node layout (two dirs up from here).
-# PY defaults to the project's uv-managed venv interpreter.
-
-COMFY_DIR ?= $(realpath $(CURDIR)/../..)
-PY ?= $(CURDIR)/.venv/bin/python
-PYTEST ?= $(PY) -m pytest
-
-UNAME_S := $(shell uname -s)
-UNAME_M := $(shell uname -m)
-IS_MACOS_ARM := $(filter Darwin,$(UNAME_S))$(filter arm64,$(UNAME_M))
-
-.PHONY: help test-unit test-smoke test-m2 bench bench-rerun ci-tier0 ci-tier1 clean check-macos-arm
-
-help:
-	@echo "ComfyUI-CoreMLSuite — make targets"
-	@echo ""
-	@echo "  test-unit       Tier 0: pure-logic pytest, runs anywhere, seconds"
-	@echo "  test-smoke      Tier 1: synthetic micro-UNet ct.convert + load (macOS-ARM, minutes)"
-	@echo "  test-m2         Tier 2: pytest -m m2 against a real Core ML UNet (Apple Silicon + ANE)"
-	@echo "  bench           Run bench/run.py against a converted .mlmodelc"
-	@echo ""
-	@echo "Vars: COMFY_DIR (default: $(COMFY_DIR)), PY (default: $(PY))"
-
-check-macos-arm:
-	@if [ -z "$(IS_MACOS_ARM)" ]; then \
-		echo "this target requires macOS on Apple Silicon (got $(UNAME_S)/$(UNAME_M))"; \
-		exit 2; \
-	fi
-
-# Tier 0 — Linux-safe pure logic. Should not import comfy/coremltools.
-test-unit:
-	$(PYTEST) -m unit tests/
-
-# Tier 1 — macOS-ARM smoke. Converts a synthetic UNet through coremltools to
-# catch API breakage without needing a real SD checkpoint or the ANE.
-test-smoke: check-macos-arm
-	$(PYTEST) -m smoke tests/
-
-# Tier 2 — full Apple Silicon path: integration + m2 golden + bench.
-# Requires a converted .mlmodelc (see bench/scripts/convert_sd15.py).
-test-m2: check-macos-arm
-	$(PYTEST) -m m2 tests/
-
-# Bench harness. Override MODEL=/path/to/.mlmodelc for an explicit model.
-MODEL ?= $(COMFY_DIR)/models/unet/v1-5-pruned-emaonly_1x512x512_se_unet.mlmodelc
-COMPUTE_UNITS ?= CPU_AND_NE CPU_AND_GPU
-REPEATS ?= 30
-ASSUMED_STEPS ?= 20
-bench: check-macos-arm
-	$(PY) bench/run.py \
-		--model "$(MODEL)" \
-		--compute-units $(COMPUTE_UNITS) \
-		--repeats $(REPEATS) \
-		--assumed-steps $(ASSUMED_STEPS)
-
-# Phase 6: quantization tradeoff matrix. Expects the {none, 8, 6, 4}
-# variants to already exist on disk (run `make convert-quant` first or
-# QUANT_NBITS=8 .venv/bin/python bench/scripts/convert_sd15.py).
-bench-quant: check-macos-arm
-	$(PY) bench/scripts/quant_matrix.py
-
-convert-quant: check-macos-arm
-	@for n in 8 6 4; do \
-		echo "=== converting nbits=$$n ==="; \
-		QUANT_NBITS=$$n $(PY) bench/scripts/convert_sd15.py || exit 1; \
-	done
-
-# What CI actually invokes — same as test-unit but echoes the env capture
-# alongside so failed runs land with diagnostics.
-ci-tier0:
-	@echo "## env (Tier 0)" && $(PY) --version && uv pip freeze --python "$(PY)" 2>/dev/null | head -50 || true
-	$(MAKE) test-unit
-
-ci-tier1: check-macos-arm
-	@echo "## env (Tier 1)" && $(PY) --version && uv pip freeze --python "$(PY)" 2>/dev/null | head -50 || true
-	$(MAKE) test-smoke
-
-clean:
-	rm -rf .pytest_cache tests/m2/_latest_generated.png pytestdebug.log
diff --git a/README.md b/README.md
index 7571fa1..1c48973 100644
--- a/README.md
+++ b/README.md
@@ -370,21 +370,21 @@ The models used in this workflow are available at the following links:
 
 ![sdxl](./assets/sdxl_conversion.png?raw=true)
 
-## Quantization (Phase 6, opt-in)
+## Quantization (opt-in)
 
 The `Core ML Converter` and `Core ML LCM Converter` nodes accept an
 optional `quantize_nbits` dropdown that runs k-means weight palettization
 (`coremltools.optimize.coreml.palettize_weights`) on the UNet before save.
 
-Values: `none` (default — no quantization, identical to pre-Phase-6
+Values: `none` (default — no quantization, identical to unquantized
 behavior and filenames), `8`, `6`, `4`. The number is appended to the
 .mlpackage stem as `_q<bits>` so quantized and unquantized variants
 coexist on disk and in cache.
 
 ### SD1.5 1×512×512 SPLIT_EINSUM tradeoffs (M2 Pro, ANE)
 
-Measured by `bench/scripts/quant_matrix.py` (20 forward passes, fixed
-seed for the PSNR comparison):
+Measured with 20 UNet forward passes at a fixed seed for the PSNR
+comparison:
 
 | nbits | size (MB) | size vs none | fwd median (ms) | PSNR vs `none` (dB) |
 |---|---:|---:|---:|---:|
@@ -408,7 +408,7 @@ PSNR is comfortably higher (the sampler averages over 20 steps).
   margin, `none` if you want bit-identical output for golden testing.
 
 The default stays `none` so existing workflows produce byte-for-byte
-identical output — the Phase 6 anchor (`tests/m2/test_golden_image.py`)
+identical output — the golden-image anchor (`tests/m2/test_golden_image.py`)
 verifies this on every Tier 2 run.
 
 ## Limitations
diff --git a/bench/env/baseline-1e5791d.txt b/bench/env/baseline-1e5791d.txt
deleted file mode 100644
index ad2819d..0000000
--- a/bench/env/baseline-1e5791d.txt
+++ /dev/null
@@ -1,134 +0,0 @@
-# Phase 1 baseline environment capture
-timestamp_utc: 2026-05-24T14:36:51Z
-repo_sha: 1e5791d1087b4dfb095e7761f05ac4e82c92cf7a
-repo_branch: modernize/phase5-toolchain-bump
-comfyui_dir: /Users/aszczepanski/dev/ComfyUI
-comfyui_sha: ab5413351eee61f3d7f10c74e75286df0058bb18
-python_bin: /Users/aszczepanski/dev/ComfyUI/custom_nodes/ComfyUI-CoreMLSuite/.venv/bin/python
-python_version: Python 3.12.11
-platform_machine: arm64
-platform_uname: Darwin mac-2.home 25.1.0 Darwin Kernel Version 25.1.0: Mon Oct 20 19:33:00 PDT 2025; root:xnu-12377.41.6~2/RELEASE_ARM64_T6020 arm64
-macos_product: macOS
-macos_version: 26.1
-macos_build: 25B78
-
-## python-coreml-stable-diffusion (resolved)
-Name: python-coreml-stable-diffusion
-Version: 1.1.0
-Location: /Users/aszczepanski/dev/ComfyUI/custom_nodes/ComfyUI-CoreMLSuite/.venv/lib/python3.12/site-packages
-Requires: coremltools, diffusers, diffusionkit, huggingface-hub, invisible-watermark, matplotlib, numpy, pytest, safetensors, scikit-learn, scipy, torch, transformers
-Required-by:
-
-## python-coreml-stable-diffusion installed git metadata
-{"url":"https://github.com/apple/ml-stable-diffusion.git","vcs_info":{"vcs":"git","commit_id":"e5d960c41a6a4ab200b8db379194127607b1c590","requested_revision":"e5d960c41a6a4ab200b8db379194127607b1c590"}}
-
-## coremltools version
-scikit-learn version 1.7.2 is not supported. Minimum required version: 0.17. Maximum required version: 1.5.1. Disabling scikit-learn conversion API.
-Torch version 2.7.1 has not been tested with coremltools. You may run into unexpected errors. Torch 2.7.0 is the most recent version that has been tested.
-coremltools 9.0
-
-## torch version
-torch 2.7.1
-
-## numpy version
-numpy 1.26.4
-
-## freeze
-accelerate==1.13.0
-aiohappyeyeballs==2.6.2
-aiohttp==3.13.5
-aiosignal==1.4.0
-annotated-types==0.7.0
-antlr4-python3-runtime==4.9.3
-argmaxtools==0.1.23
-attrs==26.1.0
-beartype==0.22.9
-cattrs==26.1.0
-certifi==2026.5.20
-cffi==2.0.0
-charset-normalizer==3.4.7
-click==8.4.1
-comfyui-frontend-package==1.14.6
-contourpy==1.3.2
-coremltools==9.0
-cycler==0.12.1
-diffusers==0.30.2
-diffusionkit==0.4.0
-einops==0.8.2
-filelock==3.29.0
-fonttools==4.63.0
-frozenlist==1.8.0
-fsspec==2026.4.0
-gitdb==4.0.12
-gitpython==3.1.50
-huggingface-hub==0.24.6
-idna==3.16
-importlib-metadata==9.0.0
-iniconfig==2.3.0
-invisible-watermark==0.2.0
-jaxtyping==0.3.9
-jinja2==3.1.6
-joblib==1.5.3
-kiwisolver==1.5.0
-kornia==0.8.3
-kornia-rs==0.1.14
-markupsafe==3.0.3
-matplotlib==3.10.9
-mlx==0.31.2
-mlx-metal==0.31.2
-mpmath==1.3.0
-multidict==6.7.1
-networkx==3.6.1
-numpy==1.26.4
-omegaconf==2.3.0
-opencv-python==4.11.0.86
-overrides==7.7.0
-packaging==26.2
-peft==0.13.2
-pillow==12.2.0
-platformdirs==4.9.6
-pluggy==1.6.0
-propcache==0.5.2
-protobuf==7.35.0
-psutil==7.2.2
-pyaml==26.2.1
-pycparser==3.0
-pydantic==2.13.4
-pydantic-core==2.46.4
-pygments==2.20.0
-pyparsing==3.3.2
-pytest==9.0.3
-python-coreml-stable-diffusion @ git+https://github.com/apple/ml-stable-diffusion.git@e5d960c41a6a4ab200b8db379194127607b1c590
-python-dateutil==2.9.0.post0
-pywavelets==1.8.0
-pyyaml==6.0.3
-regex==2026.5.9
-requests==2.34.2
-safetensors==0.7.0
-scikit-learn==1.7.2
-scipy==1.15.3
-sentencepiece==0.2.1
-sentry-sdk==2.60.0
-setuptools==81.0.0
-six==1.17.0
-smmap==5.0.3
-soundfile==0.13.1
-spandrel==0.4.2
-sympy==1.14.0
-tabulate==0.10.0
-threadpoolctl==3.6.0
-tokenizers==0.19.1
-torch==2.7.1
-torchaudio==2.11.0
-torchsde==0.2.6
-torchvision==0.22.1
-tqdm==4.67.3
-trampoline==0.1.2
-transformers==4.44.2
-typing-extensions==4.15.0
-typing-inspection==0.4.2
-urllib3==2.7.0
-wadler-lindig==0.1.7
-wandb==0.27.0
-yarl==1.24.2
-zipp==4.1.0
diff --git a/bench/env/baseline-ef2a18c.txt b/bench/env/baseline-ef2a18c.txt
deleted file mode 100644
index f8bc2b9..0000000
--- a/bench/env/baseline-ef2a18c.txt
+++ /dev/null
@@ -1,129 +0,0 @@
-# Phase 1 baseline environment capture
-timestamp_utc: 2026-05-22T13:06:37Z
-repo_sha: ef2a18cff3d8b563ae053d27c1c5c233bd05d47f
-repo_branch: modernize/phase1-freeze-baseline
-comfyui_dir: /Users/aszczepanski/dev/ComfyUI
-comfyui_sha: ab5413351eee61f3d7f10c74e75286df0058bb18
-python_bin: /Users/aszczepanski/dev/ComfyUI/custom_nodes/ComfyUI-CoreMLSuite/.venv/bin/python
-python_version: Python 3.11.14
-platform_machine: arm64
-platform_uname: Darwin mac-2.home 25.1.0 Darwin Kernel Version 25.1.0: Mon Oct 20 19:33:00 PDT 2025; root:xnu-12377.41.6~2/RELEASE_ARM64_T6020 arm64
-macos_product: macOS
-macos_version: 26.1
-macos_build: 25B78
-
-## python-coreml-stable-diffusion (resolved)
-Name: python-coreml-stable-diffusion
-Version: 1.1.0
-Location: /Users/aszczepanski/dev/ComfyUI/custom_nodes/ComfyUI-CoreMLSuite/.venv/lib/python3.11/site-packages
-Requires: coremltools, diffusers, diffusionkit, huggingface-hub, invisible-watermark, matplotlib, numpy, pytest, safetensors, scikit-learn, scipy, torch, transformers
-Required-by:
-
-## python-coreml-stable-diffusion installed git metadata
-{"url":"https://github.com/apple/ml-stable-diffusion.git","vcs_info":{"vcs":"git","commit_id":"e5d960c41a6a4ab200b8db379194127607b1c590","requested_revision":"e5d960c41a6a4ab200b8db379194127607b1c590"}}
-
-## coremltools version
-scikit-learn version 1.7.2 is not supported. Minimum required version: 0.17. Maximum required version: 1.5.1. Disabling scikit-learn conversion API.
-Minimum required torch version for importing coremltools.optimize.torch is 2.1.0. Got torch version 2.0.1.
-coremltools 8.2
-
-## torch version
-torch 2.0.1
-
-## numpy version
-numpy 1.23.5
-
-## freeze
-accelerate==1.13.0
-aiohappyeyeballs==2.6.2
-aiohttp==3.13.5
-aiosignal==1.4.0
-annotated-types==0.7.0
-antlr4-python3-runtime==4.9.3
-argmaxtools==0.1.23
-attrs==26.1.0
-beartype==0.22.9
-cattrs==26.1.0
-certifi==2026.5.20
-charset-normalizer==3.4.7
-click==8.4.1
-comfyui-frontend-package==1.14.6
-contourpy==1.3.2
-coremltools==8.2
-cycler==0.12.1
-diffusers==0.30.2
-diffusionkit==0.4.0
-einops==0.8.2
-filelock==3.29.0
-fonttools==4.63.0
-frozenlist==1.8.0
-fsspec==2026.4.0
-gitdb==4.0.12
-gitpython==3.1.50
-huggingface-hub==0.24.6
-idna==3.16
-importlib-metadata==9.0.0
-iniconfig==2.3.0
-invisible-watermark==0.2.0
-jaxtyping==0.3.9
-jinja2==3.1.6
-joblib==1.5.3
-kiwisolver==1.5.0
-kornia==0.8.3
-kornia-rs==0.1.14
-markupsafe==3.0.3
-matplotlib==3.10.9
-mlx==0.31.2
-mlx-metal==0.31.2
-mpmath==1.3.0
-multidict==6.7.1
-networkx==3.6.1
-numpy==1.23.5
-omegaconf==2.3.0
-opencv-python==4.11.0.86
-overrides==7.7.0
-packaging==26.2
-peft==0.13.2
-pillow==12.2.0
-platformdirs==4.9.6
-pluggy==1.6.0
-propcache==0.5.2
-protobuf==7.35.0
-psutil==7.2.2
-pyaml==26.2.1
-pydantic==2.13.4
-pydantic-core==2.46.4
-pygments==2.20.0
-pyparsing==3.3.2
-pytest==9.0.3
-python-coreml-stable-diffusion @ git+https://github.com/apple/ml-stable-diffusion.git@e5d960c41a6a4ab200b8db379194127607b1c590
-python-dateutil==2.9.0.post0
-pywavelets==1.8.0
-pyyaml==6.0.3
-regex==2026.5.9
-requests==2.34.2
-safetensors==0.7.0
-scikit-learn==1.7.2
-scipy==1.15.3
-sentencepiece==0.2.1
-sentry-sdk==2.60.0
-setuptools==81.0.0
-six==1.17.0
-smmap==5.0.3
-sympy==1.14.0
-tabulate==0.10.0
-threadpoolctl==3.6.0
-tokenizers==0.19.1
-torch==2.0.1
-torchsde==0.2.6
-torchvision==0.15.2
-tqdm==4.67.3
-trampoline==0.1.2
-transformers==4.44.2
-typing-extensions==4.15.0
-typing-inspection==0.4.2
-urllib3==2.7.0
-wadler-lindig==0.1.7
-wandb==0.27.0
-yarl==1.24.2
-zipp==4.1.0
diff --git a/bench/env/capture.sh b/bench/env/capture.sh
deleted file mode 100755
index aec3d9b..0000000
--- a/bench/env/capture.sh
+++ /dev/null
@@ -1,100 +0,0 @@
-#!/usr/bin/env bash
-# Phase 1 baseline environment capture.
-#
-# Default mode: query the project's .venv via `uv pip` (matches how uv builds
-# the venv — without a bundled pip executable). For a non-uv setup, point
-# PYTHON_BIN at the right interpreter; pip freeze then falls back to
-# `python -m pip` if available, otherwise `uv pip`.
-#
-# Usage:
-#   bash bench/env/capture.sh                 # uses .venv (uv-managed)
-#   PYTHON_BIN=apple_env/bin/python bash bench/env/capture.sh
-#
-# Writes bench/env/baseline-<gitsha>.txt with: this repo's git sha,
-# ComfyUI's git sha (override path with COMFY_DIR=...), python + macOS
-# versions, the resolved git commit of python-coreml-stable-diffusion as
-# installed, and the full freeze of the active venv (for fresh-venv
-# reproducibility).
-set -euo pipefail
-
-REPO_ROOT="$(git rev-parse --show-toplevel)"
-cd "$REPO_ROOT"
-
-SHA_SHORT="$(git rev-parse --short HEAD)"
-OUT_DIR="bench/env"
-OUT="${OUT_DIR}/baseline-${SHA_SHORT}.txt"
-mkdir -p "$OUT_DIR"
-
-COMFY_DIR="${COMFY_DIR:-$(cd ../.. && pwd)}"
-PYTHON_BIN="${PYTHON_BIN:-$REPO_ROOT/.venv/bin/python}"
-
-if [ ! -x "$PYTHON_BIN" ]; then
-  echo "PYTHON_BIN not executable: $PYTHON_BIN" >&2
-  exit 1
-fi
-
-# Prefer `python -m pip` (works in classic venvs). Fall back to `uv pip`,
-# which queries any interpreter without needing pip installed in it.
-run_freeze() {
-  if "$PYTHON_BIN" -m pip --version >/dev/null 2>&1; then
-    "$PYTHON_BIN" -m pip freeze
-  elif command -v uv >/dev/null 2>&1; then
-    uv pip freeze --python "$PYTHON_BIN"
-  else
-    echo "neither pip nor uv available for freeze"
-  fi
-}
-
-run_show_msd() {
-  if "$PYTHON_BIN" -m pip --version >/dev/null 2>&1; then
-    "$PYTHON_BIN" -m pip show python-coreml-stable-diffusion 2>/dev/null || echo "not-installed"
-  elif command -v uv >/dev/null 2>&1; then
-    uv pip show --python "$PYTHON_BIN" python-coreml-stable-diffusion 2>/dev/null || echo "not-installed"
-  else
-    echo "neither pip nor uv available for show"
-  fi
-}
-
-{
-  echo "# Phase 1 baseline environment capture"
-  echo "timestamp_utc: $(date -u +%Y-%m-%dT%H:%M:%SZ)"
-  echo "repo_sha: $(git rev-parse HEAD)"
-  echo "repo_branch: $(git rev-parse --abbrev-ref HEAD)"
-  echo "comfyui_dir: ${COMFY_DIR}"
-  echo "comfyui_sha: $(git -C "$COMFY_DIR" rev-parse HEAD 2>/dev/null || echo unknown)"
-  echo "python_bin: ${PYTHON_BIN}"
-  echo "python_version: $("$PYTHON_BIN" --version 2>&1)"
-  echo "platform_machine: $(uname -m)"
-  echo "platform_uname: $(uname -a)"
-  if command -v sw_vers >/dev/null 2>&1; then
-    echo "macos_product: $(sw_vers -productName)"
-    echo "macos_version: $(sw_vers -productVersion)"
-    echo "macos_build: $(sw_vers -buildVersion)"
-  fi
-  echo
-  echo "## python-coreml-stable-diffusion (resolved)"
-  run_show_msd
-  echo
-  echo "## python-coreml-stable-diffusion installed git metadata"
-  DIST_INFO="$("$PYTHON_BIN" -c "import importlib.metadata as m; d=m.distribution('python-coreml-stable-diffusion'); print(d._path)" 2>/dev/null || true)"
-  if [ -n "${DIST_INFO}" ] && [ -f "${DIST_INFO}/direct_url.json" ]; then
-    cat "${DIST_INFO}/direct_url.json"
-    echo
-  else
-    echo "direct_url.json not found"
-  fi
-  echo
-  echo "## coremltools version"
-  "$PYTHON_BIN" -c "import coremltools; print('coremltools', coremltools.__version__)" 2>&1 || true
-  echo
-  echo "## torch version"
-  "$PYTHON_BIN" -c "import torch; print('torch', torch.__version__)" 2>&1 || true
-  echo
-  echo "## numpy version"
-  "$PYTHON_BIN" -c "import numpy; print('numpy', numpy.__version__)" 2>&1 || true
-  echo
-  echo "## freeze"
-  run_freeze
-} > "$OUT"
-
-echo "wrote $OUT"
diff --git a/bench/env/pytest-unit-1e5791d.txt b/bench/env/pytest-unit-1e5791d.txt
deleted file mode 100644
index f0e81e6..0000000
--- a/bench/env/pytest-unit-1e5791d.txt
+++ /dev/null
@@ -1,98 +0,0 @@
-============================= test session starts ==============================
-platform darwin -- Python 3.12.11, pytest-9.0.3, pluggy-1.6.0 -- /Users/aszczepanski/dev/ComfyUI/custom_nodes/ComfyUI-CoreMLSuite/.venv/bin/python
-cachedir: .pytest_cache
-rootdir: /Users/aszczepanski/dev/ComfyUI/custom_nodes/ComfyUI-CoreMLSuite
-configfile: pyproject.toml
-plugins: jaxtyping-0.3.9
-collecting ... collected 88 items
-
-tests/unit/test_characterization_controlnet.py::test_expand_inputs_doubles_singleton_numpy PASSED [  1%]
-tests/unit/test_characterization_controlnet.py::test_expand_inputs_doubles_singleton_torch PASSED [  2%]
-tests/unit/test_characterization_controlnet.py::test_expand_inputs_doubles_singleton_list PASSED [  3%]
-tests/unit/test_characterization_controlnet.py::test_expand_inputs_skips_already_batched PASSED [  4%]
-tests/unit/test_characterization_controlnet.py::test_expand_inputs_preserves_unknown_value_types PASSED [  5%]
-tests/unit/test_characterization_controlnet.py::test_no_control_returns_zero_fp16_for_residuals PASSED [  6%]
-tests/unit/test_characterization_controlnet.py::test_no_control_returns_empty_when_no_residuals PASSED [  7%]
-tests/unit/test_characterization_controlnet.py::test_extract_residual_kwargs_empty_when_model_has_no_residual_inputs PASSED [  9%]
-tests/unit/test_characterization_controlnet.py::test_extract_residual_kwargs_none_control_returns_no_control_shapes PASSED [ 10%]
-tests/unit/test_characterization_controlnet.py::test_extract_residual_kwargs_flattens_output_then_middle_and_casts_fp16 PASSED [ 11%]
-tests/unit/test_characterization_controlnet.py::test_chunk_control_none_returns_list_of_nones_with_length_target PASSED [ 12%]
-tests/unit/test_characterization_controlnet.py::test_chunk_control_shapes_after_chunking[1-2-1] PASSED [ 13%]
-tests/unit/test_characterization_controlnet.py::test_chunk_control_shapes_after_chunking[2-2-1] PASSED [ 14%]
-tests/unit/test_characterization_controlnet.py::test_chunk_control_shapes_after_chunking[3-2-2] PASSED [ 15%]
-tests/unit/test_characterization_controlnet.py::test_chunk_control_shapes_after_chunking[4-2-2] PASSED [ 17%]
-tests/unit/test_characterization_controlnet.py::test_chunk_control_shapes_after_chunking[5-3-2] PASSED [ 18%]
-tests/unit/test_characterization_controlnet.py::test_chunk_control_shapes_after_chunking[9-4-3] PASSED [ 19%]
-tests/unit/test_characterization_controlnet.py::test_chunk_control_preserves_keys_order PASSED [ 20%]
-tests/unit/test_characterization_controlnet.py::test_chunk_control_zero_pads_remainder PASSED [ 21%]
-tests/unit/test_characterization_inputs.py::test_coreml_kwargs_sd15_shapes_and_fp16 PASSED [ 22%]
-tests/unit/test_characterization_inputs.py::test_coreml_kwargs_sd15_with_controlnet_emits_residuals PASSED [ 23%]
-tests/unit/test_characterization_inputs.py::test_coreml_kwargs_sd15_without_controlnet_zero_fills_residuals PASSED [ 25%]
-tests/unit/test_characterization_inputs.py::test_coreml_kwargs_lcm_adds_timestep_cond PASSED [ 26%]
-tests/unit/test_characterization_inputs.py::test_coreml_kwargs_lcm_skips_timestep_cond_when_not_provided PASSED [ 27%]
-tests/unit/test_characterization_inputs.py::test_coreml_kwargs_sdxl_base_emits_time_ids_and_text_embeds PASSED [ 28%]
-tests/unit/test_characterization_inputs.py::test_coreml_kwargs_sdxl_refiner_uses_len5_time_ids PASSED [ 29%]
-tests/unit/test_characterization_inputs.py::test_chunks_sd15_pad_to_batch2_returns_one_chunk PASSED [ 30%]
-tests/unit/test_characterization_inputs.py::test_chunks_sd15_with_controlnet_chunks_residuals_too PASSED [ 31%]
-tests/unit/test_characterization_inputs.py::test_chunks_lcm_carries_timestep_cond_per_chunk PASSED [ 32%]
-tests/unit/test_characterization_inputs.py::test_chunks_sdxl_base_propagates_time_ids_and_text_embeds PASSED [ 34%]
-tests/unit/test_characterization_inputs.py::test_chunks_sdxl_refiner_uses_len5_time_ids PASSED [ 35%]
-tests/unit/test_characterization_inputs.py::test_chunks_sdxl_synthesizes_zero_time_ids_when_caller_omits PASSED [ 36%]
-tests/unit/test_characterization_inputs.py::test_chunks_splits_batch_into_multiple_target2_chunks PASSED [ 37%]
-tests/unit/test_characterization_inputs.py::test_chunks_timestep_is_broadcast_from_first_value PASSED [ 38%]
-tests/unit/test_characterization_latents.py::test_chunk_batch_passthrough_when_shape_matches PASSED [ 39%]
-tests/unit/test_characterization_latents.py::test_chunk_batch_pads_single_chunk_when_input_smaller PASSED [ 40%]
-tests/unit/test_characterization_latents.py::test_chunk_batch_splits_exact_multiple PASSED [ 42%]
-tests/unit/test_characterization_latents.py::test_chunk_batch_pads_remainder_chunk PASSED [ 43%]
-tests/unit/test_characterization_latents.py::test_chunk_batch_pad_region_is_zero[1-4-1] PASSED [ 44%]
-tests/unit/test_characterization_latents.py::test_chunk_batch_pad_region_is_zero[3-2-2] PASSED [ 45%]
-tests/unit/test_characterization_latents.py::test_chunk_batch_pad_region_is_zero[5-3-2] PASSED [ 46%]
-tests/unit/test_characterization_latents.py::test_chunk_batch_pad_region_is_zero[9-4-3] PASSED [ 47%]
-tests/unit/test_characterization_latents.py::test_merge_chunks_exact_concat PASSED [ 48%]
-tests/unit/test_characterization_latents.py::test_merge_chunks_truncates_padding PASSED [ 50%]
-tests/unit/test_characterization_latents.py::test_merge_chunks_singleton_returns_equal_copy_when_shape_matches PASSED [ 51%]
-tests/unit/test_characterization_out_name.py::test_attention_suffix[SPLIT_EINSUM-se] PASSED [ 52%]
-tests/unit/test_characterization_out_name.py::test_attention_suffix[SPLIT_EINSUM_V2-se2] PASSED [ 53%]
-tests/unit/test_characterization_out_name.py::test_attention_suffix[ORIGINAL-orig] PASSED [ 54%]
-tests/unit/test_characterization_out_name.py::test_includes_batch_and_size PASSED [ 55%]
-tests/unit/test_characterization_out_name.py::test_appends_cn_suffix_when_controlnet_support_true PASSED [ 56%]
-tests/unit/test_characterization_out_name.py::test_drops_extension_at_first_period PASSED [ 57%]
-tests/unit/test_characterization_out_name.py::test_replaces_spaces_with_underscores PASSED [ 59%]
-tests/unit/test_characterization_out_name.py::test_single_lora PASSED    [ 60%]
-tests/unit/test_characterization_out_name.py::test_multiple_loras_sorted PASSED [ 61%]
-tests/unit/test_characterization_out_name.py::test_lora_plus_controlnet PASSED [ 62%]
-tests/unit/test_characterization_out_name.py::test_sdxl_1024_original_gpu PASSED [ 63%]
-tests/unit/test_characterization_out_name.py::test_lora_names_from_params_sorts_by_name PASSED [ 64%]
-tests/unit/test_characterization_out_name.py::test_lora_names_from_params_empty_list PASSED [ 65%]
-tests/unit/test_characterization_sdxl_options.py::test_build_time_ids_base_defaults PASSED [ 67%]
-tests/unit/test_characterization_sdxl_options.py::test_build_time_ids_base_respects_overrides PASSED [ 68%]
-tests/unit/test_characterization_sdxl_options.py::test_build_time_ids_refiner_defaults PASSED [ 69%]
-tests/unit/test_characterization_sdxl_options.py::test_build_time_ids_refiner_respects_aesthetic_score PASSED [ 70%]
-tests/unit/test_characterization_sdxl_options.py::test_build_time_ids_neither_base_nor_refiner_returns_len4 PASSED [ 71%]
-tests/unit/test_characterization_sdxl_options.py::test_text_embeds_concat_pos_then_neg PASSED [ 72%]
-tests/unit/test_characterization_sdxl_options.py::test_wrapper_captures_time_ids_text_embeds_refiner_via_closure PASSED [ 73%]
-tests/unit/test_characterization_sdxl_options.py::test_wrapper_returns_zero_when_context_missing PASSED [ 75%]
-tests/unit/test_characterization_sdxl_options.py::test_wrapper_refiner_truncates_context_to_g_clip PASSED [ 76%]
-tests/unit/test_chunks.py::test_batch_chunking[1] PASSED                 [ 77%]
-tests/unit/test_chunks.py::test_batch_chunking[2] PASSED                 [ 78%]
-tests/unit/test_chunks.py::test_batch_chunking[4] PASSED                 [ 79%]
-tests/unit/test_chunks.py::test_batch_chunking[5] PASSED                 [ 80%]
-tests/unit/test_chunks.py::test_batch_chunking[9] PASSED                 [ 81%]
-tests/unit/test_chunks.py::test_merge_chunks[1] PASSED                   [ 82%]
-tests/unit/test_chunks.py::test_merge_chunks[2] PASSED                   [ 84%]
-tests/unit/test_chunks.py::test_merge_chunks[4] PASSED                   [ 85%]
-tests/unit/test_chunks.py::test_merge_chunks[5] PASSED                   [ 86%]
-tests/unit/test_chunks.py::test_merge_chunks[9] PASSED                   [ 87%]
-tests/unit/test_chunks.py::test_chunking_controlnet[1-2-1] PASSED        [ 88%]
-tests/unit/test_chunks.py::test_chunking_controlnet[1-1-1] PASSED        [ 89%]
-tests/unit/test_chunks.py::test_chunking_controlnet[2-2-1] PASSED        [ 90%]
-tests/unit/test_chunks.py::test_chunking_controlnet[3-2-2] PASSED        [ 92%]
-tests/unit/test_chunks.py::test_chunking_controlnet[4-2-2] PASSED        [ 93%]
-tests/unit/test_chunks.py::test_chunking_controlnet[5-3-2] PASSED        [ 94%]
-tests/unit/test_chunks.py::test_chunking_controlnet[9-4-3] PASSED        [ 95%]
-tests/unit/test_chunks.py::test_chunking_no_control PASSED               [ 96%]
-tests/unit/test_chunks.py::test_chunking_inputs PASSED                   [ 97%]
-tests/unit/test_controlnet.py::test_no_control PASSED                    [ 98%]
-tests/unit/test_tier0_purity.py::test_no_framework_modules_loaded_by_unit_tier PASSED [100%]
-
-============================== 88 passed in 1.83s ==============================
diff --git a/bench/env/pytest-unit-ef2a18c.txt b/bench/env/pytest-unit-ef2a18c.txt
deleted file mode 100644
index 984defe..0000000
--- a/bench/env/pytest-unit-ef2a18c.txt
+++ /dev/null
@@ -1,39 +0,0 @@
-============================= test session starts ==============================
-platform darwin -- Python 3.11.14, pytest-9.0.3, pluggy-1.6.0 -- /Users/aszczepanski/dev/ComfyUI/custom_nodes/ComfyUI-CoreMLSuite/.venv/bin/python
-cachedir: .pytest_cache
-rootdir: /Users/aszczepanski/dev/ComfyUI
-configfile: pytest.ini
-plugins: jaxtyping-0.3.9
-collecting ... collected 20 items
-
-tests/unit/test_chunks.py::test_batch_chunking[1] PASSED
-tests/unit/test_chunks.py::test_batch_chunking[2] PASSED
-tests/unit/test_chunks.py::test_batch_chunking[4] PASSED
-tests/unit/test_chunks.py::test_batch_chunking[5] PASSED
-tests/unit/test_chunks.py::test_batch_chunking[9] PASSED
-tests/unit/test_chunks.py::test_merge_chunks[1] PASSED
-tests/unit/test_chunks.py::test_merge_chunks[2] PASSED
-tests/unit/test_chunks.py::test_merge_chunks[4] PASSED
-tests/unit/test_chunks.py::test_merge_chunks[5] PASSED
-tests/unit/test_chunks.py::test_merge_chunks[9] PASSED
-tests/unit/test_chunks.py::test_chunking_controlnet[1-2-1] PASSED
-tests/unit/test_chunks.py::test_chunking_controlnet[1-1-1] PASSED
-tests/unit/test_chunks.py::test_chunking_controlnet[2-2-1] PASSED
-tests/unit/test_chunks.py::test_chunking_controlnet[3-2-2] PASSED
-tests/unit/test_chunks.py::test_chunking_controlnet[4-2-2] PASSED
-tests/unit/test_chunks.py::test_chunking_controlnet[5-3-2] PASSED
-tests/unit/test_chunks.py::test_chunking_controlnet[9-4-3] PASSED
-tests/unit/test_chunks.py::test_chunking_no_control PASSED
-tests/unit/test_chunks.py::test_chunking_inputs PASSED
-tests/unit/test_controlnet.py::test_no_control PASSED
-
-=============================== warnings summary ===============================
-<frozen importlib._bootstrap>:241
-  <frozen importlib._bootstrap>:241: DeprecationWarning: builtin type SwigPyPacked has no __module__ attribute
-
-<frozen importlib._bootstrap>:241
-  <frozen importlib._bootstrap>:241: DeprecationWarning: builtin type SwigPyObject has no __module__ attribute
-
--- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
-======================== 20 passed, 2 warnings in 2.81s ========================
-sys:1: DeprecationWarning: builtin type swigvarlink has no __module__ attribute
diff --git a/bench/prompts.json b/bench/prompts.json
deleted file mode 100644
index 6a4017e..0000000
--- a/bench/prompts.json
+++ /dev/null
@@ -1,19 +0,0 @@
-{
-  "_comment": "Fixed prompt set for Phase 1 baseline reproducibility. Do not edit without bumping the bench version: changing prompts invalidates cross-baseline PSNR comparisons.",
-  "seed": 0,
-  "negative_prompt": "blurry, low quality, text, watermark",
-  "prompts": [
-    {
-      "id": "p01_portrait",
-      "text": "a portrait of an astronaut in a sunflower field, soft natural light, 35mm film"
-    },
-    {
-      "id": "p02_cityscape",
-      "text": "a futuristic city skyline at sunset, glass towers, volumetric lighting, cinematic"
-    },
-    {
-      "id": "p03_still_life",
-      "text": "a still life of fruit on a wooden table, oil painting, dramatic chiaroscuro"
-    }
-  ]
-}
diff --git a/bench/results/1e5791d.json b/bench/results/1e5791d.json
deleted file mode 100644
index a3e1fcd..0000000
--- a/bench/results/1e5791d.json
+++ /dev/null
@@ -1,103 +0,0 @@
-{
-  "git_sha": "1e5791d",
-  "timestamp": "2026-05-24T14:37:16+00:00",
-  "host": {
-    "machine": "arm64",
-    "system": "Darwin",
-    "release": "25.1.0",
-    "python": "3.12.11"
-  },
-  "tool_versions": {
-    "coremltools": "9.0",
-    "numpy": "1.26.4"
-  },
-  "settings": {
-    "repeats": 30,
-    "assumed_steps": 20,
-    "input_seed": 0,
-    "compute_units": [
-      "CPU_AND_NE",
-      "CPU_AND_GPU"
-    ]
-  },
-  "results": [
-    {
-      "label": "v1-5-pruned-emaonly_1x512x512_se_unet-CPU_AND_NE",
-      "model_path": "/Users/aszczepanski/dev/ComfyUI/models/unet/v1-5-pruned-emaonly_1x512x512_se_unet.mlmodelc",
-      "compute_unit": "CPU_AND_NE",
-      "error": null,
-      "model_size_bytes": 1720796055,
-      "load_time_s": 0.2086,
-      "kind": "sd15",
-      "expected_inputs": {
-        "sample": [
-          1,
-          4,
-          64,
-          64
-        ],
-        "timestep": [
-          1
-        ],
-        "encoder_hidden_states": [
-          1,
-          768,
-          1,
-          77
-        ]
-      },
-      "warmup_ms": 219.471,
-      "forward_ms": {
-        "mean": 199.435,
-        "std": 5.333,
-        "min": 196.825,
-        "median": 197.493,
-        "n": 30
-      },
-      "est_e2e_ms": 3949.9,
-      "peak_rss_mb": 399.0
-    },
-    {
-      "label": "v1-5-pruned-emaonly_1x512x512_se_unet-CPU_AND_GPU",
-      "model_path": "/Users/aszczepanski/dev/ComfyUI/models/unet/v1-5-pruned-emaonly_1x512x512_se_unet.mlmodelc",
-      "compute_unit": "CPU_AND_GPU",
-      "error": null,
-      "model_size_bytes": 1720796055,
-      "load_time_s": 1.2334,
-      "kind": "sd15",
-      "expected_inputs": {
-        "sample": [
-          1,
-          4,
-          64,
-          64
-        ],
-        "timestep": [
-          1
-        ],
-        "encoder_hidden_states": [
-          1,
-          768,
-          1,
-          77
-        ]
-      },
-      "warmup_ms": 1392.959,
-      "forward_ms": {
-        "mean": 281.433,
-        "std": 11.822,
-        "min": 268.273,
-        "median": 278.577,
-        "n": 30
-      },
-      "est_e2e_ms": 5571.5,
-      "peak_rss_mb": 2113.5
-    }
-  ],
-  "notes": [
-    "forward_ms is the latency of a single Core ML UNet forward pass with synthetic inputs; with classifier-free guidance one sampler step typically maps to one batched forward (cond+uncond) \u2014 interpret est_e2e_ms accordingly.",
-    "peak_rss_mb is process RSS only and excludes ANE/wired GPU memory; treat as a floor.",
-    "quality_psnr is null unless --reference-images and --candidate-images are provided.",
-    "Latency varies run-to-run; 'min' is the most reproducible figure for comparisons."
-  ]
-}
\ No newline at end of file
diff --git a/bench/results/1e5791d.md b/bench/results/1e5791d.md
deleted file mode 100644
index 1b854df..0000000
--- a/bench/results/1e5791d.md
+++ /dev/null
@@ -1,19 +0,0 @@
-# CoreMLSuite baseline — `1e5791d`
-
-- timestamp: 2026-05-24T14:37:16+00:00
-- host: arm64 / Darwin 25.1.0
-- python: 3.12.11
-- coremltools: 9.0, numpy: 1.26.4
-- settings: repeats=30, input_seed=0, assumed_steps=20
-
-| label | kind | compute | size (MB) | load (s) | warmup (ms) | fwd median (ms) | fwd min (ms) | est e2e (ms) | peak RSS (MB) | error |
-|---|---|---|---|---|---|---|---|---|---|---|
-| v1-5-pruned-emaonly_1x512x512_se_unet-CPU_AND_NE | sd15 | CPU_AND_NE | 1641.1 | 0.2086 | 219.471 | 197.493 | 196.825 | 3949.9 | 399.0 |  |
-| v1-5-pruned-emaonly_1x512x512_se_unet-CPU_AND_GPU | sd15 | CPU_AND_GPU | 1641.1 | 1.2334 | 1392.959 | 278.577 | 268.273 | 5571.5 | 2113.5 |  |
-
-## Notes
-
-- forward_ms is the latency of a single Core ML UNet forward pass with synthetic inputs; with classifier-free guidance one sampler step typically maps to one batched forward (cond+uncond) — interpret est_e2e_ms accordingly.
-- peak_rss_mb is process RSS only and excludes ANE/wired GPU memory; treat as a floor.
-- quality_psnr is null unless --reference-images and --candidate-images are provided.
-- Latency varies run-to-run; 'min' is the most reproducible figure for comparisons.
diff --git a/bench/results/ef2a18c.json b/bench/results/ef2a18c.json
deleted file mode 100644
index b5f52ae..0000000
--- a/bench/results/ef2a18c.json
+++ /dev/null
@@ -1,103 +0,0 @@
-{
-  "git_sha": "ef2a18c",
-  "timestamp": "2026-05-22T13:07:25+00:00",
-  "host": {
-    "machine": "arm64",
-    "system": "Darwin",
-    "release": "25.1.0",
-    "python": "3.11.14"
-  },
-  "tool_versions": {
-    "coremltools": "8.2",
-    "numpy": "1.23.5"
-  },
-  "settings": {
-    "repeats": 30,
-    "assumed_steps": 20,
-    "input_seed": 0,
-    "compute_units": [
-      "CPU_AND_NE",
-      "CPU_AND_GPU"
-    ]
-  },
-  "results": [
-    {
-      "label": "v1-5-pruned-emaonly_1x512x512_se_unet-CPU_AND_NE",
-      "model_path": "/Users/aszczepanski/dev/ComfyUI/models/unet/v1-5-pruned-emaonly_1x512x512_se_unet.mlmodelc",
-      "compute_unit": "CPU_AND_NE",
-      "error": null,
-      "model_size_bytes": 1720795925,
-      "load_time_s": 0.2841,
-      "kind": "sd15",
-      "expected_inputs": {
-        "sample": [
-          1,
-          4,
-          64,
-          64
-        ],
-        "timestep": [
-          1
-        ],
-        "encoder_hidden_states": [
-          1,
-          768,
-          1,
-          77
-        ]
-      },
-      "warmup_ms": 225.201,
-      "forward_ms": {
-        "mean": 197.126,
-        "std": 0.476,
-        "min": 196.812,
-        "median": 196.969,
-        "n": 30
-      },
-      "est_e2e_ms": 3939.4,
-      "peak_rss_mb": 349.7
-    },
-    {
-      "label": "v1-5-pruned-emaonly_1x512x512_se_unet-CPU_AND_GPU",
-      "model_path": "/Users/aszczepanski/dev/ComfyUI/models/unet/v1-5-pruned-emaonly_1x512x512_se_unet.mlmodelc",
-      "compute_unit": "CPU_AND_GPU",
-      "error": null,
-      "model_size_bytes": 1720795925,
-      "load_time_s": 19.959,
-      "kind": "sd15",
-      "expected_inputs": {
-        "sample": [
-          1,
-          4,
-          64,
-          64
-        ],
-        "timestep": [
-          1
-        ],
-        "encoder_hidden_states": [
-          1,
-          768,
-          1,
-          77
-        ]
-      },
-      "warmup_ms": 558.865,
-      "forward_ms": {
-        "mean": 270.275,
-        "std": 1.312,
-        "min": 267.398,
-        "median": 270.259,
-        "n": 30
-      },
-      "est_e2e_ms": 5405.2,
-      "peak_rss_mb": 2132.9
-    }
-  ],
-  "notes": [
-    "forward_ms is the latency of a single Core ML UNet forward pass with synthetic inputs; with classifier-free guidance one sampler step typically maps to one batched forward (cond+uncond) \u2014 interpret est_e2e_ms accordingly.",
-    "peak_rss_mb is process RSS only and excludes ANE/wired GPU memory; treat as a floor.",
-    "quality_psnr is null unless --reference-images and --candidate-images are provided.",
-    "Latency varies run-to-run; 'min' is the most reproducible figure for comparisons."
-  ]
-}
\ No newline at end of file
diff --git a/bench/results/ef2a18c.md b/bench/results/ef2a18c.md
deleted file mode 100644
index 9db2035..0000000
--- a/bench/results/ef2a18c.md
+++ /dev/null
@@ -1,19 +0,0 @@
-# CoreMLSuite baseline — `ef2a18c`
-
-- timestamp: 2026-05-22T13:07:25+00:00
-- host: arm64 / Darwin 25.1.0
-- python: 3.11.14
-- coremltools: 8.2, numpy: 1.23.5
-- settings: repeats=30, input_seed=0, assumed_steps=20
-
-| label | kind | compute | size (MB) | load (s) | warmup (ms) | fwd median (ms) | fwd min (ms) | est e2e (ms) | peak RSS (MB) | error |
-|---|---|---|---|---|---|---|---|---|---|---|
-| v1-5-pruned-emaonly_1x512x512_se_unet-CPU_AND_NE | sd15 | CPU_AND_NE | 1641.1 | 0.2841 | 225.201 | 196.969 | 196.812 | 3939.4 | 349.7 |  |
-| v1-5-pruned-emaonly_1x512x512_se_unet-CPU_AND_GPU | sd15 | CPU_AND_GPU | 1641.1 | 19.959 | 558.865 | 270.259 | 267.398 | 5405.2 | 2132.9 |  |
-
-## Notes
-
-- forward_ms is the latency of a single Core ML UNet forward pass with synthetic inputs; with classifier-free guidance one sampler step typically maps to one batched forward (cond+uncond) — interpret est_e2e_ms accordingly.
-- peak_rss_mb is process RSS only and excludes ANE/wired GPU memory; treat as a floor.
-- quality_psnr is null unless --reference-images and --candidate-images are provided.
-- Latency varies run-to-run; 'min' is the most reproducible figure for comparisons.
diff --git a/bench/results/quant_matrix_0bbd8d8.json b/bench/results/quant_matrix_0bbd8d8.json
deleted file mode 100644
index 8055500..0000000
--- a/bench/results/quant_matrix_0bbd8d8.json
+++ /dev/null
@@ -1,49 +0,0 @@
-{
-  "git_sha": "0bbd8d8",
-  "compute_unit": "CPU_AND_NE",
-  "repeats": 20,
-  "input_seed": 0,
-  "ckpt_stem": "v1-5-pruned-emaonly",
-  "rows": [
-    {
-      "nbits": "none",
-      "path": "/Users/aszczepanski/dev/ComfyUI/models/unet/v1-5-pruned-emaonly_1x512x512_se_unet.mlmodelc",
-      "size_bytes": 1720796055,
-      "load_time_s": 0.1997,
-      "fwd_ms_median": 196.651,
-      "fwd_ms_min": 196.588,
-      "psnr_db_vs_none": null,
-      "size_ratio_vs_none": 1.0
-    },
-    {
-      "nbits": "8",
-      "path": "/Users/aszczepanski/dev/ComfyUI/models/unet/v1-5-pruned-emaonly_1x512x512_se_q8_unet.mlmodelc",
-      "size_bytes": 861946375,
-      "load_time_s": 0.1475,
-      "fwd_ms_median": 185.223,
-      "fwd_ms_min": 183.373,
-      "psnr_db_vs_none": 53.54,
-      "size_ratio_vs_none": 0.501
-    },
-    {
-      "nbits": "6",
-      "path": "/Users/aszczepanski/dev/ComfyUI/models/unet/v1-5-pruned-emaonly_1x512x512_se_q6_unet.mlmodelc",
-      "size_bytes": 647068433,
-      "load_time_s": 0.1457,
-      "fwd_ms_median": 181.091,
-      "fwd_ms_min": 180.977,
-      "psnr_db_vs_none": 40.2,
-      "size_ratio_vs_none": 0.376
-    },
-    {
-      "nbits": "4",
-      "path": "/Users/aszczepanski/dev/ComfyUI/models/unet/v1-5-pruned-emaonly_1x512x512_se_q4_unet.mlmodelc",
-      "size_bytes": 432280915,
-      "load_time_s": 0.1427,
-      "fwd_ms_median": 178.953,
-      "fwd_ms_min": 178.753,
-      "psnr_db_vs_none": 27.47,
-      "size_ratio_vs_none": 0.251
-    }
-  ]
-}
\ No newline at end of file
diff --git a/bench/results/quant_matrix_0bbd8d8.md b/bench/results/quant_matrix_0bbd8d8.md
deleted file mode 100644
index 0e14a57..0000000
--- a/bench/results/quant_matrix_0bbd8d8.md
+++ /dev/null
@@ -1,12 +0,0 @@
-# Quantization tradeoff matrix — `0bbd8d8`
-
-- compute unit: CPU_AND_NE
-- repeats per variant: 20
-- ckpt: v1-5-pruned-emaonly, 1x512x512, attn=se
-
-| nbits | size (MB) | size vs none | load (s) | fwd median (ms) | fwd min (ms) | PSNR vs none (dB) | error |
-|---|---|---|---|---|---|---|---|
-| none | 1641.1 | 1.0 | 0.1997 | 196.651 | 196.588 | - |  |
-| 8 | 822.0 | 0.501 | 0.1475 | 185.223 | 183.373 | 53.54 |  |
-| 6 | 617.1 | 0.376 | 0.1457 | 181.091 | 180.977 | 40.2 |  |
-| 4 | 412.3 | 0.251 | 0.1427 | 178.953 | 178.753 | 27.47 |  |
diff --git a/bench/results/smoke/ef2a18c/E2E-1.5-CoreML_00002_.png b/bench/results/smoke/ef2a18c/E2E-1.5-CoreML_00002_.png
deleted file mode 100644
index a561e16..0000000
Binary files a/bench/results/smoke/ef2a18c/E2E-1.5-CoreML_00002_.png and /dev/null differ
diff --git a/bench/results/smoke/ef2a18c/report.json b/bench/results/smoke/ef2a18c/report.json
deleted file mode 100644
index e86e3be..0000000
--- a/bench/results/smoke/ef2a18c/report.json
+++ /dev/null
@@ -1,10 +0,0 @@
-{
-  "git_sha": "ef2a18c",
-  "seed": 42,
-  "ckpt_name": "v1-5-pruned-emaonly.safetensors",
-  "workflow": "tests/integration/workflows/e2e-1.5-basic-conversion.json",
-  "coreml_image": "bench/results/smoke/ef2a18c/E2E-1.5-CoreML_00002_.png",
-  "mps_image": null,
-  "psnr_db": null,
-  "wall_seconds": 21.1
-}
\ No newline at end of file
diff --git a/bench/run.py b/bench/run.py
deleted file mode 100644
index 37223c1..0000000
--- a/bench/run.py
+++ /dev/null
@@ -1,417 +0,0 @@
-#!/usr/bin/env python3
-"""
-bench/run.py — Phase 1 baseline benchmark harness for ComfyUI-CoreMLSuite.
-
-WHAT THIS MEASURES (the deterministic, easy-to-trust numbers)
--------------------------------------------------------------
-For each converted Core ML UNet model in the matrix, in-process:
-  * model load time (includes mlpackage compile; mlmodelc is precompiled)
-  * model file size on disk
-  * warmup (first forward) time
-  * steady-state UNet forward latency (mean/std/min/median over N repeats)
-  * an end-to-end estimate (median forward * assumed sampler steps)
-  * peak process RSS during the run
-
-It drives the Core ML model DIRECTLY with synthetic inputs shaped from the
-model's own `expected_inputs`. This isolates the ANE/Core ML UNet latency —
-the exact quantity that quantization (Phase 6) and the coremltools upgrade
-(Phase 5) are expected to move — WITHOUT needing a live ComfyUI server, CLIP,
-VAE, or a checkpoint. It is therefore the cleanest, most reproducible signal.
-
-WHAT THIS DOES NOT MEASURE
---------------------------
-  * Image quality. PSNR is a separate, decoupled concern: pass --reference-images
-    and --candidate-images (e.g. the E2E-1.5-MPS / E2E-1.5-CoreML PNGs your
-    existing integration workflow already produces) and PSNR is computed per
-    matching filename. Without them, quality_psnr is null (and that's honest).
-  * ANE / wired GPU memory. peak_rss_mb is *process* RSS only; treat as a coarse
-    floor, not the true device footprint.
-
-REQUIREMENTS
-------------
-  * Run on Apple Silicon (macOS) with the suite's deps installed
-    (coremltools + apple/ml-stable-diffusion providing `python_coreml_stable_diffusion`).
-  * numpy is required. psutil and Pillow are optional (graceful fallback).
-
-USAGE
------
-  # Auto-discover every .mlmodelc/.mlpackage under a dir, sweep compute units:
-  python bench/run.py --models-dir /path/to/ComfyUI/models/unet
-
-  # Explicit matrix (recommended for a stable, committed baseline):
-  python bench/run.py --matrix bench/matrix.json --models-dir /path/to/models/unet
-
-  # With quality from already-produced images:
-  python bench/run.py --models-dir ... \
-      --reference-images /path/to/mps_pngs --candidate-images /path/to/coreml_pngs
-
-matrix.json format (a list of configs):
-  [
-    {"label": "sd15-se",  "model": "dreamshaper_8_1x512x512_se.mlmodelc", "compute_unit": "CPU_AND_NE"},
-    {"label": "sdxl-orig","model": "sdxl_base_1x1024x1024_orig.mlmodelc",  "compute_unit": "CPU_AND_GPU"}
-  ]
-Each "model" is resolved against --models-dir unless it is an absolute path.
-
-Output: bench/results/<gitsha>.json  and  bench/results/<gitsha>.md
-"""
-
-import argparse
-import dataclasses
-import datetime as _dt
-import glob
-import json
-import os
-import platform
-import statistics
-import subprocess
-import sys
-import time
-from pathlib import Path
-from typing import Any, Optional
-
-import numpy as np
-
-# ----- optional deps (graceful) ---------------------------------------------
-try:
-    import psutil  # type: ignore
-
-    _PROC = psutil.Process(os.getpid())
-except Exception:  # pragma: no cover - optional
-    psutil = None
-    _PROC = None
-
-DEFAULT_COMPUTE_UNITS = ["CPU_AND_NE", "CPU_AND_GPU"]
-COREML_EXTS = (".mlmodelc", ".mlpackage")
-
-
-# ----- small helpers --------------------------------------------------------
-def git_sha() -> str:
-    try:
-        out = subprocess.check_output(
-            ["git", "rev-parse", "--short", "HEAD"], stderr=subprocess.DEVNULL
-        )
-        return out.decode().strip() or "nogit"
-    except Exception:
-        return "nogit"
-
-
-def coremltools_version() -> str:
-    try:
-        import coremltools as ct  # noqa: WPS433 (local import is intentional)
-
-        return getattr(ct, "__version__", "unknown")
-    except Exception as exc:  # pragma: no cover
-        return f"import-failed: {exc}"
-
-
-def dir_size_bytes(path: Path) -> int:
-    """Core ML models are directories (.mlmodelc / .mlpackage)."""
-    if path.is_file():
-        return path.stat().st_size
-    total = 0
-    for root, _dirs, files in os.walk(path):
-        for f in files:
-            try:
-                total += (Path(root) / f).stat().st_size
-            except OSError:
-                pass
-    return total
-
-
-def sample_rss_mb() -> Optional[float]:
-    if _PROC is None:
-        return None
-    try:
-        return _PROC.memory_info().rss / (1024 * 1024)
-    except Exception:
-        return None
-
-
-def detect_kind(expected_inputs: dict) -> str:
-    """Mirror the suite's detection WITHOUT importing comfy (keeps bench light)."""
-    if "time_ids" in expected_inputs and "text_embeds" in expected_inputs:
-        try:
-            n = expected_inputs["time_ids"]["shape"][1]
-        except Exception:
-            n = -1
-        return "sdxl_base" if n == 6 else "sdxl_refiner" if n == 5 else "sdxl_unknown"
-    if "timestep_cond" in expected_inputs:
-        return "lcm"
-    return "sd15"
-
-
-def make_inputs(expected_inputs: dict, rng: np.random.Generator) -> dict[str, np.ndarray]:
-    """Build synthetic fp16 inputs matching the model's expected shapes.
-
-    Values are random; they do not affect latency in any meaningful way, and
-    we feed fp16 to match how CoreMLModelWrapper feeds the model at runtime.
-    """
-    inputs: dict[str, np.ndarray] = {}
-    for name, spec in expected_inputs.items():
-        shape = tuple(int(d) for d in spec["shape"])
-        inputs[name] = rng.standard_normal(shape).astype(np.float16)
-    return inputs
-
-
-def psnr(img_a: "np.ndarray", img_b: "np.ndarray") -> float:
-    mse = float(np.mean((img_a.astype(np.float64) - img_b.astype(np.float64)) ** 2))
-    if mse == 0:
-        return 100.0
-    return 20.0 * float(np.log10(255.0 / np.sqrt(mse)))
-
-
-# ----- core measurement -----------------------------------------------------
-@dataclasses.dataclass
-class Config:
-    label: str
-    model_path: str
-    compute_unit: str
-
-
-def resolve_matrix(args) -> list[Config]:
-    configs: list[Config] = []
-
-    if args.matrix:
-        entries = json.loads(Path(args.matrix).read_text())
-        for e in entries:
-            model = e["model"]
-            if not os.path.isabs(model):
-                if not args.models_dir:
-                    raise SystemExit("matrix uses relative model names but --models-dir not given")
-                model = os.path.join(args.models_dir, model)
-            configs.append(Config(e.get("label", Path(model).stem), model, e.get("compute_unit", "CPU_AND_NE")))
-        return configs
-
-    if args.model:
-        for m in args.model:
-            for cu in args.compute_units:
-                configs.append(Config(f"{Path(m).stem}-{cu}", m, cu))
-        return configs
-
-    if args.models_dir:
-        found: list[str] = []
-        for ext in COREML_EXTS:
-            found += glob.glob(os.path.join(args.models_dir, f"*{ext}"))
-        found = sorted(set(found))
-        if not found:
-            raise SystemExit(f"No {COREML_EXTS} models found under {args.models_dir}")
-        for m in found:
-            for cu in args.compute_units:
-                configs.append(Config(f"{Path(m).stem}-{cu}", m, cu))
-        return configs
-
-    raise SystemExit("Provide one of: --matrix, --model, or --models-dir")
-
-
-def measure(cfg: Config, args, rng: np.random.Generator) -> dict[str, Any]:
-    from python_coreml_stable_diffusion.coreml_model import CoreMLModel  # local import: M2 only
-
-    result: dict[str, Any] = {
-        "label": cfg.label,
-        "model_path": cfg.model_path,
-        "compute_unit": cfg.compute_unit,
-        "error": None,
-    }
-    path = Path(cfg.model_path)
-    if not path.exists():
-        result["error"] = "model path does not exist"
-        return result
-
-    sources = "compiled" if cfg.model_path.endswith(".mlmodelc") else "packages"
-    peak_rss = sample_rss_mb()
-
-    try:
-        result["model_size_bytes"] = dir_size_bytes(path)
-
-        t0 = time.perf_counter()
-        model = CoreMLModel(cfg.model_path, cfg.compute_unit, sources)
-        result["load_time_s"] = round(time.perf_counter() - t0, 4)
-        peak_rss = max(filter(None, [peak_rss, sample_rss_mb()]), default=None)
-
-        expected = dict(model.expected_inputs)
-        result["kind"] = detect_kind(expected)
-        result["expected_inputs"] = {k: list(v["shape"]) for k, v in expected.items()}
-
-        # Pre-generate all input sets so timing excludes array allocation.
-        warm_in = make_inputs(expected, rng)
-        step_inputs = [make_inputs(expected, rng) for _ in range(args.repeats)]
-
-        # Warmup (first forward — ANE prepares its graph here).
-        t0 = time.perf_counter()
-        out = model(**warm_in)
-        result["warmup_ms"] = round((time.perf_counter() - t0) * 1000.0, 3)
-        if not (isinstance(out, dict) and "noise_pred" in out):
-            result["error"] = "unexpected output (no 'noise_pred')"
-        peak_rss = max(filter(None, [peak_rss, sample_rss_mb()]), default=None)
-
-        # Steady-state forward latency.
-        times_ms: list[float] = []
-        for inp in step_inputs:
-            t0 = time.perf_counter()
-            model(**inp)
-            times_ms.append((time.perf_counter() - t0) * 1000.0)
-            rss = sample_rss_mb()
-            if rss is not None:
-                peak_rss = rss if peak_rss is None else max(peak_rss, rss)
-
-        result["forward_ms"] = {
-            "mean": round(statistics.fmean(times_ms), 3),
-            "std": round(statistics.pstdev(times_ms), 3) if len(times_ms) > 1 else 0.0,
-            "min": round(min(times_ms), 3),
-            "median": round(statistics.median(times_ms), 3),
-            "n": len(times_ms),
-        }
-        result["est_e2e_ms"] = round(result["forward_ms"]["median"] * args.assumed_steps, 1)
-        result["peak_rss_mb"] = round(peak_rss, 1) if peak_rss is not None else None
-
-    except Exception as exc:  # keep one bad model from killing the whole run
-        result["error"] = f"{type(exc).__name__}: {exc}"
-
-    return result
-
-
-def compute_quality(reference_dir: str, candidate_dir: str) -> dict[str, Any]:
-    try:
-        from PIL import Image  # type: ignore
-    except Exception as exc:  # pragma: no cover
-        return {"error": f"Pillow not available: {exc}"}
-
-    ref = {p.name: p for p in Path(reference_dir).glob("*.png")}
-    cand = {p.name: p for p in Path(candidate_dir).glob("*.png")}
-    common = sorted(set(ref) & set(cand))
-    pairs: dict[str, Any] = {}
-    for name in common:
-        try:
-            a = np.array(Image.open(ref[name]).convert("RGB"))
-            b = np.array(Image.open(cand[name]).convert("RGB"))
-            if a.shape != b.shape:
-                pairs[name] = {"error": f"shape mismatch {a.shape} vs {b.shape}"}
-                continue
-            pairs[name] = {"psnr_db": round(psnr(a, b), 2)}
-        except Exception as exc:
-            pairs[name] = {"error": str(exc)}
-    return {
-        "reference_dir": reference_dir,
-        "candidate_dir": candidate_dir,
-        "matched": len(common),
-        "pairs": pairs,
-    }
-
-
-# ----- reporting ------------------------------------------------------------
-def to_markdown(report: dict[str, Any]) -> str:
-    lines = [
-        f"# CoreMLSuite baseline — `{report['git_sha']}`",
-        "",
-        f"- timestamp: {report['timestamp']}",
-        f"- host: {report['host']['machine']} / {report['host']['system']} {report['host']['release']}",
-        f"- python: {report['host']['python']}",
-        f"- coremltools: {report['tool_versions']['coremltools']}, numpy: {report['tool_versions']['numpy']}",
-        f"- settings: repeats={report['settings']['repeats']}, "
-        f"input_seed={report['settings']['input_seed']}, "
-        f"assumed_steps={report['settings']['assumed_steps']}",
-        "",
-        "| label | kind | compute | size (MB) | load (s) | warmup (ms) | fwd median (ms) | fwd min (ms) | est e2e (ms) | peak RSS (MB) | error |",
-        "|---|---|---|---|---|---|---|---|---|---|---|",
-    ]
-    for r in report["results"]:
-        if r.get("error") and "forward_ms" not in r:
-            lines.append(
-                f"| {r['label']} | - | {r['compute_unit']} | - | - | - | - | - | - | - | {r['error']} |"
-            )
-            continue
-        size_mb = round(r.get("model_size_bytes", 0) / (1024 * 1024), 1)
-        fwd = r.get("forward_ms", {})
-        lines.append(
-            f"| {r['label']} | {r.get('kind','?')} | {r['compute_unit']} | {size_mb} | "
-            f"{r.get('load_time_s','-')} | {r.get('warmup_ms','-')} | {fwd.get('median','-')} | "
-            f"{fwd.get('min','-')} | {r.get('est_e2e_ms','-')} | {r.get('peak_rss_mb','-')} | "
-            f"{r.get('error') or ''} |"
-        )
-
-    q = report.get("quality")
-    if q and q.get("pairs"):
-        lines += ["", "## Quality (PSNR vs reference)", "", "| image | PSNR (dB) |", "|---|---|"]
-        for name, val in q["pairs"].items():
-            lines.append(f"| {name} | {val.get('psnr_db', val.get('error','?'))} |")
-
-    if report.get("notes"):
-        lines += ["", "## Notes", ""] + [f"- {n}" for n in report["notes"]]
-    return "\n".join(lines) + "\n"
-
-
-def main(argv: Optional[list[str]] = None) -> int:
-    ap = argparse.ArgumentParser(description="CoreMLSuite Phase-1 baseline benchmark")
-    ap.add_argument("--matrix", help="path to matrix.json (list of {label, model, compute_unit})")
-    ap.add_argument("--models-dir", help="dir holding .mlmodelc/.mlpackage models")
-    ap.add_argument("--model", action="append", help="explicit model path (repeatable)")
-    ap.add_argument("--compute-units", nargs="+", default=DEFAULT_COMPUTE_UNITS,
-                    help="compute units to sweep in auto/--model mode")
-    ap.add_argument("--repeats", type=int, default=30, help="steady-state forward passes")
-    ap.add_argument("--assumed-steps", type=int, default=20,
-                    help="sampler steps used ONLY for the est_e2e_ms estimate")
-    ap.add_argument("--input-seed", type=int, default=0, help="seed for synthetic input generation")
-    ap.add_argument("--reference-images", help="dir of reference PNGs (e.g. MPS) for optional PSNR")
-    ap.add_argument("--candidate-images", help="dir of candidate PNGs (e.g. CoreML) for optional PSNR")
-    ap.add_argument("--out-dir", default="bench/results", help="where to write <gitsha>.json/.md")
-    args = ap.parse_args(argv)
-
-    sha = git_sha()
-    rng = np.random.default_rng(args.input_seed)
-    configs = resolve_matrix(args)
-
-    notes = [
-        "forward_ms is the latency of a single Core ML UNet forward pass with synthetic "
-        "inputs; with classifier-free guidance one sampler step typically maps to one "
-        "batched forward (cond+uncond) — interpret est_e2e_ms accordingly.",
-        "peak_rss_mb is process RSS only and excludes ANE/wired GPU memory; treat as a floor.",
-        "quality_psnr is null unless --reference-images and --candidate-images are provided.",
-        "Latency varies run-to-run; 'min' is the most reproducible figure for comparisons.",
-    ]
-    if psutil is None:
-        notes.append("psutil not installed -> peak_rss_mb is null. `pip install psutil` to capture it.")
-
-    print(f"[bench] git={sha} coremltools={coremltools_version()} configs={len(configs)}", file=sys.stderr)
-
-    results = []
-    for cfg in configs:
-        print(f"[bench] measuring {cfg.label} ({cfg.compute_unit}) ...", file=sys.stderr)
-        results.append(measure(cfg, args, rng))
-
-    report: dict[str, Any] = {
-        "git_sha": sha,
-        "timestamp": _dt.datetime.now(_dt.timezone.utc).isoformat(timespec="seconds"),
-        "host": {
-            "machine": platform.machine(),
-            "system": platform.system(),
-            "release": platform.release(),
-            "python": platform.python_version(),
-        },
-        "tool_versions": {"coremltools": coremltools_version(), "numpy": np.__version__},
-        "settings": {
-            "repeats": args.repeats,
-            "assumed_steps": args.assumed_steps,
-            "input_seed": args.input_seed,
-            "compute_units": args.compute_units,
-        },
-        "results": results,
-        "notes": notes,
-    }
-
-    if args.reference_images and args.candidate_images:
-        report["quality"] = compute_quality(args.reference_images, args.candidate_images)
-
-    out_dir = Path(args.out_dir)
-    out_dir.mkdir(parents=True, exist_ok=True)
-    json_path = out_dir / f"{sha}.json"
-    md_path = out_dir / f"{sha}.md"
-    json_path.write_text(json.dumps(report, indent=2))
-    md_path.write_text(to_markdown(report))
-
-    print(f"[bench] wrote {json_path}", file=sys.stderr)
-    print(f"[bench] wrote {md_path}", file=sys.stderr)
-    return 0
-
-
-if __name__ == "__main__":
-    raise SystemExit(main())
diff --git a/bench/scripts/convert_sd15.py b/bench/scripts/convert_sd15.py
deleted file mode 100644
index 5e9388e..0000000
--- a/bench/scripts/convert_sd15.py
+++ /dev/null
@@ -1,99 +0,0 @@
-#!/usr/bin/env python3
-"""bench/scripts/convert_sd15.py — Phase 1 baseline UNet conversion.
-
-Converts a SD1.5 checkpoint to a Core ML UNet (.mlmodelc) for the bench harness.
-Bypasses the ComfyUI node graph and calls coreml_suite.converter directly so
-the conversion is reproducible from a single command.
-
-Run from the repo root with the project venv:
-    .venv/bin/python bench/scripts/convert_sd15.py
-
-Environment overrides:
-    COMFY_DIR=...                       (default: ../..)
-    CKPT_NAME=v1-5-pruned-emaonly.safetensors
-    ATTN=SPLIT_EINSUM                   (SPLIT_EINSUM | SPLIT_EINSUM_V2 | ORIGINAL)
-    HEIGHT=512 WIDTH=512 BATCH_SIZE=1
-    CONTROLNET=0
-"""
-import logging
-import os
-import sys
-from pathlib import Path
-
-logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")
-log = logging.getLogger("convert_sd15")
-
-REPO_ROOT = Path(__file__).resolve().parents[2]
-COMFY_DIR = Path(os.environ.get("COMFY_DIR", REPO_ROOT.parents[1])).resolve()
-
-if str(COMFY_DIR) not in sys.path:
-    sys.path.insert(0, str(COMFY_DIR))
-if str(REPO_ROOT) not in sys.path:
-    sys.path.insert(0, str(REPO_ROOT))
-
-import folder_paths  # noqa: E402  (ComfyUI module — needs sys.path above)
-
-# Ensure ComfyUI's folder_paths is pointed at the real install for checkpoints.
-folder_paths.base_path = str(COMFY_DIR)
-folder_paths.add_model_folder_path("checkpoints", str(COMFY_DIR / "models" / "checkpoints"))
-folder_paths.add_model_folder_path("unet", str(COMFY_DIR / "models" / "unet"))
-
-from coreml_suite import converter  # noqa: E402
-from coreml_suite.config import ModelVersion  # noqa: E402
-
-
-def main() -> int:
-    ckpt_name = os.environ.get("CKPT_NAME", "v1-5-pruned-emaonly.safetensors")
-    attn = os.environ.get("ATTN", "SPLIT_EINSUM")
-    height = int(os.environ.get("HEIGHT", "512"))
-    width = int(os.environ.get("WIDTH", "512"))
-    batch_size = int(os.environ.get("BATCH_SIZE", "1"))
-    controlnet = os.environ.get("CONTROLNET", "0") not in ("0", "false", "False", "")
-    quantize_nbits = os.environ.get("QUANT_NBITS", "none")
-
-    ckpt_path = folder_paths.get_full_path("checkpoints", ckpt_name)
-    if not ckpt_path:
-        log.error("checkpoint not found: %s under %s", ckpt_name, COMFY_DIR / "models" / "checkpoints")
-        return 2
-
-    from coreml_suite.core.naming import compose_out_name
-
-    out_name = compose_out_name(
-        ckpt_name=ckpt_name,
-        batch_size=batch_size,
-        width=width,
-        height=height,
-        controlnet_support=controlnet,
-        attention_implementation=attn,
-        quantize_nbits=quantize_nbits,
-    )
-    unet_out_path = converter.get_out_path("unet", out_name)
-
-    log.info("repo_root=%s comfy_dir=%s", REPO_ROOT, COMFY_DIR)
-    log.info("ckpt=%s out_name=%s", ckpt_path, out_name)
-    log.info(
-        "attn=%s size=%dx%d batch=%d controlnet=%s quant=%s",
-        attn, width, height, batch_size, controlnet, quantize_nbits,
-    )
-
-    converter.convert(
-        ckpt_path=ckpt_path,
-        model_version=ModelVersion.SD15,
-        unet_out_path=unet_out_path,
-        batch_size=batch_size,
-        sample_size=(height // 8, width // 8),
-        controlnet_support=controlnet,
-        lora_weights=[],
-        attn_impl=attn,
-        config_path=None,
-        quantize_nbits=quantize_nbits,
-    )
-
-    target_path = converter.compile_model(out_path=unet_out_path, out_name=out_name, submodule_name="unet")
-    log.info("compiled: %s", target_path)
-    print(target_path)
-    return 0
-
-
-if __name__ == "__main__":
-    raise SystemExit(main())
diff --git a/bench/scripts/quant_matrix.py b/bench/scripts/quant_matrix.py
deleted file mode 100644
index d35456a..0000000
--- a/bench/scripts/quant_matrix.py
+++ /dev/null
@@ -1,188 +0,0 @@
-#!/usr/bin/env python3
-"""bench/scripts/quant_matrix.py — Phase 6 quantization tradeoff matrix.
-
-For each variant in {none, 8, 6, 4}, load the corresponding .mlmodelc,
-run a synthetic forward pass with a fixed seed, and record:
-  - on-disk size
-  - load time
-  - forward-pass median latency
-  - PSNR of noise_pred vs the `none` baseline (proxy for output drift)
-
-Writes a JSON + Markdown summary into bench/results/quant_matrix_<sha>.{json,md}.
-
-Assumes converted .mlmodelc files exist under $COMFY_DIR/models/unet/
-following the convention from coreml_suite.core.naming.compose_out_name
-(e.g. v1-5-pruned-emaonly_1x512x512_se_unet.mlmodelc for the baseline,
-v1-5-pruned-emaonly_1x512x512_se_q4_unet.mlmodelc for 4-bit, etc.).
-"""
-import json
-import os
-import statistics
-import subprocess
-import sys
-import time
-from pathlib import Path
-from typing import Any
-
-import numpy as np
-
-REPO_ROOT = Path(__file__).resolve().parents[2]
-COMFY_DIR = Path(os.environ.get("COMFY_DIR", REPO_ROOT.parents[1])).resolve()
-MODELS_DIR = COMFY_DIR / "models" / "unet"
-
-VARIANTS = ["none", "8", "6", "4"]
-CKPT_STEM = os.environ.get("CKPT_STEM", "v1-5-pruned-emaonly")
-BATCH = int(os.environ.get("BATCH", "1"))
-WIDTH = int(os.environ.get("WIDTH", "512"))
-HEIGHT = int(os.environ.get("HEIGHT", "512"))
-ATTN_SUFFIX = os.environ.get("ATTN_SUFFIX", "se")
-COMPUTE_UNIT = os.environ.get("COMPUTE_UNIT", "CPU_AND_NE")
-REPEATS = int(os.environ.get("REPEATS", "20"))
-INPUT_SEED = int(os.environ.get("INPUT_SEED", "0"))
-
-
-def variant_path(nbits: str) -> Path:
-    quant_suffix = f"_q{nbits}" if nbits != "none" else ""
-    name = f"{CKPT_STEM}_{BATCH}x{WIDTH}x{HEIGHT}_{ATTN_SUFFIX}{quant_suffix}_unet.mlmodelc"
-    return MODELS_DIR / name
-
-
-def dir_size_bytes(path: Path) -> int:
-    total = 0
-    for root, _dirs, files in os.walk(path):
-        for f in files:
-            try:
-                total += (Path(root) / f).stat().st_size
-            except OSError:
-                pass
-    return total
-
-
-def psnr(a: np.ndarray, b: np.ndarray) -> float:
-    """Float PSNR for noise_pred tensors normalized to a [-1, 1]-ish range."""
-    a = a.astype(np.float64)
-    b = b.astype(np.float64)
-    peak = max(float(np.max(np.abs(a))), float(np.max(np.abs(b))), 1.0)
-    mse = float(np.mean((a - b) ** 2))
-    if mse == 0:
-        return 100.0
-    return 20.0 * float(np.log10(peak / np.sqrt(mse)))
-
-
-def measure_variant(
-    nbits: str, ref_output: "np.ndarray | None"
-) -> "tuple[dict[str, Any], np.ndarray | None]":
-    from python_coreml_stable_diffusion.coreml_model import CoreMLModel
-
-    path = variant_path(nbits)
-    result: dict[str, Any] = {"nbits": nbits, "path": str(path)}
-    if not path.exists():
-        result["error"] = f"model not found at {path}"
-        return result, None
-    result["size_bytes"] = dir_size_bytes(path)
-
-    rng = np.random.default_rng(INPUT_SEED)
-    t0 = time.perf_counter()
-    model = CoreMLModel(str(path), COMPUTE_UNIT, "compiled")
-    result["load_time_s"] = round(time.perf_counter() - t0, 4)
-
-    expected = dict(model.expected_inputs)
-    fixed_inputs = {
-        name: rng.standard_normal(tuple(int(d) for d in spec["shape"])).astype(np.float16)
-        for name, spec in expected.items()
-    }
-
-    # Warmup
-    out = model(**fixed_inputs)
-    times = []
-    for _ in range(REPEATS):
-        # Fresh randomness per repeat for steady-state timing, but keep
-        # output capture deterministic (we run fixed_inputs at end).
-        live = {
-            name: rng.standard_normal(tuple(int(d) for d in spec["shape"])).astype(np.float16)
-            for name, spec in expected.items()
-        }
-        t0 = time.perf_counter()
-        model(**live)
-        times.append((time.perf_counter() - t0) * 1000.0)
-    result["fwd_ms_median"] = round(statistics.median(times), 3)
-    result["fwd_ms_min"] = round(min(times), 3)
-
-    # Deterministic forward for PSNR comparison across variants.
-    deterministic_out = model(**fixed_inputs)["noise_pred"]
-    if ref_output is None:
-        result["psnr_db_vs_none"] = None
-    else:
-        result["psnr_db_vs_none"] = round(psnr(ref_output, deterministic_out), 2)
-    return result, deterministic_out
-
-
-def main() -> int:
-    try:
-        sha = subprocess.check_output(
-            ["git", "-C", str(REPO_ROOT), "rev-parse", "--short", "HEAD"],
-            stderr=subprocess.DEVNULL,
-        ).decode().strip()
-    except Exception:
-        sha = "nogit"
-
-    out_dir = REPO_ROOT / "bench" / "results"
-    out_dir.mkdir(parents=True, exist_ok=True)
-
-    print(f"[quant-matrix] git={sha} compute_unit={COMPUTE_UNIT} repeats={REPEATS}", file=sys.stderr)
-
-    rows: list[dict[str, Any]] = []
-    ref_output: np.ndarray | None = None
-    for v in VARIANTS:
-        print(f"[quant-matrix] measuring nbits={v} ...", file=sys.stderr)
-        try:
-            row, out_tensor = measure_variant(v, ref_output)
-            if v == "none":
-                ref_output = out_tensor
-            rows.append(row)
-        except Exception as exc:
-            rows.append({"nbits": v, "error": f"{type(exc).__name__}: {exc}"})
-
-    baseline_size = next((r["size_bytes"] for r in rows if r.get("nbits") == "none" and "size_bytes" in r), None)
-    for r in rows:
-        if "size_bytes" in r and baseline_size:
-            r["size_ratio_vs_none"] = round(r["size_bytes"] / baseline_size, 3)
-
-    report = {
-        "git_sha": sha,
-        "compute_unit": COMPUTE_UNIT,
-        "repeats": REPEATS,
-        "input_seed": INPUT_SEED,
-        "ckpt_stem": CKPT_STEM,
-        "rows": rows,
-    }
-    (out_dir / f"quant_matrix_{sha}.json").write_text(json.dumps(report, indent=2))
-
-    lines = [
-        f"# Quantization tradeoff matrix — `{sha}`",
-        "",
-        f"- compute unit: {COMPUTE_UNIT}",
-        f"- repeats per variant: {REPEATS}",
-        f"- ckpt: {CKPT_STEM}, {BATCH}x{WIDTH}x{HEIGHT}, attn={ATTN_SUFFIX}",
-        "",
-        "| nbits | size (MB) | size vs none | load (s) | fwd median (ms) | fwd min (ms) | PSNR vs none (dB) | error |",
-        "|---|---|---|---|---|---|---|---|",
-    ]
-    for r in rows:
-        if "error" in r and "size_bytes" not in r:
-            lines.append(f"| {r.get('nbits','?')} | - | - | - | - | - | - | {r['error']} |")
-            continue
-        size_mb = round(r.get("size_bytes", 0) / (1024 * 1024), 1)
-        lines.append(
-            f"| {r['nbits']} | {size_mb} | {r.get('size_ratio_vs_none','-')} | "
-            f"{r.get('load_time_s','-')} | {r.get('fwd_ms_median','-')} | "
-            f"{r.get('fwd_ms_min','-')} | {r.get('psnr_db_vs_none') if r.get('psnr_db_vs_none') is not None else '-'} | "
-            f"{r.get('error','')} |"
-        )
-    (out_dir / f"quant_matrix_{sha}.md").write_text("\n".join(lines) + "\n")
-    print(f"[quant-matrix] wrote {out_dir / f'quant_matrix_{sha}.md'}", file=sys.stderr)
-    return 0
-
-
-if __name__ == "__main__":
-    raise SystemExit(main())
diff --git a/bench/scripts/smoke_image.py b/bench/scripts/smoke_image.py
deleted file mode 100644
index 74fa1fa..0000000
--- a/bench/scripts/smoke_image.py
+++ /dev/null
@@ -1,163 +0,0 @@
-#!/usr/bin/env python3
-"""bench/scripts/smoke_image.py — Phase 1 known-good-workflow image smoke test.
-
-Posts a Stable Diffusion 1.5 workflow (CLIP → Core ML UNet → VAE → PNG) to a
-locally running ComfyUI server, waits for the queue to drain, then copies the
-generated CoreML + MPS reference images into bench/results/smoke/<gitsha>/
-and writes a small report.json with the PSNR between them. This is the Phase 1
-"known-good workflow still produces an image [M2-ANE]" artifact.
-
-Prereqs:
-  - ComfyUI server running locally on $COMFY_HOST:$COMFY_PORT (default
-    http://localhost:8188), started against the project .venv so it picks up
-    the pinned ml-stable-diffusion + Core ML Suite nodes.
-  - The Core ML UNet has already been converted (see convert_sd15.py); the
-    workflow's Core ML Converter node will skip if the .mlmodelc exists.
-
-Run from the repo root:
-    .venv/bin/python bench/scripts/smoke_image.py
-"""
-import json
-import os
-import shutil
-import subprocess
-import sys
-import time
-import urllib.error
-import urllib.request
-from pathlib import Path
-
-import numpy as np
-from PIL import Image
-
-REPO_ROOT = Path(__file__).resolve().parents[2]
-COMFY_DIR = Path(os.environ.get("COMFY_DIR", REPO_ROOT.parents[1])).resolve()
-COMFY_HOST = os.environ.get("COMFY_HOST", "localhost")
-COMFY_PORT = int(os.environ.get("COMFY_PORT", "8188"))
-COMFY_URL = f"http://{COMFY_HOST}:{COMFY_PORT}"
-
-CKPT_NAME = os.environ.get("CKPT_NAME", "v1-5-pruned-emaonly.safetensors")
-WORKFLOW_PATH = Path(os.environ.get(
-    "WORKFLOW_PATH",
-    REPO_ROOT / "tests" / "integration" / "workflows" / "e2e-1.5-basic-conversion.json",
-))
-TIMEOUT_S = int(os.environ.get("TIMEOUT_S", "1800"))  # 30min cap
-
-
-def git_sha() -> str:
-    try:
-        return subprocess.check_output(
-            ["git", "-C", str(REPO_ROOT), "rev-parse", "--short", "HEAD"],
-            stderr=subprocess.DEVNULL,
-        ).decode().strip()
-    except Exception:
-        return "nogit"
-
-
-def http_get_json(path: str) -> dict:
-    with urllib.request.urlopen(f"{COMFY_URL}{path}", timeout=60) as r:
-        return json.loads(r.read().decode())
-
-
-def http_post_json(path: str, payload: dict) -> dict:
-    data = json.dumps(payload).encode("utf-8")
-    req = urllib.request.Request(
-        f"{COMFY_URL}{path}", data=data,
-        headers={"Content-Type": "application/json"}, method="POST",
-    )
-    # Long timeout: first POST blocks while the server warms model loaders +
-    # the Core ML compile-on-load pass (~60-90s on a cold cache).
-    with urllib.request.urlopen(req, timeout=300) as r:
-        return json.loads(r.read().decode())
-
-
-def psnr(a: np.ndarray, b: np.ndarray) -> float:
-    mse = float(np.mean((a.astype(np.float64) - b.astype(np.float64)) ** 2))
-    if mse == 0:
-        return 100.0
-    return 20.0 * float(np.log10(255.0 / np.sqrt(mse)))
-
-
-def find_image(out_dir: Path, prefix: str) -> Path | None:
-    matches = sorted(out_dir.glob(f"{prefix}_*.png"), reverse=True)
-    return matches[0] if matches else None
-
-
-def main() -> int:
-    sha = git_sha()
-    out_root = REPO_ROOT / "bench" / "results" / "smoke" / sha
-    out_root.mkdir(parents=True, exist_ok=True)
-
-    workflow = json.loads(WORKFLOW_PATH.read_text())
-    # Point both checkpoint nodes at the maintainer's available SD1.5 ckpt.
-    for nid in ("4", "10"):
-        if nid in workflow:
-            workflow[nid]["inputs"]["ckpt_name"] = CKPT_NAME
-    # Fixed seed for reproducibility (Phase 1 baseline).
-    seed = int(os.environ.get("SEED", "42"))
-    for nid in ("3", "11"):
-        if nid in workflow and "seed" in workflow[nid].get("inputs", {}):
-            workflow[nid]["inputs"]["seed"] = seed
-    # Phase 1 only needs a Core ML image. The reference MPS branch (nodes
-    # 3/8/9) trips a known MPS f16/f32 mismatch on torch 2.0.1 + macOS 26+,
-    # which would crash the whole server. Strip those nodes so the queue
-    # only runs the Core ML path; SKIP_MPS=0 to opt back in.
-    if os.environ.get("SKIP_MPS", "1") not in ("0", "false", "False"):
-        for nid in ("3", "8", "9"):
-            workflow.pop(nid, None)
-
-    print(f"[smoke] git={sha} server={COMFY_URL} ckpt={CKPT_NAME} seed={seed}", file=sys.stderr)
-
-    resp = http_post_json("/prompt", {"prompt": workflow})
-    print(f"[smoke] queued: {resp}", file=sys.stderr)
-
-    t0 = time.time()
-    while True:
-        if time.time() - t0 > TIMEOUT_S:
-            print(f"[smoke] TIMEOUT after {TIMEOUT_S}s waiting for queue drain", file=sys.stderr)
-            return 2
-        try:
-            q = http_get_json("/prompt")
-            remaining = q.get("exec_info", {}).get("queue_remaining", -1)
-            if remaining == 0:
-                break
-        except (urllib.error.URLError, json.JSONDecodeError) as exc:
-            print(f"[smoke] poll error: {exc}", file=sys.stderr)
-        time.sleep(2)
-
-    comfy_out = COMFY_DIR / "output"
-    coreml_png = find_image(comfy_out, "E2E-1.5-CoreML")
-    if coreml_png is None:
-        print(f"[smoke] missing Core ML image under {comfy_out}", file=sys.stderr)
-        return 3
-    coreml_dst = out_root / coreml_png.name
-    shutil.copy2(coreml_png, coreml_dst)
-
-    mps_png = find_image(comfy_out, "E2E-1.5-MPS")
-    mps_dst = None
-    psnr_db = None
-    if mps_png is not None:
-        mps_dst = out_root / mps_png.name
-        shutil.copy2(mps_png, mps_dst)
-        a = np.array(Image.open(coreml_dst).convert("RGB"))
-        b = np.array(Image.open(mps_dst).convert("RGB"))
-        if a.shape == b.shape:
-            psnr_db = round(psnr(a, b), 2)
-
-    report = {
-        "git_sha": sha,
-        "seed": seed,
-        "ckpt_name": CKPT_NAME,
-        "workflow": str(WORKFLOW_PATH.relative_to(REPO_ROOT)),
-        "coreml_image": str(coreml_dst.relative_to(REPO_ROOT)),
-        "mps_image": str(mps_dst.relative_to(REPO_ROOT)) if mps_dst else None,
-        "psnr_db": psnr_db,
-        "wall_seconds": round(time.time() - t0, 1),
-    }
-    (out_root / "report.json").write_text(json.dumps(report, indent=2))
-    print(json.dumps(report, indent=2))
-    return 0
-
-
-if __name__ == "__main__":
-    raise SystemExit(main())
diff --git a/constraints/comfy-ceiling.txt b/constraints/comfy-ceiling.txt
index c582dec..013d2b6 100644
--- a/constraints/comfy-ceiling.txt
+++ b/constraints/comfy-ceiling.txt
@@ -3,14 +3,14 @@
 #
 # ComfyUI's requirements.txt requests bare `torch`/`torchvision`/`torchaudio`
 # and `numpy>=1.25.0`, which would float past the versions coremltools 9 /
-# apple-ml-stable-diffusion have been validated against (see docs/deps.md).
+# apple-ml-stable-diffusion have been validated against.
 # These constraints cap the resolution so the canary keeps testing the same
 # toolchain the suite actually ships.
 #
 # If upstream ComfyUI ever hard-requires something beyond these bounds, the
 # install FAILS — and that failure is the signal we want: it means the host
 # outgrew the pinned toolchain and coremltools / ml-stable-diffusion need a
-# deliberate bump (a Phase 5-style upgrade), not a silent float.
+# deliberate bump, not a silent float.
 torch>=2.7,<2.8
 torchvision>=0.22,<0.23
 torchaudio>=2.7,<2.8
diff --git a/coreml_suite/controlnet.py b/coreml_suite/controlnet.py
index 141b471..dbac4c4 100644
--- a/coreml_suite/controlnet.py
+++ b/coreml_suite/controlnet.py
@@ -1,4 +1,4 @@
-"""Phase 3 compatibility shim — re-exports from coreml_suite.core.controlnet."""
+"""Compatibility shim — re-exports from coreml_suite.core.controlnet."""
 from coreml_suite.core.controlnet import (
     chunk_control,
     expand_inputs,
diff --git a/coreml_suite/converter.py b/coreml_suite/converter.py
index c2a5b55..1af74cd 100644
--- a/coreml_suite/converter.py
+++ b/coreml_suite/converter.py
@@ -307,8 +307,8 @@ def convert_unet(
     gc.collect()
 
     if quantize_nbits != "none":
-        # Phase 6: opt-in k-means weight palettization. Default path
-        # (quantize_nbits="none") is byte-for-byte unchanged from Phase 5.
+        # Opt-in k-means weight palettization. The default path
+        # (quantize_nbits="none") leaves the traced UNet untouched.
         from coremltools.optimize.coreml import (
             OpPalettizerConfig,
             OptimizationConfig,
diff --git a/coreml_suite/core/controlnet.py b/coreml_suite/core/controlnet.py
index 0f063d3..72a1ba5 100644
--- a/coreml_suite/core/controlnet.py
+++ b/coreml_suite/core/controlnet.py
@@ -1,7 +1,7 @@
 """Pure helpers around the ControlNet residual inputs of the Core ML UNet.
 
-Moved from coreml_suite.controlnet in Phase 3. Behavior unchanged — Phase 2
-characterization tests cover shapes, dtype (fp16), and zero-fill fallback.
+Re-exported by coreml_suite.controlnet. Characterization tests cover
+shapes, dtype (fp16), and zero-fill fallback.
 """
 from itertools import chain
 from math import ceil
diff --git a/coreml_suite/core/inputs.py b/coreml_suite/core/inputs.py
index 9263a47..b875c5a 100644
--- a/coreml_suite/core/inputs.py
+++ b/coreml_suite/core/inputs.py
@@ -1,7 +1,6 @@
 """Pure transform from torch sampler inputs to Core ML UNet kwargs.
 
-Moved from coreml_suite.models in Phase 3. Behavior unchanged — Phase 2
-characterization tests cover SD1.5 / SDXL base / SDXL refiner / LCM
+Characterization tests cover SD1.5 / SDXL base / SDXL refiner / LCM
 variants and the chunked-batch fan-out.
 """
 import numpy as np
diff --git a/coreml_suite/core/latents.py b/coreml_suite/core/latents.py
index 58f72ce..d6ad528 100644
--- a/coreml_suite/core/latents.py
+++ b/coreml_suite/core/latents.py
@@ -1,8 +1,8 @@
 """Pure batch-chunking helpers for Core ML's fixed-shape UNet inputs.
 
-Moved from coreml_suite.latents in Phase 3. Behavior unchanged — Phase 2
-characterization tests cover the contract (padding-zero regions, truncation
-in merge_chunks, identity-passthrough when shape already matches).
+Re-exported by coreml_suite.latents. Characterization tests cover the
+contract (padding-zero regions, truncation in merge_chunks,
+identity-passthrough when shape already matches).
 """
 import torch
 
diff --git a/coreml_suite/core/naming.py b/coreml_suite/core/naming.py
index 16f2024..71d186d 100644
--- a/coreml_suite/core/naming.py
+++ b/coreml_suite/core/naming.py
@@ -1,6 +1,6 @@
 """Pure out_name composition for the Core ML UNet artifact.
 
-Extracted from CoreMLConverter.convert in Phase 3 so the filename contract
+Extracted from CoreMLConverter.convert so the filename contract
 can be tested + reused without instantiating the node. The string is the
 cache key: every workflow that references a converted .mlpackage depends
 on it staying byte-for-byte identical.
@@ -13,8 +13,8 @@ ATTN_SUFFIX = {
     "ORIGINAL": "orig",
 }
 
-# Phase 6: palettization bits. "none" = no quantization (default; keeps the
-# pre-Phase-6 filename intact so existing workflows still resolve their
+# Palettization bits. "none" = no quantization (default; keeps the
+# unquantized filename intact so existing workflows still resolve their
 # cached .mlpackage). Numeric values append a `_q<bits>` suffix.
 QUANT_NBITS_VALUES = ("none", "8", "6", "4")
 
@@ -32,7 +32,7 @@ def compose_out_name(
 ) -> str:
     """Build the .mlpackage stem from convert() parameters.
 
-    Locked behaviour (Phase 2 characterization tests):
+    Locked behaviour (characterization tests):
       - first '.' in ckpt_name wins (`a.b.c.safetensors` -> `a`)
       - spaces collapse to underscores
       - LoRA names are taken stem-only, sorted, joined with '_' and
@@ -41,7 +41,7 @@ def compose_out_name(
       - controlnet adds `_cn`
       - attn suffix is `_se` | `_se2` | `_orig`
 
-    Phase 6 addition:
+    Quantization:
       - quantize_nbits "none" (default) appends nothing — existing
         unquantized .mlpackages keep the old filename
       - "4" / "6" / "8" appends `_q<bits>` after the attn suffix
diff --git a/coreml_suite/core/sdxl.py b/coreml_suite/core/sdxl.py
index 7ea16eb..fd1e8d2 100644
--- a/coreml_suite/core/sdxl.py
+++ b/coreml_suite/core/sdxl.py
@@ -1,9 +1,9 @@
 """Pure SDXL detection + time_ids/text_embeds assembly.
 
-Moved from coreml_suite.models in Phase 3. The framework-coupled adapter
-`add_sdxl_model_options` lives in models.py and now delegates the math
-here. Phase 2 characterization tests cover base (len 6) vs refiner
-(len 5) and the closure free-vars produced by `sdxl_model_function_wrapper`.
+The framework-coupled adapter `add_sdxl_model_options` lives in models.py
+and delegates the math here. Characterization tests cover base (len 6) vs
+refiner (len 5) and the closure free-vars produced by
+`sdxl_model_function_wrapper`.
 """
 import torch
 
diff --git a/coreml_suite/latents.py b/coreml_suite/latents.py
index d2896df..b3b55d6 100644
--- a/coreml_suite/latents.py
+++ b/coreml_suite/latents.py
@@ -1,4 +1,4 @@
-"""Phase 3 compatibility shim — re-exports from coreml_suite.core.latents."""
+"""Compatibility shim — re-exports from coreml_suite.core.latents."""
 from coreml_suite.core.latents import chunk_batch, merge_chunks
 
 __all__ = ["chunk_batch", "merge_chunks"]
diff --git a/coreml_suite/models.py b/coreml_suite/models.py
index f4bc360..358d7b1 100644
--- a/coreml_suite/models.py
+++ b/coreml_suite/models.py
@@ -1,7 +1,7 @@
 """Framework-coupled glue between Core ML UNets and ComfyUI's sampler stack.
 
 Pure math (CoreMLInputs, SDXL detection, time_ids/text_embeds assembly,
-sdxl_model_function_wrapper) lives in coreml_suite.core.* after Phase 3.
+sdxl_model_function_wrapper) lives in coreml_suite.core.*.
 This module is what touches comfy.*: model_base, ModelPatcher, the
 diffusion_model wrapper, and the maintainer-facing add_sdxl_model_options
 adapter.
diff --git a/coreml_suite/nodes.py b/coreml_suite/nodes.py
index 96f1d78..00025fc 100644
--- a/coreml_suite/nodes.py
+++ b/coreml_suite/nodes.py
@@ -249,12 +249,11 @@ class CoreMLConverter(COREML_NODE):
                 "controlnet_support": ("BOOLEAN", {"default": False}),
             },
             "optional": {
-                # Phase 6: k-means weight palettization. Kept optional so
-                # workflows saved before Phase 6 still validate — ComfyUI
-                # rejects a prompt that omits any `required` input. When
-                # omitted it defaults to "none", identical to pre-Phase-6
-                # behavior and filename, so existing cached .mlpackages
-                # still resolve.
+                # k-means weight palettization. Kept optional so workflows
+                # that omit it still validate — ComfyUI rejects a prompt that
+                # omits any `required` input. When omitted it defaults to
+                # "none", identical to unquantized behavior and filename, so
+                # existing cached .mlpackages still resolve.
                 "quantize_nbits": (list(QUANT_NBITS_VALUES), {"default": "none"}),
                 "lora_params": ("LORA_PARAMS",),
             },
diff --git a/docs/ci-m2.md b/docs/ci-m2.md
deleted file mode 100644
index dee42cc..0000000
--- a/docs/ci-m2.md
+++ /dev/null
@@ -1,137 +0,0 @@
-# Self-hosted M2 runner — setup
-
-Tier 2 (ANE + integration + bench) runs on a self-hosted GitHub Actions
-runner registered against the maintainer's M-series Mac. Hosted macOS
-runners on GitHub do not expose the Apple Neural Engine, so the ANE
-half of the matrix has to live on real hardware.
-
-## One-time runner setup
-
-1. **Install dependencies on the Mac.** Python 3.12.x (matching the
-   `requires-python = ">=3.12,<3.13"` pin), `uv`, and `git`. The workflow
-   clones and manages the ComfyUI checkout itself (see step 3), so you do
-   not pre-install ComfyUI.
-
-   ```bash
-   brew install python@3.12 uv git
-   ```
-
-2. **Register the runner.** From the repo Settings → Actions → Runners
-   → New self-hosted runner, follow the macOS-ARM instructions. Add the
-   labels exactly: `self-hosted`, `macOS`, `ARM64`, `coreml` (the
-   workflow `runs-on` clause requires all four).
-
-   ```bash
-   mkdir ~/actions-runner && cd ~/actions-runner
-   curl -O -L https://github.com/actions/runner/releases/download/v2.317.0/actions-runner-osx-arm64-2.317.0.tar.gz
-   tar xzf actions-runner-osx-arm64-2.317.0.tar.gz
-   ./config.sh --url https://github.com/<owner>/<repo> \
-               --token <REGISTRATION_TOKEN> \
-               --labels self-hosted,macOS,ARM64,coreml \
-               --name "$(hostname)-m2"
-   ./svc.sh install && ./svc.sh start   # run as a launchd service
-   ```
-
-3. **Persist `COMFY_DIR` for the runner.** Point it at a **dedicated,
-   runner-owned** ComfyUI directory — **not** your personal dev checkout.
-   The workflow does `git reset --hard` on it and rewrites its
-   `custom_nodes/ComfyUI-CoreMLSuite` symlink, so it must be disposable.
-   Add it to the runner's `.env`:
-
-   ```bash
-   echo 'COMFY_DIR=/Users/<you>/actions-runner/comfyui' >> ~/actions-runner/.env
-   ```
-
-   You don't have to clone ComfyUI yourself: the `Set up ComfyUI checkout`
-   step clones it on first run, checks out the right ref (latest or the
-   pinned SHA — see *ComfyUI version under test*), installs ComfyUI's deps,
-   and symlinks `$COMFY_DIR/custom_nodes/ComfyUI-CoreMLSuite` →
-   `$GITHUB_WORKSPACE` so the server always loads the checked-out PR. If
-   `COMFY_DIR` is a real node directory rather than a symlink, the step
-   fails loudly instead of deleting it.
-
-4. **Provide the SD1.5 checkpoint.** Drop
-   `v1-5-pruned-emaonly.safetensors` (~4 GB) into
-   `$COMFY_DIR/models/checkpoints/`. This is the one heavy artifact that
-   stays as a runner-local cache; everything derived from it (the
-   `.mlmodelc` UNet variants) is converted automatically by the
-   `Convert UNet variants if missing` step and cached across runs. Override
-   the filename with `CKPT_NAME` in the runner `.env` if needed.
-
-   Order is free: do this before or after the first run. The setup step
-   initialises the ComfyUI repo **in place** (`git init`, not `git clone`,
-   which would refuse a non-empty directory) and uses `checkout -f`, which
-   never removes untracked files — so a checkpoint you dropped in first is
-   preserved.
-
-## Triggers
-
-The Tier 2 workflow (`.github/workflows/tier2.yml`) runs:
-
-- **On PR label `run-m2`** — maintainers add the label to opt a PR
-  into the ANE lane (the runner is not free; default off).
-- **Nightly at 04:00 UTC** via `schedule:`.
-- **Manually** via the workflow_dispatch button.
-
-## ComfyUI version under test
-
-Tier 2 runs a **hybrid** strategy, keyed on the trigger, so the suite tracks a
-moving host without making PRs flaky to overnight upstream drift:
-
-| Trigger | ComfyUI ref | ComfyUI deps |
-|---|---|---|
-| **schedule** (nightly) | latest `origin/master` | its own `requirements.txt`, capped by `constraints/comfy-ceiling.txt` |
-| **PR label `run-m2`** / **dispatch** | pinned `requires-comfyui` SHA | the frozen `comfy` uv group |
-
-- **Nightly = canary.** It pulls the latest ComfyUI and installs *ComfyUI's
-  own* dependency set. The frozen `comfy` uv group cannot track a moving host
-  by hand (a latest checkout needs e.g. `comfyui-frontend-package==1.44.19`,
-  `comfy_aimdo`, `alembic`, `blake3` that an older pin never listed), so latest
-  mode defers to upstream's `requirements.txt`. Upstream API or dependency
-  breakage surfaces here, in CI, instead of in a user's install.
-- **PR / dispatch = reproducible gate.** It checks out the pinned
-  `requires-comfyui` SHA and uses the frozen `comfy` group — the known-good
-  combination — so a PR fails for its own reasons, not because ComfyUI moved.
-
-The resolved ComfyUI SHA and mode are written to the job step summary (and
-`COMFY_SHA` in the env), so any failure names the exact commit it hit.
-
-**The toolchain ceiling is deliberate.** `constraints/comfy-ceiling.txt` caps
-`torch<2.8` / `numpy<2` / `coremltools 9` while installing latest ComfyUI's
-requirements. If upstream ever hard-requires something past those bounds the
-nightly install *fails on purpose* — that is the signal that coremltools /
-`ml-stable-diffusion` need a deliberate Phase-5-style bump, not a silent float
-that would break the ANE path (see `docs/deps.md`).
-
-`pyproject.toml`'s `requires-comfyui` is both the PR-gate ref and the
-published-compatibility declaration for the Comfy registry. Bump it once a
-newer ComfyUI is validated (the nightly canary is what tells you it's safe).
-
-## Artifacts
-
-- Bench JSON/MD are uploaded as `bench-results`.
-- M2 golden image regressions surface as test failures in
-  `tests/m2/test_golden_image.py`; diff PNG is written next to the
-  golden under `tests/m2/_latest_generated.png` (gitignored).
-
-## When the runner is down
-
-If the maintainer's Mac is offline, the workflow queues until the
-runner comes back. Cancel a stuck run from the Actions UI; the gate is
-not blocking by default (Tier 0 + Tier 1 carry PR status). Tier 2 is
-"good to merge once it goes green," not "blocked until then."
-
-## Replacing the integration e2e
-
-The legacy `tests/integration/test_basic_conversion_1_5.py` checked
-CoreML output against an MPS reference image at PSNR > 25 dB. That
-reference path is broken on macOS 26.x with torch 2.0.1 (see Phase 1
-Gate report). Phase 4 moves the same coverage to
-`tests/m2/test_golden_image.py`, which:
-
-- runs the Core ML pipeline only (no MPS reference),
-- asserts SHA256 against `tests/m2/goldens/sd15_seed42.sha256`,
-- falls back to PSNR ≥ 40 dB if the hash drifts.
-
-This removes the human-eyeball dependency: a regression is now a
-numerical fail, not a "looks different to me."
diff --git a/docs/deps.md b/docs/deps.md
deleted file mode 100644
index 3606707..0000000
--- a/docs/deps.md
+++ /dev/null
@@ -1,163 +0,0 @@
-# Dependency decisions
-
-## Phase 5: ml-stable-diffusion compatibility under the bumped toolchain
-
-### Scope
-
-Bumped: Python 3.11 → 3.12, torch 2.0.1 → 2.7.1, coremltools 8.2 → 9.0.
-**Not bumped:** numpy stays in the 1.24..1.x range. Phase 5 spec asked
-for numpy 2 as a goal; in practice the bump is decoupled — see
-"Why not numpy 2" below.
-
-### Problem
-
-`python_coreml_stable_diffusion` pins three transitive dependencies in
-its `setup.py` that block any modern install:
-
-```python
-install_requires=[
-    "coremltools>=8.0",
-    "diffusers[torch]==0.30.2",
-    "transformers==4.44.2",
-    "huggingface-hub==0.24.6",
-    "numpy<1.24",
-    ...
-]
-```
-
-The exact-pin lines on diffusers/transformers/huggingface-hub block the
-Python-3.12 / torch-2.7 / coremltools-9 combo. The numpy ceiling at
-1.24 also blocks newer numpy.
-
-The upstream `main` branch as of 2026-05-24 (commit `e12202c1f`) has
-the same pins as our Phase 1 `e5d960c4` SHA — no working alternative
-exists upstream. No maintained fork on PyPI either.
-
-### Decision
-
-**Keep the SHA pinned at `e5d960c4`** (Phase 1 baseline) and apply
-`[tool.uv] override-dependencies` to relax the four blocking pins:
-
-```toml
-[tool.uv]
-override-dependencies = [
-    "numpy>=1.24,<2",
-    "diffusers>=0.30",
-    "transformers>=4.44",
-    "huggingface-hub>=0.24",
-]
-```
-
-### Why not bump the SHA, vendor the code, or fork?
-
-- **Bump SHA:** the only newer SHA on `main` (`e12202c1f`) has identical
-  pins. No upstream fix.
-- **Vendor the imports:** our code only uses `unet.UNet2DConditionModel`,
-  `unet.UNet2DConditionModelXL`, `unet.AttentionImplementations`,
-  `unet.calculate_conv2d_output_shape`, and `coreml_model.CoreMLModel`.
-  Copying these into `coreml_suite/_vendor/` would work but adds
-  hundreds of lines of code, owns a maintenance burden we don't want
-  yet, and detaches us from upstream bug fixes.
-- **Public fork:** would also need maintenance + CI to keep current with
-  Apple's main.
-
-The override is the smallest workable patch: it trusts that the API
-surface we touch (`unet.*` + `coreml_model.CoreMLModel`) is stable
-across the bumped versions — verified empirically by Tier 1 (synthetic
-UNet round-trip through `ct.convert`) and Tier 2 (real SD1.5 conversion
-+ golden image PSNR ≥ 25 dB).
-
-### Why not numpy 2
-
-The Phase 5 spec asked for numpy 2 alongside the Python / coremltools /
-torch bumps. Tier 0 + Tier 1 are happy on numpy 2 — none of our own
-modules touch numpy in a way that breaks. The SD UNet conversion path
-is not:
-
-1. `coremltools 9.0 (and 8.3) `_cast(int)` in
-   `converters/mil/frontend/torch/ops.py` does
-   `dtype(x.val)` where `x.val` is a numpy ndarray with a single
-   element. Under numpy 2 that raises `TypeError: only 0-dimensional
-   arrays can be converted to Python scalars`. Fixable via a
-   `np.ndarray.item()` shim.
-2. After patching (1), `view` then trips
-   `mb.cast(x=shape, dtype="int32")` where `shape` is a Python list of
-   non-scalar `Var`s — a code path the upstream guard
-   `all([isinstance(dim, Var) and len(dim.shape) == 0 for dim in shape])`
-   doesn't cover. Different bug class; needs a separate, larger shim
-   on the `view` op (and probably more elsewhere — each patch reveals
-   the next).
-
-Neither bug is a Phase-5 deliverable. None of our code paths require
-numpy 2. **Decision: hold numpy at >=1.24,<2 and unblock the rest of
-the bump.** numpy 2 stays as an explicit follow-up once coremltools
-ships an upstream fix (or we sign up for the shim work).
-
-### Tier 2 PSNR threshold
-
-Phase 2 anchored the golden image as a SHA256 + 40 dB PSNR fallback.
-The toolchain bump changes the bit-exact output: the SD1.5 baseline at
-seed=42 lands at ~29 dB against the Phase 2 PNG. The image is
-visually identical (same composition, same colors, sub-pixel drift),
-just numerically different — typical for a coremltools/torch upgrade.
-
-`tests/m2/test_golden_image.py` was bumped to a 25 dB threshold and the
-golden anchor was re-captured against the bumped toolchain. Refactor PRs
-(Phase 3-style "should not change math") should raise the threshold via
-`GOLDEN_PSNR_MIN_DB=40` (or higher); future toolchain bumps can repeat
-the Phase 5 dance and recapture.
-
-### Bench diff vs Phase 1
-
-| metric | Phase 1 (ct8.2/np1.23/py3.11) | Phase 5 (ct9.0/np1.26/py3.12/torch2.7) | delta |
-|---|---:|---:|---:|
-| SD1.5 NE fwd median (ms) | 196.97 | 197.32 | +0.2% |
-| SD1.5 GPU fwd median (ms) | 270.26 | 272.01 | +0.6% |
-| Model size (MB) | 1641 | 1641 | 0 |
-
-Within run-to-run noise (Phase 1 reproducibility check landed at
-+0.0% / -0.3% across two consecutive runs). The bump is performance-
-neutral on the SD1.5 baseline.
-
-### Rollback
-
-Reverting Phase 5 = restoring the Phase 1 pin set:
-
-```toml
-requires-python = ">=3.11,<3.12"
-dependencies = [
-    "python-coreml-stable-diffusion @ git+https://github.com/apple/ml-stable-diffusion.git@e5d960c41a6a4ab200b8db379194127607b1c590",
-    "torch==2.0.1",
-    "coremltools==8.2",
-    "numpy<1.25",
-    "overrides",
-    "diffusers>=0.22",
-    "peft>=0.6.2",
-    "omegaconf>=2.3",
-]
-# Drop the [tool.uv] override-dependencies block entirely.
-# Restore tests/m2/test_golden_image.py GOLDEN_PSNR_MIN_DB to 40.
-# Restore tests/m2/goldens/sd15_seed42.{png,sha256} from the
-# `modernize/phase4-tiered-ci` branch tip (the Phase 2 anchor).
-```
-
-Concretely: `git revert <Phase 5 commit>` followed by
-`rm -rf .venv && uv venv --python 3.11 && uv sync` will restore the
-Phase 1 environment. The Phase 1 `bench/results/ef2a18c.json` is the
-performance reference the rollback gets you back to.
-
-## Symbols we depend on from ml-stable-diffusion
-
-If we ever do need to vendor (option above), this is the surface to
-copy:
-
-| Import path | Used in |
-|---|---|
-| `python_coreml_stable_diffusion.unet.UNet2DConditionModel` | `coreml_suite/converter.py` |
-| `python_coreml_stable_diffusion.unet.UNet2DConditionModelXL` | `coreml_suite/converter.py` |
-| `python_coreml_stable_diffusion.unet.AttentionImplementations` | `coreml_suite/converter.py`, `coreml_suite/nodes.py` |
-| `python_coreml_stable_diffusion.unet.calculate_conv2d_output_shape` | `coreml_suite/converter.py` |
-| `python_coreml_stable_diffusion.coreml_model.CoreMLModel` | `coreml_suite/nodes.py` |
-
-Plus the LCM-specific conversion helpers in
-`coreml_suite/lcm/converter.py` (same `unet.*` symbols).
diff --git a/docs/spikes/01-multifunction-models.md b/docs/spikes/01-multifunction-models.md
deleted file mode 100644
index a1b4b07..0000000
--- a/docs/spikes/01-multifunction-models.md
+++ /dev/null
@@ -1,122 +0,0 @@
-# Spike 1 — MultiFunction models
-
-**Question (from spec):** Can one `.mlpackage` hold multiple batch sizes /
-controlnet-on-off variants sharing weights, replacing the
-"reconvert-per-config / encode-in-filename" approach?
-
-**Verdict:** Prototype for **controlnet-on/off only**. Use enumerated shapes
-(spike 2), not multifunction, for batch size. Hard-gated behind **macOS 15 /
-iOS 18**. Medium effort, medium payoff.
-
-## What it is
-
-MultiFunction packs several "functions" into one `mlprogram` `.mlpackage` and
-**deduplicates identical weights across them**, so disk and in-memory footprint
-shrink rather than just bundling N models.
-
-- Introduced: **coremltools 8.0** (2024-09-16). Unchanged through 9.0.
-- Runtime minimum: **macOS 15 / iOS 18** (enforced — `save_multifunction` forces
-  `spec_version = max(spec_version, iOS18)`).
-- `mlprogram` only (all of this suite's `.mlpackage`s already are).
-
-Build a multifunction package:
-
-```python
-from coremltools.utils import MultiFunctionDescriptor, save_multifunction
-
-desc = MultiFunctionDescriptor()
-desc.add_function(model_path="unet.mlpackage",
-                  source_function_name="main", target_function_name="plain")
-desc.add_function(model_path="unet_cn.mlpackage",
-                  source_function_name="main", target_function_name="controlnet")
-desc.default_function_name = "plain"          # mandatory
-save_multifunction(desc, "unet_multi.mlpackage")
-```
-
-Load a specific function (the save-time `CPU_ONLY` default does **not** bind
-runtime compute units — reload with the unit you want):
-
-```python
-import coremltools as ct
-m = ct.models.MLModel("unet_multi.mlpackage",
-                      function_name="controlnet",
-                      compute_units=ct.ComputeUnit.CPU_AND_NE)
-```
-
-Swift runtime selection: `MLModelConfiguration.functionName`.
-
-## Does it actually share weights?
-
-Yes, genuinely — not just a bundle. `save_multifunction` runs the MIL pass
-`const_deduplication` across functions. Dedup is **value-based (weight-hash)**,
-not reference-based: two *independently converted* packages dedup their
-bit-identical tensors at merge time.
-
-**Critical caveat for this suite:** dedup only fires if the shared tensors are
-**byte-identical**. Different palettization/quantization (this repo's
-`quantize_nbits` k-means path, `converter.py:309-325`) between two functions
-defeats the hash match — you silently get a bundle, not a shared-weight model,
-with no error. To benefit, convert the shared UNet trunk with identical
-precision for both functions.
-
-Apple's documented benefit is **smaller size + shared in-memory weights** (load
-base once, switch functions cheaply). No Apple source claims an ANE *compute*
-speedup beyond that memory/load win.
-
-## How the two targets map
-
-| Target | Fit | Why |
-|--------|-----|-----|
-| **controlnet-on / controlnet-off** | **Good** | Structurally the adapter case Apple pushes (WWDC24 10159/10161): shared UNet trunk + extra control inputs in one function. Different I/O signatures per function are fully supported. This is the prototype candidate. |
-| **different batch sizes** | **Awkward — wrong tool** | Batch is a *shape* difference, not a graph/weight difference. The idiomatic answer is `EnumeratedShapes` on a single function (spike 2): same weights, multiple shapes, no merge step, works on older OS targets. |
-
-Current per-config grounding: `converter.py:288-289` adds controlnet inputs via
-`add_cnet_support`; `naming.py:22-63` encodes `_cn` into the filename cache key.
-A 2-function package would replace the two separate `*_cn` / non-`cn` artifacts
-with one — but only for macOS 15+ users.
-
-## Pipeline support
-
-`python_coreml_stable_diffusion` / `apple/ml-stable-diffusion` has **no
-multifunction support** — it still emits separate `.mlpackage`s per component
-and per config (`--unet-batch-one`, separate `*_control-unet.mlpackage`). You
-would convert each variant as today, then add your **own** post-merge step
-calling `MultiFunctionDescriptor`/`save_multifunction`.
-
-## Effort / risk
-
-- **Hard OS gate (macOS 15 / iOS 18).** Single biggest adoption risk — many Mac
-  SD users lag on OS. Likely need to keep the per-config path as fallback, which
-  erodes the simplicity win. The suite currently targets macOS 13.
-- **Dedup defeated by mismatched quant** between functions (see above).
-- `default_function_name` is mandatory; `source_function_name` is almost always
-  `"main"` for `ct.convert` output — verify per package.
-- Spec version is forced up to the max of inputs (≥ iOS18); cannot keep a lower
-  target after merge.
-
-## Recommendation
-
-Low priority. If pursued: prototype **controlnet-on/off as a two-function
-package**, converting the shared trunk with identical precision so dedup fires;
-gate behind a macOS 15+ check; keep per-config reconvert as fallback. Do **not**
-use multifunction for batch size — that is spike 2's job. Measure ANE engagement
-and first-load compile time on real hardware before committing — those are the
-make-or-break unknowns.
-
-## Unverified (needs [M2-ANE])
-
-- Real-world ANE behavior of a merged SD UNet — no public report either way.
-- Whether Core ML compiles a multifunction package per-function or whole-package
-  on first load (SD UNet first-load compile is minutes — could worsen startup).
-- Exact `@available` version string for Swift `MLModelConfiguration.functionName`
-  (property confirmed; version inferred iOS 18 from the feature gate).
-
-## Sources
-
-- coremltools MultiFunction guide — https://apple.github.io/coremltools/docs-guides/source/multifunction-models.html
-- `coremltools/models/utils.py` source — https://github.com/apple/coremltools/blob/main/coremltools/models/utils.py
-- coremltools 8.0 release notes — https://github.com/apple/coremltools/releases/tag/8.0
-- WWDC24 10159 (Bring your ML/AI models to Apple silicon) — https://developer.apple.com/videos/play/wwdc2024/10159/
-- WWDC24 10161 (Deploy ML/AI models on-device with Core ML) — https://developer.apple.com/videos/play/wwdc2024/10161/
-- `MLModelConfiguration.functionName` — https://developer.apple.com/documentation/coreml/mlmodelconfiguration/functionname
-- apple/ml-stable-diffusion `torch2coreml.py` — https://github.com/apple/ml-stable-diffusion/blob/main/python_coreml_stable_diffusion/torch2coreml.py
diff --git a/docs/spikes/02-flexible-shapes.md b/docs/spikes/02-flexible-shapes.md
deleted file mode 100644
index 6f2b72f..0000000
--- a/docs/spikes/02-flexible-shapes.md
+++ /dev/null
@@ -1,135 +0,0 @@
-# Spike 2 — Flexible / enumerated shapes
-
-**Question (from spec):** Can enumerated/dynamic shapes +
-`materialize_dynamic_shape_mlmodel` remove the fixed 512×512 / per-size
-reconversion and simplify the `chunk_batch` padding?
-
-**Verdict:** Use `ct.EnumeratedShapes` over a small set of **spatial** latent
-sizes to kill per-size reconversion while keeping the ANE. Keep `chunk_batch`
-for the **batch** dimension. Medium effort, medium–high payoff. The batch logic
-is only partially eliminable and not worth the OS-gate cost.
-
-## The three mechanisms
-
-```python
-import coremltools as ct
-
-# (a) RangeDim — continuous bounded range per dim
-shape = ct.Shape(shape=(1, 4,
-        ct.RangeDim(lower_bound=64, upper_bound=128, default=64),
-        ct.RangeDim(lower_bound=64, upper_bound=128, default=64)))
-
-# (b) EnumeratedShapes — fixed finite set (≤128 shapes)
-shape = ct.EnumeratedShapes(
-        shapes=[[1,4,64,64],[1,4,96,96],[1,4,128,128]], default=[1,4,64,64])
-
-# (c) fixed — what converter.py:80 uses today
-ct.TensorType(name="sample", shape=(1,4,64,64))
-```
-
-## The ANE crux (confirmed, not inferred)
-
-coremltools maintainers confirmed in issue
-[#2370](https://github.com/apple/coremltools/issues/2370) (tested coremltools
-8.0 / macOS 15 / M3 Pro):
-
-- **`EnumeratedShapes` → runs every listed shape on the ANE.** This is the only
-  flexible option Apple endorses for an ANE model, and the only one their own
-  SD repo points to.
-- **`RangeDim` → only the *default* shape runs on ANE.** Non-default shapes fall
-  to GPU/CPU (measured ~10× slower). The `ReshapeFrequency.Infrequent`
-  optimization hint (iOS 17.4+) lets RangeDim use the ANE for arbitrary shapes,
-  but **each new shape triggers an ANE recompile** (latency spike).
-
-coremltools FAQ, verbatim: *"Use `EnumeratedShapes` for best performance ...
-The converted model will run on the NE, unless the conversion introduces dynamic
-layers not supported on the NE, such as converting a static reshape to a fully
-dynamic reshape."*
-
-## `materialize_dynamic_shape_mlmodel`
-
-Alternative path: convert once with symbolic dims, then bake out N fixed-shape
-**functions** into one weight-shared package.
-
-```python
-from coremltools.utils import materialize_dynamic_shape_mlmodel
-materialize_dynamic_shape_mlmodel(
-    dynamic_shape_mlmodel,
-    {"size_64": {"sample": (1,4,64,64)}, "size_96": {"sample": (1,4,96,96)}},
-    "unet_materialized.mlpackage")
-```
-
-- Introduced coremltools **8.0**. Tensor I/O only (SD UNet is `MLMultiArray` — OK).
-- Deployment target: materializing **>1 function** (or non-`main` source) emits a
-  multifunction model → forced to **iOS 18 / macOS 15**. Materializing exactly
-  **one** function from `"main"` stays unifunction at the **original** target
-  (macOS 13 stays viable for the single-shape case).
-- Materialized functions are fully static → run on ANE like normal fixed-shape
-  models, with shared weights across functions.
-
-## Does this kill `chunk_batch` or just per-size reconversion?
-
-Current grounding:
-- `converter.py:266-271` converts for a fixed `(batch, C, H, W)` `sample_shape`.
-- `latents.py:10-35` `chunk_batch` pads/splits the **batch** dim to the fixed
-  converted batch size.
-
-| Approach | ANE | Kills per-size reconv? | Kills `chunk_batch`? | Min target | Effort |
-|----------|-----|------------------------|----------------------|------------|--------|
-| `EnumeratedShapes` (spatial only) | Yes (all shapes) | **Yes** | No | macOS 13 | Med — `group_norm` hurdle |
-| `EnumeratedShapes` (spatial × batch) | Yes | Yes | Partially | **macOS 15** (multi-input) | Med–High |
-| `RangeDim` | Default only / `Infrequent`=recompile | Yes but slow | Partially | iOS 17.4 (hint) | Low code, bad ANE |
-| `materialize_dynamic_shape_mlmodel` | Yes (static funcs) | Yes (1 package) | No | macOS 15 if >1 func | Med |
-
-- **Per-size (spatial H/W) reconversion → realistically eliminable** via
-  `EnumeratedShapes` over a curated size set, keeping the ANE.
-- **`chunk_batch` batch padding → only partially eliminable, higher risk.**
-  Reasons: (1) batch must co-vary across multiple UNet inputs (`sample`,
-  `encoder_hidden_states`, `timestep`, SDXL/LCM/cnet extras), which hits the
-  **pre-iOS 18 single-`EnumeratedShapes`-input limit** — a hard wall unless you
-  target macOS 15+; (2) enumeration only covers the exact batch sizes you list,
-  so you still need `chunk_batch`-style fallback for unlisted counts;
-  (3) batch × spatial multiplies the enumerated-shape count, inflating compile
-  time and package size.
-
-## What breaks / risk
-
-- **`group_norm` conversion failure** is documented for flexible-shape SD UNet
-  (ml-stable-diffusion [#70](https://github.com/apple/ml-stable-diffusion/issues/70)):
-  flexible shapes failed at `group_norm`, and `SPLIT_EINSUM` with non-standard
-  dims caused kernel panics. Flexible *spatial* dims on the SD UNet are known to
-  be fiddly, not turnkey. (This suite defaults to `SPLIT_EINSUM` —
-  `converter.py:339`.)
-- One **unreproduced** report in #2370 of `EnumeratedShapes` on `mlprogram`
-  running only the default shape on ANE; the maintainer could not reproduce.
-  Treat the documented "all shapes on ANE" as expected but **benchmark on target
-  hardware** — this is the single biggest verification gap.
-
-## Recommendation
-
-Highest-value direction of the three packaging spikes. If pursued:
-`EnumeratedShapes` over a curated set of latent **spatial** sizes (e.g. 64²/96²/
-128²), keep batch fixed at the CFG-doubled value, keep `chunk_batch` for batch.
-Stays on macOS 13. Budget time for the `group_norm` conversion hurdle and
-**prove ANE residency on real hardware first** before removing any per-size
-reconversion. `materialize_dynamic_shape_mlmodel` is a clean fallback for
-"ship N fixed sizes in one weight-shared package" but forces macOS 15 once you
-bake >1 function.
-
-## Unverified (needs [M2-ANE])
-
-- `EnumeratedShapes`-on-`mlprogram` ANE residency for a *real* SD UNet
-  (the #2370 regression — unreproduced by maintainer).
-- That materialized functions run on ANE (strongly implied; no benchmark found).
-- `.mlmodelc` size / compile-time growth per enumerated shape (qualitative only).
-- Whether batch-as-flexible-dim is ANE-safe for the SD UNet (no API prohibition,
-  no positive confirmation).
-
-## Sources
-
-- coremltools flexible inputs guide — https://apple.github.io/coremltools/docs-guides/source/flexible-inputs.html
-- coremltools #2370 (Flexible shapes on Neural Engine) — https://github.com/apple/coremltools/issues/2370
-- coremltools #2386 (docs fix) — https://github.com/apple/coremltools/pull/2386
-- `materialize_dynamic_shape_mlmodel` source — https://github.com/apple/coremltools/blob/main/coremltools/models/utils.py
-- coremltools 8.0 release notes — https://github.com/apple/coremltools/releases/tag/8.0
-- ml-stable-diffusion #70 (flexible-shape group_norm failure) — https://github.com/apple/ml-stable-diffusion/issues/70
diff --git a/docs/spikes/03-mlx-interop.md b/docs/spikes/03-mlx-interop.md
deleted file mode 100644
index bda9e04..0000000
--- a/docs/spikes/03-mlx-interop.md
+++ /dev/null
@@ -1,106 +0,0 @@
-# Spike 3 — MLX interop decision
-
-**Question (from spec):** Should the suite stay pure Core ML/ANE, or add an
-interop path with DiffusionKit / ComfyUI-MLX for SD3/Flux (GPU/MLX)? Recommend,
-don't implement.
-
-**Verdict:** **Stay pure Core ML/ANE.** Document an off-ramp (point users to
-existing MLX nodes for Flux); do **not** absorb MLX into the suite. The niche
-MLX cannot reach — low-power, background, GPU-free — is exactly this suite's
-moat. The optional on-brand build is **SD3 Medium on ANE** (upstream-supported),
-plus a power benchmark to make the moat quantitative.
-
-## The two are complements, not substitutes
-
-MLX runs on **GPU + CPU only — it cannot target the ANE** (confirmed: MLX
-devices are `mx.cpu` / `mx.gpu`; the M5 "Neural Accelerators" are inside the
-GPU via Metal 4, *not* the ANE). So MLX structurally cannot serve this suite's
-niche.
-
-| Axis | Core ML / ANE (this suite) | MLX (mflux / DiffusionKit) |
-|------|----------------------------|----------------------------|
-| Compute target | **ANE** (or GPU) | GPU + CPU only |
-| Power draw | **Very low** (~5 W GEMM-level on M4) | High (~24 W; saturates GPU) |
-| Thermal / fanless / background | **Wins** — cool, doesn't block GPU | Pins the GPU |
-| Concurrent with GPU work | **Yes** (game/render/train running) | No — competes for GPU |
-| Raw speed | Competitive on SD1.5/SDXL | **Faster** for Flux/SDXL when GPU free |
-| Model coverage | SD1.5 / SDXL / **SD3** | **Flux**, FLUX.2, Qwen-Image, Z-Image; no SD3 |
-| Conversion friction | High (convert+chunk+palettize ahead) | Low (load HF weights, quant on the fly) |
-
-The power figures are **GEMM-level, not full-pipeline diffusion** — directional,
-not citable per-image. See "Unverified".
-
-## Ecosystem reality (2024–2026)
-
-- **DiffusionKit** (argmaxinc) — MIT. Python side runs SD3 + Flux on **MLX**;
-  Swift side does SD3 on Core ML (upstreamed to apple/ml-stable-diffusion).
-  **Repo archived (read-only) March 2026** — effectively end-of-life. Building
-  interop on it means depending on a dead repo.
-- **mflux** (filipstrand) — MIT, **active** (v0.17.x, Apr 2026, ~2k stars). The
-  healthy MLX-image engine: FLUX.1/FLUX.2, Qwen-Image, Z-Image, 4/8-bit quant.
-  No SD3. **GPU only.**
-- **Mflux-ComfyUI** (raysers) — MIT, wraps mflux; the most-starred Comfy MLX-Flux
-  node. The Flux-on-Mac-ComfyUI niche is **already occupied** here.
-- The original `thoddnn/ComfyUI-MLX` that blogs cite ("70% faster") **no longer
-  exists** (404, deliberately removed). Treat those claims as referencing a dead
-  repo.
-
-## Can SD3 / Flux run on ANE via Core ML?
-
-- **apple/ml-stable-diffusion** — MIT, 17.8k stars. Not abandoned but coasting:
-  **last tagged release 1.1.1 (May 2024)**, main branch commits through
-  **July 2025**. Maintained at a trickle.
-- **SD3 on Core ML/ANE: YES.** The README has a `--sd3-version` conversion path
-  (MMDiT, 16-ch VAE, T5), upstreamed from DiffusionKit. SD3 Medium ~2B params —
-  within ANE feasibility with palettization/chunking, same playbook as SDXL.
-- **Flux on Core ML/ANE: effectively NO.** Flux is ~12B params (~24 GB fp16
-  transformer + ~9 GB T5). ANE wants aggressive chunking (this suite already
-  bisects the 1.72 GB SD1.5 UNet into sub-1 GB chunks via `bisect_model`).
-  Flux would need heavy low-bit palettization **and** a dozen-plus chunks just to
-  fit, with precision risk on MMDiT attention. No maintained community Core ML
-  Flux exists. Flux on Mac in practice = MLX/GPU (mflux).
-
-## Maintenance burden if we went interop (rejected)
-
-- Two heavy new deps: **MLX** (large native, Apple-silicon-only, moves fast) +
-  **mflux** (DiffusionKit is archived, so mflux is the only live option). Both
-  violate the spec's "no new runtime dependencies without flagging" rule
-  (`MODERNIZATION_SPEC.md` §0.7) and would coexist with the already-fragile
-  `python_coreml_stable_diffusion` git dependency — doubling the moving surface.
-- The interop would be a thin re-wrap of `Mflux-ComfyUI`, which already exists
-  and is MIT. Zero differentiation for real maintenance cost.
-
-## Recommendation
-
-| Option | Effort | Payoff | Verdict |
-|--------|--------|--------|---------|
-| **A. Stay pure Core ML/ANE; optionally extend to SD3 Medium** | Low–Med | Med | **Do** |
-| **B. Document an off-ramp** in README (Flux/SD3-on-GPU → mflux / Mflux-ComfyUI; this suite = low-power / background / GPU-free / SD1.5/SDXL/SD3 on ANE) | Trivial | High | **Do** |
-| C. Build MLX interop inside the suite (mflux dep) | High | Low | **Reject** |
-| D. Build Flux-on-ANE Core ML conversion | Very high (research) | Low/uncertain | **Reject** |
-
-Net: stay pure, lean into the moat, point users elsewhere for Flux. The market
-already built the MLX-Flux bridge; this suite's value is the orthogonal axis MLX
-provably cannot reach. The one new build worth considering is **SD3 Medium on
-ANE** (feasible, upstream-supported, on-brand) plus a **power benchmark** to
-make the moat quantitative.
-
-## Unverified
-
-- A clean full-pipeline **diffusion** joules-per-image ANE-vs-MLX-GPU benchmark
-  (only GEMM-level ~5 W vs ~24 W on M4 exists). `bench/run.py` already measures
-  `CPU_AND_NE` vs `CPU_AND_GPU` — generating this would be the highest-value
-  artifact to prove the moat.
-- Any working community Core ML Flux conversion (none found).
-- Whether Argmax has a successor to DiffusionKit (archived Mar 2026, no announced
-  replacement; the live MLX-Flux torch is carried by mflux).
-
-## Sources
-
-- argmaxinc/DiffusionKit (archived) — https://github.com/argmaxinc/DiffusionKit
-- apple/ml-stable-diffusion (+ releases) — https://github.com/apple/ml-stable-diffusion / https://github.com/apple/ml-stable-diffusion/releases
-- filipstrand/mflux — https://github.com/filipstrand/mflux
-- raysers/Mflux-ComfyUI — https://github.com/raysers/Mflux-ComfyUI
-- MLX (unified memory / devices) — https://ml-explore.github.io/mlx/
-- Apple ML research, LLMs with MLX on M5 — https://machinelearning.apple.com/research/exploring-llms-mlx-m5
-- NPU vs GPU power (GEMM-level, directional) — arXiv 2511.13450
diff --git a/docs/spikes/README.md b/docs/spikes/README.md
deleted file mode 100644
index f464ce9..0000000
--- a/docs/spikes/README.md
+++ /dev/null
@@ -1,49 +0,0 @@
-# Phase 7 — Strategic spikes (research only)
-
-Research-only investigations per `MODERNIZATION_SPEC.md` Phase 7. None of these
-are merged behavior changes. Each spike delivers a written recommendation with
-effort/payoff so the maintainer can choose which (if any) become real phases.
-
-| # | Spike | Verdict | Effort | Payoff |
-|---|-------|---------|--------|--------|
-| 1 | [MultiFunction models](01-multifunction-models.md) | Prototype for **controlnet-on/off only**; use enumerated shapes for batch | Medium | Medium (gated by macOS 15) |
-| 2 | [Flexible / enumerated shapes](02-flexible-shapes.md) | `EnumeratedShapes` for **spatial size**; keep `chunk_batch` for batch | Medium | Medium–High |
-| 3 | [MLX interop decision](03-mlx-interop.md) | **Stay pure Core ML/ANE.** Document an off-ramp; do not absorb MLX | Trivial (docs) | High (positioning) |
-
-## Cross-cutting finding
-
-Spikes 1 and 2 both run into the same wall: the high-value packaging features
-(`MultiFunctionDescriptor`, `materialize_dynamic_shape_mlmodel`, multi-input
-`EnumeratedShapes`) require **macOS 15 / iOS 18** at runtime, while the suite
-today targets **macOS 13** (`coreml_suite/converter.py:110`). Adopting either
-means either raising the floor (drops older-OS users — a real cost for a
-community ComfyUI pack) or keeping the current per-config path as a fallback
-(erodes the "one artifact" simplicity win). This OS gate, not the API, is the
-dominant decision factor.
-
-## Gate 7 report
-
-```
-## Gate 7 report
-- Spike 1 (MultiFunction): recommendation written — prototype controlnet-on/off
-  as 2-function package; NOT for batch size. Gated macOS 15. No code merged.
-- Spike 2 (flexible shapes): recommendation written — EnumeratedShapes for
-  spatial size, retains ANE per coremltools maintainer confirmation; keep
-  chunk_batch for batch. No code merged.
-- Spike 3 (MLX interop): recommendation written — stay pure Core ML/ANE,
-  document off-ramp to mflux/Mflux-ComfyUI for Flux. Reject in-suite interop.
-  No code merged.
-- Open verification gaps requiring on-hardware [M2-ANE] measurement are listed
-  per spike under "Unverified".
-```
-
-## Verification debt (must measure on [M2-ANE] before any promotion)
-
-- EnumeratedShapes ANE residency for a *real* SD UNet (one unreproduced
-  coremltools regression report exists — spike 2).
-- Multifunction SD UNet ANE behavior + first-load compile time (no public
-  report either way — spike 1).
-- Full-pipeline diffusion power draw ANE vs MLX-GPU (only GEMM-level figures
-  exist — spike 3). `bench/run.py` already measures `CPU_AND_NE` vs
-  `CPU_AND_GPU`, so this is the highest-value artifact to make the moat
-  quantitative.
diff --git a/pyproject.toml b/pyproject.toml
index 8eb088b..79aeea4 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -6,11 +6,11 @@ license = "MIT"
 requires-python = ">=3.12,<3.13"
 packages = [{ include = "coreml_suite" }]
 dependencies = [
-    # Phase 5 toolchain bump: Python 3.12, coremltools 9, torch 2.7.
+    # Python 3.12, coremltools 9, torch 2.7.
     # numpy stays in the 1.24..1.x range — none of our modules need
     # numpy 2, and coremltools + ml-stable-diffusion's SD UNet trace
     # hit hard bugs under numpy 2 (`_cast` int(ndarray) strictness and
-    # `view` mixed-Var shape lists). See docs/deps.md.
+    # `view` mixed-Var shape lists).
     # torch 2.7 is the latest version coremltools 9's PyTorch frontend
     # has been tested against.
     "python-coreml-stable-diffusion @ git+https://github.com/apple/ml-stable-diffusion.git@e5d960c41a6a4ab200b8db379194127607b1c590",
@@ -31,8 +31,7 @@ Repository = "https://github.com/aszc-dev/ComfyUI-CoreMLSuite"
 PublisherId = "aszc-dev"
 DisplayName = "ComfyUI-CoreMLSuite"
 Icon = ""
-# Pinned to the ComfyUI commit this Phase 1 baseline was validated against.
-# Bump together with the toolchain upgrade in Phase 5.
+# Pinned to the ComfyUI commit this toolchain was validated against.
 requires-comfyui = "==ab5413351eee61f3d7f10c74e75286df0058bb18"
 
 [dependency-groups]
@@ -41,7 +40,7 @@ dev = [
     "psutil>=7.2.2",
 ]
 # ComfyUI runtime deps that aren't part of our package's runtime contract
-# but are needed to spin up the ComfyUI server for Tier 2 / smoke_image.
+# but are needed to spin up the ComfyUI server for Tier 2 integration tests.
 # Kept in a uv group so `uv sync --group comfy` brings them in without
 # polluting the published metadata (and without re-bumping our torch pin
 # via `uv pip install -r ComfyUI/requirements.txt`, which would float to
@@ -63,11 +62,11 @@ comfy = [
 ]
 
 [tool.uv]
-# Phase 5 toolchain bump: ml-stable-diffusion's setup.py hard-pins
-# numpy<1.24, diffusers==0.30.2 and transformers==4.44.2, which blocks
-# the modern torch / coremltools combo on Python 3.12. Override the
-# four blocking pins; the .unet / .coreml_model symbols we actually
-# import (see docs/deps.md) are stable across the bumped versions.
+# ml-stable-diffusion's setup.py hard-pins numpy<1.24, diffusers==0.30.2
+# and transformers==4.44.2, which blocks the modern torch / coremltools
+# combo on Python 3.12. Override the four blocking pins; the .unet /
+# .coreml_model symbols we actually import are stable across the bumped
+# versions.
 override-dependencies = [
     "numpy>=1.24,<2",
     "diffusers>=0.30",
@@ -76,14 +75,13 @@ override-dependencies = [
 ]
 
 [tool.pytest.ini_options]
-# Phase 2: tier markers gate which environment a test needs.
-# - unit: framework-free pure-logic tests (Tier 0; future Phase 3 will run
-#   these without ComfyUI on Linux). For now they may still transitively
-#   import comfy.
+# Tier markers gate which environment a test needs.
+# - unit: framework-free pure-logic tests (Tier 0; run without ComfyUI on
+#   Linux).
 # - m2:   needs an Apple Silicon Mac with the Neural Engine (Tier 2),
-#   typically the maintainer's self-hosted runner or local M-series box.
+#   typically a self-hosted runner or local M-series box.
 # - smoke: lightweight checks that need Apple Silicon + coremltools but no
-#   ANE/real model (reserved for Phase 4 Tier 1 smoke harness).
+#   ANE/real model (Tier 1).
 markers = [
     "unit: framework-free unit test (Tier 0)",
     "m2: requires Apple Silicon + Neural Engine (Tier 2)",
diff --git a/tests/conftest.py b/tests/conftest.py
index 1587338..5fd4a7c 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -1,14 +1,12 @@
 """Pytest bootstrap for ComfyUI-CoreMLSuite tests.
 
-- Adds the ComfyUI checkout to sys.path so production modules that
-  transitively import `comfy.*` resolve when pytest is invoked from this
-  package's root. Phase 3 will split pure logic into a comfy-free core and
-  this hack can go.
+- Adds the ComfyUI checkout to sys.path so the framework-coupled modules
+  that transitively import `comfy.*` resolve when pytest is invoked from
+  this package's root.
 - Auto-applies tier markers based on the directory a test lives in, so
   individual files don't have to repeat @pytest.mark.unit / .m2.
-- Skips the maintainer's in-progress test scaffolds so they don't break
-  collection (they reference modules / venv layouts that aren't part of
-  Phase 2 scope).
+- Skips in-progress test scaffolds so they don't break collection (they
+  reference modules / venv layouts that aren't part of this suite).
 """
 import sys
 from pathlib import Path
@@ -23,9 +21,8 @@ for p in (str(COMFY_DIR), str(REPO_ROOT)):
         sys.path.insert(0, p)
 
 
-# Skip WIP test scaffolds left in the tree by the maintainer; they import
-# modules (coreml_suite.experiments, convert_apple) that are not part of
-# Phase 2 scope.
+# Skip WIP test scaffolds; they import modules (coreml_suite.experiments,
+# convert_apple) that are not part of this suite.
 collect_ignore_glob = [
     "unit/test_experiments.py",
     "unit/test_unet_conversion.py",
diff --git a/tests/m2/test_golden_image.py b/tests/m2/test_golden_image.py
index 7bb4dcc..a43495e 100644
--- a/tests/m2/test_golden_image.py
+++ b/tests/m2/test_golden_image.py
@@ -1,4 +1,4 @@
-"""Phase 2 [M2-ANE] golden-image anchor.
+"""[M2-ANE] golden-image anchor.
 
 Runs the e2e SD1.5 + CoreML workflow against a local ComfyUI server, fetches
 the generated PNG, and asserts both:
@@ -14,8 +14,8 @@ regression.
 The 25 dB default reflects empirical drift between Core ML toolchains
 (8.2/np1.23/py3.11/torch2.0 → 9.0/np1.26/py3.12/torch2.7 measured at
 ~29 dB on the SD1.5 baseline; the image is visually identical at that
-level). Bump the threshold up for refactor PRs (Phase 3-style "should
-not change math"), down for toolchain PRs.
+level). Bump the threshold up for refactor PRs ("should not change
+math"), down for toolchain PRs.
 
 Skips entirely on non-Apple-Silicon hosts or when the server / converted
 model is missing, so the unit lane on Linux still passes.
@@ -111,8 +111,7 @@ def _post_workflow_and_collect_png() -> bytes:
     for nid in ("3", "11"):
         if nid in workflow and "seed" in workflow[nid].get("inputs", {}):
             workflow[nid]["inputs"]["seed"] = SEED
-    # Drop the MPS reference branch — torch 2.0.1 MPS path is broken on macOS 26
-    # (see Phase 1 gate report). Only the Core ML pipeline is needed here.
+    # Drop the MPS reference branch — only the Core ML pipeline is needed here.
     for nid in ("3", "8", "9"):
         workflow.pop(nid, None)
 
diff --git a/tests/unit/test_characterization_controlnet.py b/tests/unit/test_characterization_controlnet.py
index bb46258..dfdcecb 100644
--- a/tests/unit/test_characterization_controlnet.py
+++ b/tests/unit/test_characterization_controlnet.py
@@ -1,4 +1,4 @@
-"""Phase 2 characterization tests for coreml_suite.controlnet.
+"""Characterization tests for coreml_suite.controlnet.
 
 Locks shapes + dtypes + zero-fill behavior of expand_inputs / no_control /
 extract_residual_kwargs / chunk_control. These pure helpers feed the Core ML
diff --git a/tests/unit/test_characterization_inputs.py b/tests/unit/test_characterization_inputs.py
index 8209b1d..2f47b5f 100644
--- a/tests/unit/test_characterization_inputs.py
+++ b/tests/unit/test_characterization_inputs.py
@@ -1,10 +1,10 @@
-"""Phase 2 characterization tests for coreml_suite.models.CoreMLInputs.
+"""Characterization tests for coreml_suite.models.CoreMLInputs.
 
 Locks the shape transforms applied by chunks() and coreml_kwargs() for the
 four model variants the suite supports: SD1.5, LCM (SD1.5 + timestep_cond),
 SDXL base (time_ids len 6), and SDXL refiner (time_ids len 5).
 
-These contracts feed the Core ML UNet at runtime; if Phase 3 silently
+These contracts feed the Core ML UNet at runtime; if a refactor silently
 re-shapes them, generation breaks.
 """
 import numpy as np
diff --git a/tests/unit/test_characterization_latents.py b/tests/unit/test_characterization_latents.py
index 24e3df7..0ad40a0 100644
--- a/tests/unit/test_characterization_latents.py
+++ b/tests/unit/test_characterization_latents.py
@@ -1,7 +1,7 @@
-"""Phase 2 characterization tests for coreml_suite.latents.
+"""Characterization tests for coreml_suite.latents.
 
 Locks the *current* behavior of chunk_batch / merge_chunks — including the
-zero-pad regions and the truncation in merge — so the Phase 3 refactor
+zero-pad regions and the truncation in merge — so a refactor
 cannot silently shift either contract.
 """
 import pytest
@@ -110,7 +110,7 @@ def test_merge_chunks_truncates_padding():
 
 def test_merge_chunks_singleton_returns_equal_copy_when_shape_matches():
     """A singleton chunk list still goes through torch.cat, so we get a new
-    tensor equal to the input — locked here because Phase 3 might be tempted
+    tensor equal to the input — locked here because a refactor might be tempted
     to short-circuit and accidentally return the same object."""
     x = _const_tensor(2)
     out = merge_chunks([x], x.shape)
diff --git a/tests/unit/test_characterization_out_name.py b/tests/unit/test_characterization_out_name.py
index dce672b..a246a50 100644
--- a/tests/unit/test_characterization_out_name.py
+++ b/tests/unit/test_characterization_out_name.py
@@ -1,9 +1,9 @@
-"""Phase 2 characterization tests, Phase 3 re-pointed.
+"""Characterization tests for the .mlpackage filename composition.
 
-After Phase 3 the .mlpackage filename composition is the pure
+The filename composition is the pure
 coreml_suite.core.naming.compose_out_name function. CoreMLConverter.convert
-calls it; the previous Phase 2 test had to monkey-patch heavy converter
-internals just to capture the string, which made the test framework-coupled.
+calls it; testing the pure function avoids monkey-patching heavy converter
+internals just to capture the string.
 """
 import pytest
 
@@ -145,11 +145,11 @@ def test_lora_names_from_params_empty_list():
     assert lora_names_from_params([]) == []
 
 
-# ---------- Phase 6: quantize_nbits suffix ---------------------------------
+# ---------- quantize_nbits suffix ------------------------------------------
 
 
 def test_quantize_nbits_none_appends_nothing():
-    """'none' is the default and must keep the pre-Phase-6 filename so
+    """'none' is the default and must keep the unquantized filename so
     existing cached .mlpackages still resolve."""
     out = compose_out_name(
         ckpt_name="dreamshaper_8.safetensors",
diff --git a/tests/unit/test_characterization_sdxl_options.py b/tests/unit/test_characterization_sdxl_options.py
index c51c374..85cbd1b 100644
--- a/tests/unit/test_characterization_sdxl_options.py
+++ b/tests/unit/test_characterization_sdxl_options.py
@@ -1,6 +1,6 @@
-"""Phase 2 characterization tests, Phase 3 re-pointed.
+"""Characterization tests for the SDXL options math.
 
-After Phase 3 the SDXL time_ids / text_embeds math lives in
+The SDXL time_ids / text_embeds math lives in
 coreml_suite.core.sdxl as pure builders. The framework adapter
 add_sdxl_model_options (in models.py) is exercised separately by the m2
 golden image test; here we just lock the pure math.
@@ -95,7 +95,7 @@ def test_wrapper_captures_time_ids_text_embeds_refiner_via_closure():
 
 def test_wrapper_returns_zero_when_context_missing():
     """When c_crossattn is None the wrapper short-circuits to zeros_like(x).
-    Locked here because Phase 3 mustn't change this default."""
+    Locked here because the refactor mustn't change this default."""
     wrapper = sdxl_model_function_wrapper(torch.zeros(2, 6), torch.zeros(2, 1280))
     x = torch.randn(2, 4, 16, 16)
     out = wrapper(
diff --git a/tests/unit/test_chunks.py b/tests/unit/test_chunks.py
index 8ead7ed..f1f1dd7 100644
--- a/tests/unit/test_chunks.py
+++ b/tests/unit/test_chunks.py
@@ -1,8 +1,7 @@
-"""Original smoke tests, re-pointed at coreml_suite.core.* in Phase 3.
+"""Smoke tests for the pure batch-chunking helpers in coreml_suite.core.
 
-Drops the `from comfy.model_management import get_torch_device` import
-(replaced with torch.device('cpu') so Tier 0 runs without ComfyUI) and the
-dead `model_config` fixture (defined in Phase 1 but never consumed).
+Uses torch.device('cpu') instead of comfy.model_management.get_torch_device
+so Tier 0 runs without ComfyUI.
 """
 import pytest
 import torch
diff --git a/tests/unit/test_tier0_purity.py b/tests/unit/test_tier0_purity.py
index ffa0690..cff52f5 100644
--- a/tests/unit/test_tier0_purity.py
+++ b/tests/unit/test_tier0_purity.py
@@ -1,4 +1,4 @@
-"""Phase 3 gate: prove the Tier-0 lane is framework-free.
+"""Gate: prove the Tier-0 lane is framework-free.
 
 In a pure `pytest -m unit` run, none of the banned runtime modules
 (comfy, coremltools, python_coreml_stable_diffusion, folder_paths,