* fix(inspect): show the prompts and settings the legacy parser finds
The workspace inspector took its prompts from the new server resolver,
which refused workflows where several samplers share one prompt (base
pass, upscale pass, FaceDetailer), and showed "Multiple possible
prompts" for files the legacy gallery and admin pages read correctly.
The inspector now runs the same extractGenerationParams those pages
use and shows its positive, negative, seed, steps, CFG, sampler and
model whenever it finds them. Prompts recorded at execution or saved
in the Library are never replaced; when the legacy parser finds
nothing, the server result stays.
Read-only check over a 38,602-file output folder: prompts shown went
from 1,798 to 7,542 files, and every one of the 7,236 positives and
9,421 negatives the legacy parser finds is shown identically.
* fix(inspect): keep definite server prompts over the legacy parse
The legacy parser follows ConditioningZeroOut back to the positive text
node and shows the positive as the negative. Apply its result only to
roles and settings the server could not decide, so a definite answer
such as an explicitly empty negative is never replaced.
Node callers (the corpus facts tool and the parity test support) now
pass the parser explicitly instead of relying on a browser global.
On a 38,602-file output folder, prompts shown stay at 7,542 files. Where
the server and the legacy parser both answer, they differ only in
surrounding whitespace (793 roles), except 227 zeroed negatives the
legacy parser misread and one prompt that is a JSON block.
* fix(inspect): never fill a setting the file does not record
The legacy parser can take a FaceDetailer as the generation sampler and
report its CFG for a workflow whose guider records none. Fill a setting
from the legacy parse only when the server left it unresolved or
ambiguous; a known value or a definite "not recorded" stays.
The corpus and parity Node tools again measure extractGenerationFacts
on its own, the contract their fixture vectors describe.
* fix(viewer): destroy an inline Viewer only after its pending view settles
Viewer.js's inline destroy() only unbinds and leaves an in-flight image
view running; when it finishes it reads the destroyed item list and
throws "c is undefined". Closing the preview while an image was still
loading hit this in Firefox about once in 15 runs of the recorded-run
history test. A closed viewer is now destroyed once that view settles,
its image fails, or a 10 second bound passes.
The history test passed 60 of 60 runs in Firefox with the fix.
* fix(viewer): keep the image src until a closed Viewer is released
Viewer.js rebuilds its item list from images that still have a src. On
close, the src was removed before the viewer cleanup ran, so a pending
view found no items and threw "c is undefined" (WebKit and Firefox).
With Viewer.js, the src is now removed after the viewer is destroyed,
once its pending view settles.
The recorded-run history test passed 40 of 40 runs in WebKit and in
Firefox.
Vito
d6eba05d21