test(thumbnail): [PKG-TOOL-THUMB-02] reproduce extraction and resource lifecycle gaps

refactor/changes/2026-09-13-PKG-TOOL-THUMB-02-behavior.md

@@ -0,0 +1,79 @@
+# PKG-TOOL-THUMB-02 extraction and lifecycle baseline
+
+Started from 0338e0b8b1d9814ebe1628c7b55d08003ffa6347. No production runtime or
+distribution changes. This task reproduces behavior before the resource/module
+work in 03 and TS migration in 04. The old default-semantics conflict remains open.
+
+## Repeatable controlled tests
+
+`yarn test:thumbnail` runs the 26 frozen contract assertions and 36 extraction
+assertions against recovered CDN 3.5.31 and frozen workspace main/legacy.
+The new behavior file also runs in the normal unit/CI command.
+
+Controlled timers and deferred Blob callbacks allow assertions without arbitrary
+waiting: serial frames, interval clamps/density guards, progress/done ordering,
+repeat generation, duplicate start/download, pending callback gating, null Blob,
+late callback after destroy, replacement URLs, missing metadata, changed input,
+seek setter failure, throwing done/update listeners. These are baseline tests,
+not assertions that the defects are fixed. Future candidate regression tests
+must assert corrected resource ownership separately while preserving old evidence.
+
+The old and current implementations leak the first source URL on replacement,
+retain changed-input listeners and generated wrapper inputs, permit late update
+after destroy, and leave several callback-failure promises unsettled. Missing
+metadata rejects with an unbound receiver in 3.5.31; workspace keeps polling even
+after destruction. Seek-assignment failure rejects correctly in both; a throwing
+done listener emits error twice in 3.5.31 and once in workspace. Normal progress
+and PNG URL replacement remain ordered. These distinctions guide 03 fixes.
+
+## Actual browser evidence and limitations
+
+Run `yarn test:browser test/browser/thumbnail-tool.spec.js test/browser/thumbnail-native.spec.js`.
+The final 39 rows have zero retries/skips: **24 actual tool scenarios** in
+Chromium/Firefox, **12 Windows WebKit Blob-unavailable controls**, and **3 native
+media comparison tests**. The 12 controls do not establish successful extraction.
+Every control proves that the same file loads in native HTTP video but fails in
+native Blob video with media error 4, then reproduces the tool's corresponding
+failure and cleanup. If native Blob starts working, the full scenario runs.
+
+Actual scenarios use native file selection, metadata/seek, real canvas.toBlob,
+PNG decode and non-black opaque pixels. They verify ten updates, two complete
+generations and normal final URL cleanup. DOM drop dispatch with a real File
+reproduces the missing listener; calling the existing ondrop method does load.
+Corrupt media raises a native error without a tool error event; replacement leaks
+the first URL and a second destroy throws NotFoundError. A captured **native PNG
+callback**, delivered after destroy, still emits update and creates an unreleased
+URL. Only callback delivery is controlled; encoding and media are native.
+
+Initial docs sample testing passed 24 Chromium/Firefox cases and failed all 12
+WebKit cases. A second 24-second video-only MP4 probe still failed WebKit. The
+independent native control then demonstrated HTTP success and Blob failure for
+the original pattern, looped pattern and docs sample. An exploratory WebM HTTP
+control did not settle within its 1.5-second observation and is not used to claim
+codec support. All failed reports and trace directories remain archived. No
+production media workaround or claimed physical Safari result was introduced.
+
+`test/browser/media/thumbnail-pattern.mp4` is a 24-second stream-copy loop of the
+existing generated 8-second pattern, needed for the tool's minimum ten samples
+and maximum one-sample-per-second density. Reproduce with FFmpeg:
+
+```sh
+ffmpeg -stream_loop 2 -i test/browser/media/pattern.mp4 -t 24 -an -c:v copy -movflags +faststart test/browser/media/thumbnail-pattern.mp4
+```
+
+The local test server exposes this fixture; no dist or compiled docs files were
+edited. Hashes, individual outcomes, measurements, browser versions and CI counts
+are in `refactor/baselines/thumbnail-behavior-validation.json`. The ordinary
+browser job discovers the new spec files automatically. No hosted CI run,
+coverage improvement, successful WebKit Blob extraction or npm publication is claimed.
+
+## Handoff and rollback
+
+03 should own source URLs, generated inputs/listeners, metadata waiting and
+extraction callbacks; settle pending work on cancellation/error and stop late
+effects. Keep historical public methods/events and use a documented compatibility
+policy for delay/height defaults. 04 covers emitter provenance and strict TS;
+05 keeps actual Safari/WebKit Blob extraction and old/new core sheet integration
+open; 06 owns complete installed-package/demo checks. Reverting this task's
+dedicated commit removes only tests, fixture/server registration, scripts and
+refactor evidence; it does not roll back production code.