refactor(thumbnail): [PKG-TOOL-THUMB-04] preserve approved published defaults

refactor/changes/2026-09-15-PKG-TOOL-THUMB-04-approved-policy.md

@@ -0,0 +1,91 @@
+# PKG-TOOL-THUMB-04 approved Thumbnail policy and migration completion
+
+The user approved published 3.5.31 behavior as the default, with the unpublished
+workspace behavior explicitly selected through `compatibility: 'workspace-4.4'`.
+This completes the source/type migration task, including its earlier TypeScript,
+declaration, installed-entry and emitter checkpoints. It does not complete 05/06,
+physical-device evidence, formal reviews or publication readiness.
+
+## Implementation
+
+`src/policy.ts` selects and validates the policy. The public constructor restores
+`DEFAULTS.delay: 300` and published key order. Published mode validates/clamps
+height and delay to 10-1000, retains configured height and the selected input
+value, delays the video event and each frame, and waits twice the delay before
+done. Workspace mode keeps aspect-ratio height, synchronous video notification,
+no fixed extraction delays and input reset. The class DEFAULTS always represents
+the published policy; this boundary was explicitly approved.
+
+Source and extraction continue owning their resources separately. Source changes,
+native errors and destroy cancel pending notification, frame and completion
+timers. Frame readiness and seek completion are required in addition to the
+published delay. Cancellation settles the Promise without late update/done or
+new URLs. NaN retains the old asynchronous timer boundary rather than being
+mistaken for the workspace no-timer mode. Existing public spellings, descriptors,
+custom fields, event payloads, synchronous validation and return values remain.
+
+The authored declarations add optional compatibility/delay and exported
+Compatibility/DefaultOptions types. SheetOptions still accepts old literals
+without a delay field. Source and public constructors remain mutually assignable.
+The generated local Monaco declarations are regenerated; its real worker compiles
+the new mode/default types and executes constructor/event/cleanup code.
+
+README, architecture and type ownership guides describe both policies. No new
+dependency, lockfile or version change is needed. `test:thumbnail` and `test:unit`
+now include the new policy regression suite. Bundles and docs copies were built
+through the repository build script; source remains authoritative.
+
+## Verification
+
+Node 24.21.0, Yarn Classic 1.22.22, Windows. Evidence:
+[thumbnail-policy-validation.json](../baselines/thumbnail-policy-validation.json).
+
+- The initial new policy tests failed 6/7 against the previous implementation.
+  Final `yarn test:thumbnail`: 136 passed, no failure/skip/cancellation, 14.968 s.
+  This includes frozen historical tests and current/old TypeScript consumers.
+- Strict source tsc and scoped source/type/test/script lint pass. Five public
+  compiler/resolution modes reject all 12 invalid calls. Native CommonJS/no-interop
+  fixtures and both-direction source/declaration assignment pass.
+- Main/legacy/ESM build succeeds. Final external Yarn pack/install evidence is
+  `.cache/thumbnail-package-types-IH4Ouy`: offline frozen reinstall, installed-file
+  hashes, native require/import and 12 compiler cases pass. Five cases characterize
+  the frozen workspace missing-entry defect; seven exercise candidate consumers.
+  Installed legacy runtime/policy/lifecycle subset: 63/63, 2.770 s; the final
+  legacy bytes are identical to that tested snapshot. The final packed files are
+  checked against the working tree in the evidence recorder.
+- Three-engine source browser run: 90 passed, 0 failed/skipped/flaky, 66.423 s.
+  Installed main browser run: 30 passed, 0 failed/skipped/flaky, 33.414 s.
+  Chromium 153.0.8010.12 and Firefox 155.0 verify actual file selection, configured
+  versus aspect height, PNG pixels, repeated extraction, URL release, held native
+  encoding cancellation and latest-source-only updates. Published video/frame/done
+  delays have lower-bound assertions; owned-timer tests verify exact scheduling.
+- Windows WebKit 26.6 cannot decode the tested native Blob URLs (error 4), while
+  HTTP controls load. There are 23 annotated capability-control source cases and
+  10 installed cases. These are not successful extraction or physical Safari
+  evidence. Other passing historical tests also include exact old-defect controls.
+
+The first broad runtime attempt exposed fixtures that assumed no fixed delays;
+those now explicitly select workspace mode. A waiting worker was stopped and its
+unfinished run is not counted as passing. An initial strict error on optional
+delay was fixed through numeric normalization after runtime validation.
+
+The first browser run had 89 passes and one Chromium failure: selecting an
+identical file while published mode retains input.value did not trigger change,
+so starting a second job correctly rejected it. The replacement test now clears
+selection before choosing the file again. Both modes retain the original strict
+AbortError/latest-source assertions. Final reports and the initial trace are
+archived separately; no retry, weaker assertion or production input reset hides
+that failure.
+
+## Status and continuation
+
+PKG-TOOL-THUMB-04 is done; THUMB-COMPAT-01 is resolved for the approved default
+policy and migration boundary. No complete original npm archive or historical
+declaration has been recovered, and 05/06 retain integration/distribution work.
+Next implementation work includes the recorded installed Danmuku CPU-gap failure
+and remaining package/release preparation. Finish implementation and required
+tests, then await user direction for reviews; do not push or publish.
+
+Rollback requires reverting this task's source, declarations, generated outputs,
+tests and documentation together. Returning to the prior workspace default would
+also reverse the approved compatibility decision and must not be done silently.