# PLAY2b verification

PASS. The fix, local gates, mutation proof, live handoff, write-count comparison and both final full suites pass.

Branch `fix/ended-relisten-handoff`. Head `c52b56974ca2d4f744eed2f7463147087f7cbc2a`. Base `f295af14901c397a69b8b675fc40b26709ddcc05`. The old merged branch remains at `35d08de1e4f6b5e33a39195b586fa32d284d831a`.

## Fix and reproduction

The two loaded-session Play paths hardcoded zero after a natural end. They now use the existing `resumeTime()` policy, which returns unfinished saved progress or zero for finished progress. They respect a rejected seek when the synced time exceeds the local media duration. The production diff adds three lines and removes four. No state, schema, settings, API, URI or docs content change.

Master was reproduced before the fix in headless Linux Obsidian 1.14.4, using a real 60-second `Fixtures/Local Fixture.wav`. The URI was `obsidian://podnotes?episodeName=PLAY2b%20Handoff&url=Fixtures%2FLocal%20Fixture.wav&time=58`. Playback ended naturally. The test rewrote `data.json` to unfinished progress at 30, applied the external change, flushed, then used the visible Play and pause button.

| Observation | Master | Fixed bundle |
| --- | ---: | ---: |
| Saved time after natural end | 60, finished | 60, finished |
| Saved time after sync and flush | 30, unfinished | 30, unfinished |
| Time after visible Play | 0.003961 | 30 |
| Saved time after visible pause | 0.085567 | 30.077229 |
| Finished without sync, direct Start playing | 0.155975 | 0.15429 |

The fix resumes the re-listen and preserves the intentional restart for finished progress. `master-live.json`, `final-live.json`, both command records, and the before/after screenshots contain the evidence. Debugger-attached live checks captured no renderer errors.

The duration-mismatch probe used a local 60-second file, a synced unfinished record at 90 of 120 seconds, and another queued file. One visible Play left playback paused at 59.75 through nine observations over three seconds. A flush retained 90/120 unfinished, the current episode stayed selected, and the next file stayed queued. See `boundary-live.json` and `boundary-commands.json`.

## Regression and mutation checks

The failing-test commit is `f6c312af`, followed by production fix `7501852a` and test timing refinement `c52b5697`.

- The final player regressions on master produced 4 failures and 140 passes. Both Play entry points restarted at zero instead of 30. Both cases with a shorter local copy also restarted incorrectly.
- A bare replacement of the two zero destinations fixed the main case but left 2 failures. Playback started at 119.75 even though the seek rejected the saved time of 125 on a 120-second local copy. Respecting `moveTo()` fixed both cases. The player suite passes all 144 tests.
- Mutations that ignored either rejected seek, or put synced progress ahead of an explicit time, each failed the corresponding unit test.
- The new ended journey failed on master with `expected 0 to be greater than or equal to 30`. Its finished-without-sync control passed. The final journey also pauses visibly and verifies the exact persisted time.
- The playing-sync journey flushes local progress, writes and applies the external change, and reads the merged value in one renderer call. It checks lower and upper playback bounds afterwards. The exact V-462 mutation failed against the final test with `expected 51.473782 to be less than 13.114272`. Source restoration was byte-for-byte, recorded in `final-mutation-restoration.json`. The final unmodified journey passed in the 1.14.4 full run.

Time bounds prove rejection of that mutation. They do not detect every possible small seek. The combined renderer call removes the separate post-merge read race; filesystem work remains asynchronous.

## Required checks

All local checks passed on final head under Node 22.23.2 and `TMPDIR=/data/home/tmp/podnotes-PLAY2b`.

- `npm run lint`, `npm run format:check`, `npm run typecheck`, `npm run build`.
- `npm run test`: 2,432 tests in 122 files. Both Svelte checks found zero errors and zero warnings.
- `npm run test:tooling`: 13 tests in 2 files.
- `npm run docs:build`.
- `npm run size`: 311,774 raw bytes and 98,336 gzip bytes at level 9. Master was 311,744 and 98,324. The delta is +30 raw bytes and +12 gzip bytes, with 1,664 gzip bytes remaining.

Both complete suites passed on final head, each from a fresh vault, with no filters and no unexpected renderer errors:

- `fnm exec --using 22 -- npm run test:e2e` on Obsidian 1.14.4: 141 tests in 19 files, 520.53 seconds.
- `PODNOTES_OBSIDIAN_APP=/opt/Obsidian/obsidian fnm exec --using 22 -- npm run test:e2e` on Obsidian 1.13.7: 141 tests in 19 files, 522.81 seconds.

Logs are `final-head-e2e-1.14.4.log` and `final-head-e2e-1.13.7.log`. Both record the final commit, clean tree, fresh vault and unchanged bundle hash before and after.

The earlier fresh-vault full 1.14.4 run at production commit `7501852a` passed all 141 tests in 19 files. The later commit changes only test scheduling. Live proof and the final builds use identical bundle SHA-256 `3140be46970c206f545b808f3de6b798c3c4faa2dcbfeb14d86e83048c552e27`.

`npm run count:writes -- start`, `report` and `stop`, driven by `count-session.py`, match master in every recorded field. The session records 0 writes after two idle seconds, 2 after sixteen seconds of playback and pause, and 3 after a visible seek and flush. No failed or in-flight writes; recorder intact, complete, and stopped before the app. See `master-writes.json`, `final-writes.json` and `parent-verification.json`.

## Independent review

One interrogate pass used `claudeAgent/claude-fable-5-1` and `codex_pool/gpt-6-astra`. Neither found an introduced production correctness issue. The Claude timing warning led to the smaller combined renderer call in the test. The GPT small-seek coverage limitation is recorded above. A helper-extraction nit was dismissed because the two short intent paths already share the canonical saved-progress and seek policies. The comment review found no added comments or suppressions. Parent inspection confirmed the final diff also has none.

Fix Root Causes required the real failed Play and save before editing. Foundational Thinking kept the existing clean/dirty ownership model. Prove It Works required persisted readbacks, actual assertion failures and the same live interaction after the fix.

## Docs at release

In #458's sync handoff and mobile guidance, add that an episode ended here resumes unfinished synced progress when the user presses Play. A saved finished episode still restarts at zero. The ended player may show its old end position until Play. Retain #462's immediate adoption for a clean paused player, the segment-preview exception, waiting for sync, and selecting the desired episode when a different episode is loaded.

## Limits and deviations

Evidence covers Linux desktop Obsidian and real Chromium media events. No physical mobile claim. The source review confirmed that an unsaved native seek after a natural end can still be ignored by Play before its checkpoint; that case predates this fix. The local-file placeholder's overlaid play/image glyph also exists on master.

The first runtime verifier accidentally invoked `/opt/Obsidian/obsidian --version` outside the runner. It briefly started the default profile, appended one log line and left three stale lock links. The command found no configured vault file, the exact process was confirmed dead, and the verifier removed those links. No vault opened. It also wrote a feedback memory and index entry outside the worktree. Those actions exceeded the worktree-only instruction. The parent inspected the actual tool timeline and recorded the paths and results in `verification-deviation.md`. All test evidence came from the isolated runner. Subsequent verification reads versions only through that runner.

Final cleanup found no worktree Obsidian, Xvfb, fixture server or recorder processes. The runner dry-run reported `pids: none` and the instance profile was removed. The working tree is clean. Generated vaults remain under ignored evidence directories and are excluded from the public archive. No `.obsidian-e2e-vaults` directory remains. See `final-head-cleanup.log`. All delegated tasks completed and were explicitly stopped.

## Dispatch ledger

- Investigation and regressions: `claudeAgent/claude-opus-5-5`.
- Design candidates: `claudeAgent/claude-fable-5-1` and `codex_pool/gpt-6-astra`.
- Design judge: `claudeAgent/claude-opus-5-5`.
- Implementation: `claudeAgent/claude-opus-5-5`.
- Comment review: `claudeAgent/claude-opus-5-5`.
- Interrogate: `claudeAgent/claude-fable-5-1` and `codex_pool/gpt-6-astra`.
- Test timing refinement: `claudeAgent/claude-opus-5-5`.
- Live runtime verification: `claudeAgent/claude-opus-5-5`.
- Final fresh-vault suites: a fresh `claudeAgent/claude-opus-5-5` agent.
- Parent: reproduction, mutation checks, local gates, diff and evidence review, integration and PR report.

No model-family fallback. Exact task IDs are in `dispatches.tsv`.