# QA2 execution report

PodNotes E2E tooling uses Obsidian 1.14.4 by default on Linux and retains selectable 1.13.7. The plugin's shipped code and data contracts are unchanged.

## Revision and scope

Branch `test/e2e-hardening`. Original baseline `beda5be`; current base `3980111`. Final head `b431ed00b4a5537353d5454519d6421b59f9891e`. PR is recorded below after creation. Changes cover isolated launch tooling, fixture resolution, test rollback, journey readiness and developer verification recipes. No `src/`, dependency, package version, manifest, or published documentation content change.

Fix Root Causes shaped the local resolver, debounced core-config rollback and note-opening completion wait. Encode Lessons in Structure shaped the window guards, permanent rollback regressions and shared runtime leaf. The original 68-test runtime suite pinned plugin behavior. The final suite retains those behaviors and expands explorer/window isolation coverage.

## Obsidian versions

Official 1.14.4 archive came from `https://github.com/obsidianmd/obsidian-releases/releases/download/v1.14.4/obsidian-1.14.4.tar.gz`. Its SHA-256 is `e70b7cd63787e7f798a84fec453f5c446b91877bd3f612e8fb98e9b76cc81120`, matching the official GitHub asset digest in `official-release.json` and the downloaded archive in `download-sha256.txt`.

Installed without root under `~/.local/share/obsidian-versions/obsidian-1.14.4/`. `/opt/Obsidian` remains 1.13.7. The locked obsidian-e2e 0.13.0 runner already supports selecting the app, so no upstream version-selection PR was necessary.

```sh
npm run test:e2e
npm run stop:e2e-obsidian
PODNOTES_OBSIDIAN_APP=/opt/Obsidian/obsidian npm run test:e2e
```

`PODNOTES_OBSIDIAN_BIN` optionally selects a CLI outside the app's directory. The runner passes its resolved CLI into Vitest. Focused and full final runs deliberately set the unrelated `OBSIDIAN_BIN=/usr/bin/false` fallback to prove the selected CLI is used.

Commands ran on Linux under Xvfb with Node 22.23.2, using `fnm exec --using 22 --`. Scratch uses `/data/home/tmp/podnotes-qa2`. Linux profiles use `~/.cache/podnotes-e2e`; the measured CLI socket path is 100 bytes within the runner's 107-byte limit. macOS launch plumbing and isolated-suite selection were inspected but not executed on this Linux host.

## Root causes and proof

| Case | Before | After |
| --- | --- | --- |
| Public DNS | Journey and manual fixtures used `127.0.0.1.nip.io` and journey setup called public DNS lookup. | `podnotes.test` resolves to loopback through Chromium host resolver rules on every owned launch, including restarts. Real Settings feed loading and audio playback passed. An HTTP fixture test checks held responses and independent refreshes. |
| Closed explorer | On unchanged master, an enabled explorer with zero leaves failed with `File explorer is not open.` | The user command opens and reveals the explorer before a trusted right-click. Local-file cases pass with disabled, enabled/open and enabled/closed explorer. |
| Core-plugin leak | Baseline master suite left file-explorer enabled in `core-plugins.json`. | Per-test snapshots own runtime state and exact file bytes. Core debounced saves are drained before reload and again in the current renderer before restoring bytes. A regression appends JSON whitespace, toggles and persists the state, then checks real rollback and exact byte equality. A deliberate removed-restoration mutation fails this assertion. Appending whitespace remains distinct even after an interrupted seed. Full-run hashes are checked between runs. |
| Journey sleeps | Artless cards, loaded episode lists, delayed refreshes, playback and sync tests waited fixed intervals. | Tests wait on visible cards/lists, actual requests, explicit fixture response release, cache timestamps, settings or advancing media. Negative observations retain bounded measurement windows. |
| Extra window | On unchanged master, a hidden second BrowserWindow kept quit waiting until its 30-second timeout. | Restart rejects extra windows before closing the vault. Teardown closes and reports leaks, with window IDs/URLs in failures. Permanent tests check both behaviors. |
| Asynchronous note opening | The first batch's second 1.13.7 run failed the Markdown-rendering test. The captured note was correct, but the view was in Source mode with its preview hidden. | The test observes the original `openFile` promise before switching to preview. The temporary observer is restored in `finally`. Focused verification passed both versions; the final full series tests it repeatedly. |
| Introduced import cycle | An early candidate introduced `harness -> actions -> harness`, allowing undefined view constants in selectors. This was a development regression. | Runtime constants/readiness now live in `runtime.ts`, all callers migrate directly, and the harness imports helpers statically. The import scan reports zero static or dynamic cycles. |

Reproduction artifacts are `baseline.log`, `explorer-closed-before.log`, `windows-before.log`, and `pre-refactor/run-1.13.7-2.log`. The original Markdown Source-mode DOM was inspected by the parent, and the screenshot was inspected by the Claude reviewer, then overwritten by a later unsupported-module diagnostic because SDK capture paths depend only on test name. The latest capture was archived under `diagnostic-unsupported-module/`; it does not prove the original legacy failure. The original master baseline passed 68/68 on 1.13.7 with Vitest duration 231.59 s. That duration is a baseline characterization, not a cross-version speed comparison.

## Validation

Final full-run results and slowest tests are appended from `run-log.json`. Times are wall time for each complete npm command and build against a warm owned instance, not only Vitest. Per-version setup and shutdown are recorded separately and excluded from the per-run number. No automatic retries are configured or performed. Between each run the controller reads the actual Obsidian version from the window title, checks one window and disabled explorer, and compares the core-plugin file hash.

The six required commands were refreshed after the final timer change and passed on the exact committed content. `current-checks.json` records the final head and log hashes. The unit run passed 96 files and 1,587 tests; both Svelte checks reported zero errors and warnings. The docs build validated internal links and all 73 legacy fragments.

```sh
npm run lint
npm run format:check
npm run typecheck
npm run build
npm run test
npm run docs:build
```

An earlier runtime verifier passed seven focused cases on each version. Its earlier manual user flow added `http://podnotes.test:35398/feed.xml` in Settings, browsed Fixture Episode 1, played its 60-second WAV enclosure and paused. `data.json` recorded 21.42944 seconds and retained that position across plugin reload. Runtime errors were empty. Screenshots include Settings, episode list and player; the parent inspected Settings and player, and the verifier inspected every PNG. `verification/report.md` records exact commands and state.

## Development runs and deviations

The first stability batch is retained under `pre-refactor/`. Five consecutive 1.14.4 runs passed 71 tests in 267.098, 267.152, 287.136, 306.817 and 320.358 s. Legacy run 1 passed in 273.337 s; legacy run 2 failed in 308.055 s with 70 passed and one Markdown timeout. All had zero retries and preserved core config bytes. That batch does not count as final acceptance.

The initial import-cycle diagnostic candidates failed or were deliberately interrupted, then corrected. They are not represented as pre-existing product failures or accepted stability runs. The speculative focus-emulation override was removed because no visibility defect had been reproduced on this display.

Mechanical delegate `qa2-runtime-leaf-patch-1` on `claudeAgent/claude-opus-5-5` wrote a scratch patch and scratch copies despite a no-file-edit instruction, then removed the copies. It also reported removing two scratch directories before reuse without establishing their previous ownership. No foreign content or tracked-code change was evidenced. The parent reviewed and applied the patch as sole checkout writer.

QA2 changed no source, replaced no `/opt/Obsidian` installation, force-pushed no branch and merged no own PR. The parent stopped only owned instances. A reviewer attempted a broad `pkill -f` to end its slow regex search. That command failed; no other-unit process impact was evidenced, but the ownership constraint was not applied to the attempted stop. GitHub CLI was used because Origin was unavailable. The first batch's legacy failure required one fix round and a fresh full series.

## Reviews and dispatch ledger

- Contract grounding, documentation draft and comment review ran on `claudeAgent/claude-opus-5-5`.
- Alternative launch design ran on `codex_pool/gpt-6-astra`.
- Mechanical runtime-leaf draft ran on `claudeAgent/claude-opus-5-5`; the parent applied it.
- Initial full code reviews ran on `claudeAgent/claude-fable-5-1` and `codex_pool/gpt-6-astra`.
- Fix round 1 ran on the original Claude/GPT targets. In round 2, native Claude and its pool failed with rate limits; `codex_pool/claude-fable-5-1` restored that seat. `codex_pool/gpt-6-astra` and a separate `codex_pool/gpt-6.1-sol` fallback seat also reviewed it. Round 3 ran on `codex_pool/claude-fable-5-1` and `codex_pool/gpt-6-astra`. All accepted correctness findings were resolved in the third batch.
- Earlier real artifact verification ran on `claudeAgent/claude-opus-5-5`. The round-2 physical runs completed before its rate limit; a separate same-model `codex_pool/gpt-6.1-sol` seat audited the preserved logs and hashes. The final merged artifact verifier is recorded below after completion.
- Offline grounding ran on `codex_pool/claude-opus-5-5`, followed by the parent’s actual coordinate-race reproduction.
- The decision-trail attention audit ran on `codex_pool/claude-opus-5-5`.
- Implementation, selected-baseline reproductions and stability controller ran in-thread to preserve one writer and one runtime owner.

Decision trail `.audit/qa2.tsv` remains local and append-only. Final cross-family attention review and coordinator handoff are recorded below.


## Final additions and follow-ups

PERF1 supplied an offline J-EDGE-03 failure with both cached episodes listed but no playback or runtime errors. The captured state and source trace indicated that a failing uncached feed removed its loading banner after the harness measured click coordinates. In a real Obsidian reproduction, the test held the actual failing request, measured Keeper at (874, 388.859375), released the request and sent trusted CDP input to that old point. It hit the list background and left episode null, playing false. A fresh ordinary click then started the downloaded audio. `offline-layout-before.log` records the measurements. The final journey waits for global and selected loading indicators to disappear, and proves that Back navigation completed before selecting Keeper. It also checks the Node-side server probe on literal loopback, with HTTP 200 before stop and connection rejection after stop.

The original 172e51a full series produced three 73/73 greens, then run 4 was deliberately interrupted for the final review and coordinator flake additions. Those logs remain in `pre-round3/`; they do not count as the final series.

Final reviews used three fix rounds. GPT and Claude confirmed the final mechanisms. Native Claude and its pool hit rate limits in round 2. Both were retried as prescribed; the live catalog exposed the same Claude model through `codex_pool`, which retained model-family diversity. Additional GPT fallback review and a separate same-model runtime-artifact audit also ran. The artifact audit found a pre-existing disabled recorder directory and retained original interrupted-test fixture data. Both are inert state, not live instance leaks, and were left as found.

Master moved twice during QA. The branch merged #408 and then PERF1 #407 before the final seven-run series. QA2 itself changes no source, dependency, lockfile, manifest or build configuration relative to the current base. PERF1 changed the plugin bundle, so final manual evidence must identify the refreshed bundle separately from the retained earlier screenshots.

The approximate three-hour timebox was exceeded. Additional review findings, the coordinator's offline flake, rate-limit fallbacks, changing base commits and discarded partial stability series account for the added work. Final duration is recorded in the handoff.

Follow-ups outside this PR are the SDK's overwritten default failure captures, screenshot wrapper stdout containing a version line before JSON, and pre-existing cosmetic UI issues visible in the earlier screenshots (volume fill/label and Settings card alignment). No source UI change was folded into QA2.

Known interrupted-state leftovers are the initial disabled recorder podnotes-e2e-errors-2032567, journey-feeds-58340af5 fixture folder and Journey Latest feed, plus podnotes-e2e-errors-3104979 and podnotes-runtime-9307734f from the later deliberately interrupted run. Final acceptance uses a reused vault, not a fresh provision. The full controller checks core-file bytes between runs; its retries field is fixed at zero and reflects the inspected Vitest configuration and no automatic rerun path, rather than an independent retry counter. Explicit-app full runs verify selection; default-app selection with the environment unset is covered by focused and final manual runtime verification.


## Final core-save fix and stopping rule

The first 7cd369b series passed four full 1.14.4 runs (401.986, 332.883, 278.575 and 250.104 s), then its fifth run failed the exact-byte regression. The expected JSON plus a trailing newline was overwritten by canonical JSON. Those logs and the failure capture are preserved under `pre-core-drain/` and `core-bytes-before/`.

The rollback now drains and cancels the current renderer's core-config persistence after workspace restoration and before restoring file bytes. The earlier pre-cache drain operated in the old renderer. The regression cancels the startup timer, explicitly queues a fresh current-renderer save and observes exact bytes for 2.5 seconds after rollback. This is a bounded absence check for a delayed write, not a readiness sleep.

The coordinator explicitly authorized this additional final fix and its review beyond the skill's suggested three-round cap. The latest stopping rule requires exactly one more scheduled five-plus-two series. A fully green series gets the ready PR. If another flake appears, the PR still opens with the honest result, retained failure evidence and a separate GitHub issue per remaining flake, with no further fix round. The controller preserves each failed run's captures and attempts the remaining scheduled runs without retrying or changing code.

The core-drain review ran on `codex_pool/gpt-6-astra` and `codex_pool/claude-fable-5-1`. Claude traced the 500 ms debounce and adapter queue in both installed app bundles. Its accepted warning strengthened the regression by cancelling the startup timer before scheduling a fresh save. The removed-drain mutation then failed the exact byte assertion, while the corrected focused run passed. Fresh timer follow-up reviews ran on the same two model families.

## Attention audit responses

The cross-family trail audit by `codex_pool/claude-opus-5-5` was read in full. Its incomplete-series finding is superseded only by the seven final rows below. Earlier evidence remains developmental. The coordinator's explicit stopping rule supersedes the three-fix-round cap for the core-drain fix and one final series.

The Markdown timing concern compares a whole eleven-case test with the timeout of one 15-second wait. The test loops over eleven punctuation cases, each with note creation, open completion and preview rendering. Whole-test time alone cannot show an individual wait approaching its limit. Host contention likely affected those totals, but no causal speed claim is made. The final legacy rows give the direct result on the version that originally flaked. The natural Markdown failure and subsequent passing runs are weaker root-cause proof than the forced offline geometry reproduction; this limitation remains explicit.

Interrupted fixture and disabled recorder leftovers are listed above. No fresh-vault full-series result is claimed. Missing trail events and dispatch identifiers were added as retrospective append-only rows. Overwritten log and screenshot pointers have superseding corrections; unsupported-module captures are not original Source-mode proof. Current check logs and final revision metadata identify the latest tree. Same-model fallback seats are not family diversity. Claude through an alternate transport supplies the cross-family review.

Explicit-app full runs prove selection. The final manual user flow proves the unset-variable default. Zero retries reflects the inspected Vitest configuration and controller behavior, not an independent counter. Times cover command and build against a warm instance and should not be compared across series as performance results. A later master merge would invalidate the claim that this exact head received the seven-run series.

Final follow-up reviewers both cleared the timer delta. Claude noted that a removed-drain mutation can pass vacuously if the pre-write restore takes over 500 ms. The observed mutation failed at the expected missing-newline byte; no universal sensitivity claim is made. The correct drain cancels and flushes persistence regardless of that timing. The follow-up reviewer left one CPU-bound regex session alive; a cleanup-only reuse of the owning child stopped its exact zsh and grep PIDs and confirmed session exit143 before the series started. No broad process kill was used for that cleanup.

## Slow tests and safe speed changes

The seven-run table below records the five slowest tests in each raw log. The restart-and-restore journey waits for a natural disk checkpoint, restarts both main and renderer processes, observes paused playback and then resumes. The hide-window and popout persistence tests also wait for the normal checkpoint cadence to distinguish their event-triggered save from the next periodic save. Those real media, disk and process boundaries constrain their time. Seeding a saved checkpoint or skipping the actual restart would remove coverage. No speedup is claimed from unpaired, contended series.

Simple improvements already in this patch remove fixed journey delays. Refresh races use held HTTP responses and explicit release instead of waiting for arbitrary server delays. Cards, cache and episode lists use real readiness checks. Bounded windows remain where the assertion requires playback advancement or absence of a delayed write. Remaining slow cases merit profiling as a separate task rather than reducing coverage during the final bounded series.

QuickAdd prior art was rechecked against the shared research note and commits `65dcf2f3`, `e7724338` and `1cae501a`. Its vault-option tests avoid `app.json` writes by changing in-memory options, because delayed writes can undo restoration under load. QA2's core-plugin case deliberately exercises actual enable/disable and renderer reload, so replacing those transitions with memory-only flags would remove the defect the regression covers. The harness instead drains the owner of the delayed write before restoring exact bytes. No QuickAdd source was changed.

## Final seven-run result

All seven scheduled runs passed on `b431ed00b4a5537353d5454519d6421b59f9891e`. This is the final acceptance series after the core-save fix. No retry or code change occurred within it. Every run passed all 73 tests; actual versions, disabled explorer, zero explorer leaves, one window and exact core-config bytes all matched. Both version instances stopped, and both cleanup logs report `pids: none`. No remaining flake appeared in this final series, so no new flake issue was needed.

| Obsidian | Run | Complete command seconds | Raw result | Core bytes | Retries |
| --- | --- | ---: | --- | --- | ---: |
| 1.14.4 | 1 | 240.114 | 73 passed (73) | Preserved | 0 |
| 1.14.4 | 2 | 240.091 | 73 passed (73) | Preserved | 0 |
| 1.14.4 | 3 | 277.307 | 73 passed (73) | Preserved | 0 |
| 1.14.4 | 4 | 255.392 | 73 passed (73) | Preserved | 0 |
| 1.14.4 | 5 | 321.407 | 73 passed (73) | Preserved | 0 |
| 1.13.7 | 1 | 350.849 | 73 passed (73) | Preserved | 0 |
| 1.13.7 | 2 | 353.935 | 73 passed (73) | Preserved | 0 |

Total complete-command time was 2039.095 s, excluding version setup/shutdown and the state probes between runs. Raw results and log hashes are in `final-results-audit.json`. `run-log.json` has each controller row.

| Run | Slowest passing test | Milliseconds |
| --- | --- | ---: |
| 1.14.4 #1 | #377 J-PLAY-11: after restart the restored player keeps its saved position without autoplay | 25590 |
| 1.14.4 #1 | hiding the main window saves the playing position at once | 12621 |
| 1.14.4 #1 | leaves the playing episode, its position and the queue alone | 12271 |
| 1.14.4 #1 | keeps data.json intact through a reload right after a popout player closes (#370) | 11807 |
| 1.14.4 #1 | #355 J-PLAY-09 an episode that ends while the user browses plays the queued one | 8318 |
| 1.14.4 #2 | #377 J-PLAY-11: after restart the restored player keeps its saved position without autoplay | 24630 |
| 1.14.4 #2 | hiding the main window saves the playing position at once | 12846 |
| 1.14.4 #2 | leaves the playing episode, its position and the queue alone | 12286 |
| 1.14.4 #2 | keeps data.json intact through a reload right after a popout player closes (#370) | 11660 |
| 1.14.4 #2 | #355 J-PLAY-09 an episode that ends while the user browses plays the queued one | 8019 |
| 1.14.4 #3 | #377 J-PLAY-11: after restart the restored player keeps its saved position without autoplay | 24688 |
| 1.14.4 #3 | hiding the main window saves the playing position at once | 13025 |
| 1.14.4 #3 | leaves the playing episode, its position and the queue alone | 12285 |
| 1.14.4 #3 | keeps data.json intact through a reload right after a popout player closes (#370) | 11666 |
| 1.14.4 #3 | J-NOTE-07 the default note's Resume in PodNotes link resumes the saved position | 9636 |
| 1.14.4 #4 | #377 J-PLAY-11: after restart the restored player keeps its saved position without autoplay | 24646 |
| 1.14.4 #4 | hiding the main window saves the playing position at once | 12642 |
| 1.14.4 #4 | keeps data.json intact through a reload right after a popout player closes (#370) | 11912 |
| 1.14.4 #4 | leaves the playing episode, its position and the queue alone | 10877 |
| 1.14.4 #4 | J-PLAY-09 an episode that ends is marked played and the queue plays the next one | 7957 |
| 1.14.4 #5 | #377 J-PLAY-11: after restart the restored player keeps its saved position without autoplay | 26113 |
| 1.14.4 #5 | keeps data.json intact through a reload right after a popout player closes (#370) | 17651 |
| 1.14.4 #5 | hiding the main window saves the playing position at once | 14959 |
| 1.14.4 #5 | leaves the playing episode, its position and the queue alone | 12238 |
| 1.14.4 #5 | J-EDGE-03 offline after a restart: cached episodes browse, the downloaded episode plays, an unloaded podcast fails cleanly | 11180 |
| 1.13.7 #1 | #377 J-PLAY-11: after restart the restored player keeps its saved position without autoplay | 27996 |
| 1.13.7 #1 | keeps data.json intact through a reload right after a popout player closes (#370) | 15169 |
| 1.13.7 #1 | leaves the playing episode, its position and the queue alone | 14469 |
| 1.13.7 #1 | hiding the main window saves the playing position at once | 13567 |
| 1.13.7 #1 | J-PLAY-09 an episode that ends is marked played and the queue plays the next one | 10949 |
| 1.13.7 #2 | #377 J-PLAY-11: after restart the restored player keeps its saved position without autoplay | 32159 |
| 1.13.7 #2 | keeps data.json intact through a reload right after a popout player closes (#370) | 15394 |
| 1.13.7 #2 | hiding the main window saves the playing position at once | 14097 |
| 1.13.7 #2 | #355 audio keeps playing while the user browses the grid, an episode list and settings | 12293 |
| 1.13.7 #2 | leaves the playing episode, its position and the queue alone | 12183 |

Host contention varied. The initial sample showed load average 3.87 on 12 CPUs; a later recorded legacy-leg sample in `host-load.jsonl` showed 30.68. These are operational run times, not a controlled comparison between versions.

## Exact dispatch ledger

All task IDs below share the prefix `node:delegated-task:command%3Amcp%3A0108c973-9a6a-4a26-ab6a-b1a17d6e3f4c%3Adelegate-task%3A`. Append the task name to recover its exact ID. Raw terminal summaries and IDs are in `dispatch-ledger.json`.

| Delegated task | Provider and model | Terminal state |
| --- | --- | --- |
| qa2-architect-gpt-1 | codex_pool/gpt-6-astra | completed |
| qa2-how-contract-1 | claudeAgent/claude-opus-5-5 | completed |
| qa2-runtime-docs-1 | claudeAgent/claude-opus-5-5 | completed |
| qa2-no-comments-1 | claudeAgent/claude-opus-5-5 | completed |
| qa2-interrogate-gpt-1 | codex_pool/gpt-6-astra | completed |
| qa2-interrogate-claude-1 | claudeAgent/claude-fable-5-1 | completed |
| qa2-runtime-leaf-patch-1 | claudeAgent/claude-opus-5-5 | completed |
| qa2-fix-review-gpt-1 | codex_pool/gpt-6-astra | completed |
| qa2-runtime-verify-fix-1 | claudeAgent/claude-opus-5-5 | completed |
| qa2-fix-review-claude-1 | claudeAgent/claude-fable-5-1 | completed |
| qa2-fix-review-gpt-2 | codex_pool/gpt-6-astra | completed |
| qa2-fix-review-claude-2 | claudeAgent/claude-fable-5-1 | failed |
| qa2-runtime-verify-fix-2 | claudeAgent/claude-opus-5-5 | failed |
| qa2-fix-review-claude-2-pool | claudeAgent_claude_poolclaudeAgent_pool/claude-fable-5-1 | failed |
| qa2-runtime-verify-fix-2-pool | claudeAgent_claude_poolclaudeAgent_pool/claude-opus-5-5 | failed |
| qa2-fix-review-claude-2-gpt-fallback | codex_pool/gpt-6.1-sol | completed |
| qa2-runtime-verify-fix-2-gpt-fallback | codex_pool/gpt-6.1-sol | completed |
| qa2-fix-review-claude-2-codex-transport | codex_pool/claude-fable-5-1 | completed |
| qa2-offline-click-root-1 | codex_pool/claude-opus-5-5 | completed |
| qa2-fix-review-gpt-3 | codex_pool/gpt-6-astra | completed |
| qa2-fix-review-claude-3 | codex_pool/claude-fable-5-1 | completed |
| qa2-trail-attention-final | codex_pool/claude-opus-5-5 | completed |
| qa2-core-drain-review-gpt | codex_pool/gpt-6-astra | completed |
| qa2-core-drain-review-claude | codex_pool/claude-fable-5-1 | completed |
| qa2-core-timer-review-gpt | codex_pool/gpt-6-astra | completed |
| qa2-core-timer-review-claude | codex_pool/claude-fable-5-1 | completed |

The final current-bundle verifier is `qa2-final-runtime-b431ed0` on `codex_pool/claude-opus-5-5`, dispatched after both series instances stopped. Its terminal result is added below. The cleanup-only reuse is `node:delegated-task:command%3Amcp%3A0108c973-9a6a-4a26-ab6a-b1a17d6e3f4c%3Athread-send%3Aqa2-timer-grep-cleanup` on `codex_pool/claude-fable-5-1`. It stopped the search session it still owned and completed. Two unused runtime verifier tasks were interrupted before app use; they count as no runtime pass. Parent implementation and controller ran in-thread to retain one writer and one runtime owner.

| Unused delegated verifier | Provider and model | Terminal state |
| --- | --- | --- |
| qa2-runtime-verify-fix-3 | codex_pool/claude-opus-5-5 | interrupted |
| qa2-final-ui-after-series | codex_pool/claude-opus-5-5 | interrupted |

Final verifier limitation on the six original check logs: the logs predate the final commit by 441 to 510 seconds. The parent ran them after the last tracked edit and reports no code change before committing. The metadata records that final head and log hashes, but did not capture input hashes when the commands started. The verifier separately reran lint, format and typecheck at HEAD and rebuilt to the same bundle hash. The seven full suites ran after the commit.

## Final current-bundle verifier

`qa2-final-runtime-b431ed0` on `codex_pool/claude-opus-5-5` returned PASS at the final head. Its independent raw-log audit agreed with all seven results. The default app variables were unset. Actual app 1.14.4 ran on Linux x64, Electron 43.7.7 and Chrome 150.0.7871.250. Host tooling used Node 22.23.2; embedded renderer Node was 24.21.0. The bundle SHA256 before the verifier build, after it and inside the running app was `c5d3a6dff959abbd47ec66dc58adb952f1cc1db391f8bd1a177050752671e7fa`.

The verifier added `http://podnotes.test:35398/feed.xml` through Settings, used ribbon, palette and direct Show player paths, browsed the episode list, played the 60-second WAV and paused through the UI at 27.695497 seconds. A separate disk read recorded that exact position without forcing a save. Plugin reload retained the same data-file hash and paused API/player position. Runtime errors were empty. The parent and verifier opened all nine final PNGs, including the two post-reload captures.

After stopping its fixture server and app, the verifier restored the four files changed by the flow. The full manifest and recursive file comparison show zero differences from the original vault. Final dry-run reports `pids: none`; the parent also checked it after closing the terminal task. Other worktrees' Xvfb and fixture processes were identified and left untouched.

The manual flow did not press Play again after reload. The full suite covers actual restart and resume. Palette enable and plugin reload used direct CLI paths; feed and playback used trusted user input. Exact commands and evidence limits are in `final-verification/report.md`. The UI observations include 13 px horizontal overflow, a clipped queued title, a wrapped volume label and a status-bar sliver. They are outside this tests/tooling change and are inferred to be existing source behavior. No source fix was added.

The final verifier task is completed with no pending turns. No required work or live process remains in that lane.

Timebox record. Task start was 13:09:32 UTC. Final verification and publication preparation finished at 19:09:25 UTC, after 6.00 hours. The approximate 3-hour timebox was exceeded. The coordinator explicitly authorized the final core-save fix and one final seven-run series.