Findings for [#33283](https://git.cafe/pingdotgg/t3code/pull/33283), proof-audit-4, final W/B recording recheck, 2026-11-10.

1. **B — one timestamp label still describes the wrong event.** The current paragraph says “thread fails while inactive 0.2” before and “fails 0.3” after. At those frames the IDLE control is active, the folder is closed, and no failure indicator is present. Subsequent frames show the blue working dot before the failure indicator: before is still blue at 1.3 s and red at 1.4 s; after is still blue at 2.1 s and amber at 2.2 s. The supplied `capture-b.py` marks those early timestamps immediately before calling the driver's `send B`; `driver.mjs` dispatches a message, rather than injecting an already-completed failure. The linked W/B index correctly calls 0.21/0.26 s **dispatch** times. Change the current paragraph to “dispatch B while inactive 0.2/0.3; failure visible by 1.4/2.2.” No new recording is needed.

2. **W — the newly disclosed Resume transients are accurate, but one earlier loading transient remains undisclosed.** In [W after](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/9e8025f257545228/W-flow-after.webm), the exact 0.4 s frame has both file disclosures open, but beta reads **“Loading output…”**. Both file contents are populated at 0.5 s and remain readable in the hold. Thus “expanded 0.4” is accurate as a disclosure-state claim; it is not a fully loaded checkpoint, and the paragraph's transient disclosure currently mentions only the Resume spinner and prior-error notice. “Transient startup splashes” does not describe a file-output placeholder appearing on expansion. Add “beta briefly shows Loading output… at 0.4; both contents loaded by 0.5,” or narrate the populated expanded checkpoint at 0.5 and disclose the preceding load. This is normal first-expansion loading, not corrupt or stuck video. No recapture is needed.

3. **The previous rejection's substantive recording concerns are resolved by the current wording.** W's Resume spinner at 9.2 s is a pending action; the existing raw prior-error notice is visible at 9.5–9.7 s; the notice clears and continuation succeeds by 10.2 s. B's held notification positions are correctly described as about 1.9 s before and 2.9 s after; dismissal is in progress at 3.8/5.6 s and clean closed-folder holds are visible by 4.1/5.9 s. The approximately 0.5 s toast entry/exit motion is disclosed. These states do not justify a recapture. In particular, removing W's red notice by dismissing or hiding it before the resumed turn would change the demonstrated interaction; repeating the same interaction offers no basis to promise it disappears. It is stale in the sense that its text belongs to the previous failure, but it is actual current app rendering, not a stale recording frame. Existing [ChatView.tsx](https://git.cafe/pingdotgg/t3code/blob/68414fb0aa3321c60eee6f714c4c5b586ed82b5e/apps/web/src/components/ChatView.tsx#L981) hides the raw thread error while presenting the failed usage-limit recovery run, and the unchanged [ThreadErrorBanner.tsx](https://git.cafe/pingdotgg/t3code/blob/68414fb0aa3321c60eee6f714c4c5b586ed82b5e/apps/web/src/components/chat/ThreadErrorBanner.tsx#L72) renders remaining raw errors. This supports the lead's explanation; the successful continuation independently rules out a stuck Resume in this clip.

4. **The linked W/B evidence index still carries superseded narration.** The [W/B index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/4ea1a06e95d4db4c/evidence-index.md) still labels B's 1.53/2.68 s notification checkpoints “settled,” despite visible entry motion, and omits the W Resume transients. The current PR paragraph corrects those points, but the section also directs readers to that index for narrated timestamps. Mark that index's old narration superseded by the PR paragraph, or publish a corrected index. This is another text correction, not a capture defect. No other W/B capture contradiction was found in the section's classification, recovery, manual collapse, sidebar, folder or hover claims. Native and other accepted rows were not reopened.

| Row | Status | Remaining work |
|---|---|---|
| W | INSUFFICIENT — disclosure only | Disclose the 0.4 s file-output loading placeholder, with both contents loaded by 0.5 s. Expansion, collapse, Resume, reply, previously accepted stills/twins/masks all stand. |
| B | INSUFFICIENT — narration only | Relabel 0.2/0.3 s as dispatch, not completed failure; mark the linked index's old settled timestamps superseded. Corrected holds, dismissals, folder/hover states, previously accepted stills/twins/masks all stand. |

No recapture is required for either row. The other 14 rows retain their accepted status and existing limitations. The verdict below concerns the requested exact narration and complete transient disclosure, not a newly found adapter regression or a demand that every video frame be settled.

Verification details:

- Read the PR through `cafe pr view 33283 --repo pingdotgg/t3code --json`, extracting `data.resource.description`. Re-read at completion: description unchanged, resource version 10, head unchanged at [68414fb0aa](https://git.cafe/pingdotgg/t3code/commit/68414fb0aa3321c60eee6f714c4c5b586ed82b5e). The read-only checkout is clean at that head on the requested branch. Its source diff from the stated before commit touches only the Pi adapter and its tests.
- Downloaded all four clips with curl and decoded every frame with ffmpeg. All are 1440×1000 at 10 fps: W before 11.5 s, W after 14.9 s, B before 10.7 s, B after 13.9 s. Inspected exact frame-index samples, timestamped full-surface contact sheets, and unscaled detail samples around entry, expansion and continuation. No decoder errors were reported.
- W before: both disclosures open at 0.4, both collapsed at 5.2, no Resume at 9.4. W after: open at 0.4 with the loading caveat above, populated at 0.5, collapsed at 4.9, pending Resume at 9.2, prior-error notice 9.5–9.7, reply and “Read 2 files” by 10.2. The ordinary working/thinking presentation during continuation clears; its “Working for 1h” age follows the linked index's documented server/browser clock offset, not real elapsed recording time.
- B before: dispatch checkpoint 0.2, failure indicator visible by 1.4, notification holds about 1.9, dismisses 3.8, closed folder clean by 4.1, folder open at 6.4, hover card held at 8.8. B after: dispatch checkpoint 0.3, limit indicator visible by 2.2, notification holds about 2.9, dismisses 5.6, folder clean by 5.9, folder open at 9.3, hover card held at 12.0. Hover timestamps describe the interaction, not an already-open card.
- Still arithmetic, twins and masks were accepted in the previous audit and deliberately not recomputed. No repository edits, pushes, PR comments, browser sessions or dev servers were created.

After: exact W frame at 0.4 s, with beta expanded and “Loading output…” visible. The 0.5 s frame already contains its text.

![After: W at 0.4 seconds showing the brief beta output-loading placeholder](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/1cd8592bab00dae0/W-after-0.4s.png)

Before: B dispatch checkpoint precedes the blue working indicator and subsequent red failure indicator. These are labeled explanatory crops, not altered comparison evidence.

![Before: B at dispatch, working, and failure checkpoints](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0cd3af4d625bc879/B-before-dispatch-details.png)

After: B dispatch checkpoint precedes the blue working indicator and subsequent amber limit indicator.

![After: B at dispatch, working, and usage-limit checkpoints](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/85c5ef2f9c9dbb76/B-after-dispatch-details.png)

FRICTION:
- tool | Python Pillow was unavailable, so the initial contact-sheet command failed and its first image could not be opened | Use the already-installed ffmpeg crop, timestamp and tile filters; they produced all inspection sheets without installing anything.

VERDICT: INSUFFICIENT W B