Proof audit 2 for [#33283](https://git.cafe/pingdotgg/t3code/pull/33283), 2026-11-10. Read-only review; no source edits, pushes, or forge comments.

Findings first:

1. **B — INSUFFICIENT as the requested before/after proof.** B-hover-before.png/B-hover-after.png and both masks show substantial unrelated fixture changes: sidebar count 19→18, reordered rows, completed M/W/R/PNT threads, draft marker, active S moving into Snoozed, snoozed count, and the main-pane wake notice/banner. B-folder-before.png includes a toast that B-folder-after.png does not. This is more than an age-label difference: the hover target moves and the entire sidebar/main-pane comparison changes. Individual Limited/Usage limit reached labels and the notification recording are readable corroboration, but these full-image diffs do not isolate the row's comparison. The index correctly discloses these changes; disclosure does not repair the comparison. Recapture all three B checkpoints from matched fixture snapshots, with identical membership/order/completion/snooze/draft state, the same active control thread and toast phase, then supply twins and both unmasked diffs at 1440×1000 CSS/DPR1 dark. Record inactive failure, closed-folder attention, folder open, hover and notification.

2. **W — INSUFFICIENT for the narrated interaction claim.** The index says W-final-after.webm expands both reads during 0–4 s, then “collapse and Resume” during 4–6 s. At 5 fps around the action, the file paths/results remain expanded until continuation folds them into “Read 2 files.” There is no visible, settled manual collapse before Resume. In desktop-before-final.webm, the W segment retains “Syncing messages” and the collapse attempts do not demonstrate a resulting collapsed state. Clean W-expanded stills are usable and show the classification comparison. W-settled-before.png retains expanded reads while W-settled-after.png is grouped; this disclosed mismatch cannot prove a grouping-only diff. Recapture a clean W interaction on each side: expand both reads, pause, visibly collapse them, pause, then Resume where available. Capture matched expansion states before continuation; keep the post-continuation image separately as behavioral evidence. Retain both diffs and twins for the matched pair. This is a proof/narration gap, not a demonstrated code defect.

3. **Minor reproduction difference, N.** The independently rebuilt N-unknown-after image has identical visible text, positions and recovery state, but is not pixel-identical: Diff0 65,284/1,316,640 pixels (4.9584%); DiffT >5/255 1,683 pixels (0.1278%). Differences are concentrated on vector/control edges, rounded outlines and faint shadow/gradient pixels; text glyphs match. I cannot attribute every changed shadow pixel specifically to anti-aliasing. This is a small rendering reproducibility finding, not evidence of a different UI state. The independent twin is exactly identical. Installed Pi was 1.5.1 rather than the producer's 1.4.0; the requested browser rendering environment matches, and the real adapter received the same synthetic limit error.

4. **Nonblocking limits on broader behavior claims.** A demonstrates arm and cancel, not absence of dispatch after the original 10:42 deadline. PNT demonstrates native scheduled/snooze labels and successful manual Resume, not successful manual Wake (the device daemon command failed). Do not claim either unobserved result. These gaps do not undermine the Pi provider-error→usage-limit comparison, so they do not independently fail those rows under the requested proportional standard. The Saving state is visible in S-timer-after.webm, rather than A's short clip.

Before and after B, with unrelated fixture changes visible:

![B before and after, followed by Diff0 and DiffT](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/50e95fc2f618d8aa/B-hover-sheet.png)

W action samples at 5 fps; the expanded reads fold when continuation begins:

![W after action detail showing expanded reads followed by continuation](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b3ad59921fd70f7d/W-action-detail.png)

The row table uses SUFFICIENT for the classification comparison and preserves explicit behavior limitations. Existing accepted rows carry forward [pass 1](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/13acb87ec28dbb4f/proof-audit-1.md); U was spot-checked. New material comes from the [round-two index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/86cb34e79e545196/evidence-index.md).

| Row | Status | Evidence, scope and exact recapture if needed |
|---|---|---|
| C | SUFFICIENT, test-only | Unchanged non-limit negative control; accepted focused adapter evidence. No visual row required. |
| U | SUFFICIENT | Existing wide dark unknown-reset pair/diffs accepted; spot-check confirms raw red failure becomes amber notice/play. |
| Z | SUFFICIENT, shared visual | U renderer plus accepted zero-minute boundary test; no separate screenshot required. |
| T | SUFFICIENT | Existing wide light timed-reset pair/diffs accepted. |
| TN | SUFFICIENT | Existing 390×844 dark closed-info/wrapped-actions pair accepted. |
| R | SUFFICIENT | Existing failed-retry/unknown-reset pair accepted; supplemental continuation image shows reachable historical “Usage limit reached.” wording. |
| RT | SUFFICIENT, shared visual | T renderer plus accepted timed retry adapter evidence. |
| D | SUFFICIENT | Existing automatic-deadline recordings and clean terminal stills accepted with pass-1 overlay qualifications. |
| N | SUFFICIENT | 390×844 CSS/DPR2 dark: unknown/timed info open-close, armed labels and settled closed composer; all four quartets inspected. Independent head reproduction below. |
| A | SUFFICIENT, limited claim | 1440×1000 CSS/DPR1 dark: arm→Cancel→Resume states readable in recording and two quartets. Original-deadline no-duplicate behavior remains unproved. |
| S | SUFFICIENT | Same wide dark environment; four quartets and manual/global/combined/independent/deadline recordings show the reachable recovery UI. Independent clocks differ but descriptions and state changes remain readable. |
| M | SUFFICIENT | Same wide dark environment; draft/send→clear/play→manual continuation, historical unknown-reset warning and success; two quartets. |
| W | INSUFFICIENT | Classification shown by expanded pair, but narrated collapse is not demonstrated and settled pair uses different expansion states. Recapture clean expand→visible collapse→Resume recording and matched-state stills/twins/diffs as specified in finding 2. |
| B | INSUFFICIENT | Status/hover/folder/toast labels are visible, but all three comparisons contain substantial unrelated fixture changes. Recapture matched sidebar/main-pane/notification state with all three still checkpoints/twins/diffs and interaction recording as specified in finding 1. |
| PN0 | SUFFICIENT | Android Pixel 9, 1080×2424 physical dark: native unknown-reset card/play and list Limited label replace raw error/Failed. Twin delta is only 87 pixels in connectivity indicator, physical x954–959/y26–42. |
| PNT | SUFFICIENT, limited claim | Same native environment: dated reset, scheduled/owned-snooze links, successful manual Resume and history; two quartets. Successful Wake remains unproved. |

Provenance and image audit:

- Reviewed pushed [68414fb0aa](https://git.cafe/pingdotgg/t3code/commit/68414fb0aa3321c60eee6f714c4c5b586ed82b5e) against merge-base [633438d54f](https://git.cafe/pingdotgg/t3code/commit/633438d54f17b99f73f1a09466ecbe4a6aa0d545). The diff changes only adapter.ts and adapter.test.ts. The adapter and test are byte-identical to old captured fix [107c62505b](https://git.cafe/pingdotgg/t3code/commit/107c62505bbaa56b1f7d2c9cdd73a5fd66fc0831); old before [c6ca605630](https://git.cafe/pingdotgg/t3code/commit/c6ca6056309904f89b90b6055d797973be0a0939) is adapter-identical to this merge-base. The full provider-pi directory is not identical because main changed status.ts. Capture-column labels match these accepted identities; build SHAs are provenance claims in logs, not embedded screenshot metadata.
- Opened every new selected before/after still, both masks, and navigation sheets, including the supplemental retry-history image. The 20 checkpoint quartets have correctly sized/theme-matched originals; no selected still has ghosted/duplicated elements, loading skeleton/spinner, clipped/unrendered text, stale transient overlay or accidental blank render. Ordinary empty conversation space is intentional. Route/loading/session overlays visible in recordings were not treated as settled evidence.
- Recomputed all 40 full-image masks without alignment/cropping/masking: Diff0 max RGB delta >0; DiffT max RGB delta >5/255. Every mask matches the supplied mask byte-for-byte. [Per-checkpoint counts and verification](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e9d1ba04df4a53ed/selected-diff-results.json). Numeric agreement does not cure W/B state mismatches.
- Independently compared 40 selected before/after twin pairs: 39 have AE=0; PN0 after has the 87 status-indicator pixels above. The app content is stable in that twin. The index's claim of no settled PN0 twin is narrower than reality: a twin exists but is not exactly equal across the complete system screen.
- The “what else changed” list accounts for the visible regions, including composer focus/height, S's wake notice, continuation/history/minimap, W grouping, B fixture drift and native system indicators. Reduced motion was actually false because of the producer's wrong flag. This is a disclosed capture-condition gap; no motion/style change is introduced by this adapter PR. Differing clocks and native list membership are minor for the readable target card/list label comparison. Native logical DPR was not independently measured; physical dimensions match the declared device.

Recording inspection:

All 22 linked new/supplemental recordings decoded fully with ffmpeg and had sampled frames opened. Long clips were sampled every three seconds, short clips every second; A/W were additionally inspected at 5 fps and native action intervals at 1 fps. Six supplemental/raw recordings were inspected as supplemental material, not replacements for the selected clean captures. The 16 primary recordings were N-matched-before/after; desktop-before-final; A-final-after; S-manual-after, S-timer-after, S-deadline-after and S-independent-clean-after; M-final-after; W-final-after; B-before-final, B-final-after and B-notification-after; native-before-retry, native-after and native-after.part-002.

N's narrated info/route/arm/cancel sequence matches. A shows arming and cancellation within 5.4 s. S shows owned/combined Wake, global setting, delayed independent snooze and wake-only expiry. M shows draft switching and successful continuation. B shows the narrated labels/notification but has the fixture drift described above. PN0/native list and PNT dated/armed/snoozed/manual-resume/history states appear. Native timed detail begins nearer 64 s than the index's approximate 45 s; this timing approximation does not conceal an absent card. The failed Wake is correctly disclosed. W's collapse narration is the substantive mismatch. Historical retry heading “Usage limit reached.” is the actual renderer behavior, consistent with pass 1, and does not need a capture of unreachable retry-count wording.

Independent capture reconstruction:

Created an adjacent temporary worktree at the exact head. Installed dependencies, started isolated dev with T3CODE_DEV_SEED=0 and its own .t3, and used the uploaded seed/RPC/Pi-loopback-stub/narrow-driver scripts. Only scratch script port allocation changed to avoid reusing the producer's port. The real Pi CLI produced the unknown-reset failure through the server; the fixture was titled PN0 · Unknown reset. No fabricated DOM or live database was used. The supplied narrow driver's focus/blur/info procedure recreated N-unknown-after.

Rendering environment: Linux x86_64, headless Chrome **151.0.7922.108** (full version checked), 390×844 CSS, DPR2, 780×1688 PNG, dark, 16px root font, default browser scale, no device emulation; Date frozen 09:55 UTC, matching the delivered N procedure. Reduced-motion false matches the producer's actual environment, not its intended emulation. Waited for fonts and a clean settled route. An immediate independent twin had AE=0. Full RGB delta to delivered N-unknown-after is reported in finding 3, without masking or alignment.

![Independently rebuilt N unknown-reset capture at the reviewed head](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/d2a0517babf130d5/N-unknown-reproduced.png)

![Independent reproduction Diff0, nonzero RGB difference](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/5d21c9d4a2c7820a/N-unknown-reproduction-Diff0.png)

![Independent reproduction DiffT, maximum RGB difference greater than 5](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a525ed9be8432e96/N-unknown-reproduction-DiffT.png)

Scope: proof audit and independent runtime reproduction only. Existing tests/CI were not rerun; no new code correctness failure was established. All requested rows were assessed within the 40-minute budget. Server/stub processes were stopped using retained spawn sessions; owned browser tab/session closed, and adjacent worktree removed. Scratch is removed after this report is uploaded. Root filesystem had 106 GB available before cleanup.

FRICTION:
- tool | initial browser cookie import did not establish dev authentication; a fresh isolated pairing token entered through the pairing form worked | use the complete pairing form flow.
- tool | the shared Browser panel used a different Chrome version | use the supplied agent-browser renderer to match Chrome 151 for pixel reproduction.
- infra | a detached stub terminated with its spawning shell | retain an exec session for the stub and stop that captured session explicitly.
- provenance | selected and rejected raw N-settled captures shared basenames | resolve artifacts by their full indexed URLs, not basename alone.
- environment | installed Pi was 1.5.1 while producer reported 1.4.0 | pin the provider binary version alongside the browser environment for exact runtime reproduction.

VERDICT: INSUFFICIENT W, B