Proof audit of round-2 captures for [#25761](https://git.cafe/pingdotgg/t3code/pull/25761), head [bbb5f62a6f](https://github.com/pingdotgg/t3code/commit/bbb5f62a6ff0346b82fd9779187098bec572088a), before [71646e031a](https://github.com/pingdotgg/t3code/commit/71646e031a6dc9db1309b3b9dff597f872cf2fd3). Read-only; no source edits, pushes, comments, host actions or fixture resets. No round-1 re-audit or whole-matrix verdict.

Findings:

1. C03 lacks the recorded initial first-page response needed to prove its cursor-only precondition. In both accepted C03-before-retake-rpc.json and C03-after-retake-rpc.json, the first recorded pullRequests.list request is cursor request64, answered with8,7,6,5. All no-cursor99-row requests in these logs occur after close starts. Initial UI text is virtualized:41 rendered list rows, ending9, not a complete99-row response. The uploaded supplemental archive contains scripts and inspection sheets, no missing initial response. Thus “PR5 absent from first99” is plausible and consistent with the cursor boundary but not directly captured as required. Retain the initial99-row response on each side before Load more. The actual cursor response, close action, Closed detail and final placement difference are captured.

2. A03/D01 are diagnostic only. A03-D01-detail-before.png and -after.png plus both detail diffs show OS clock/status-icon differences and a one-level Changed files separator difference outside this PR. List stills/diffs show the OS clock difference. Both videos begin with unequal ages, and the bottom CTA begins Work on PR before versus Continue thread after. Native clock, DPR/logical viewport and reply/paint instrumentation are uncontrolled/unmeasured as the index itself discloses. D01 never executes a list long-press/context-menu action. These are not clean native comparison proof. This is a limit already honestly disclosed, not an undisclosed claim of success.

3. Loading frames are excluded from settled proof: A04-before.mp4 near0–1s comments skeleton; B14-before.mp4 near2s detail skeleton/3s comments skeleton; B15-before.mp4 near4s and14s comments loading; C02-before.mp4 near19s temporary list loading/blanking. They show transitions, not settled endpoints. No accepted web endpoint still contains a skeleton, spinner, stale overlay, ghost/duplicate element, wrong theme/viewport, or unexpected blank content. Empty stopped-thread chat surfaces and intentional hover cards are correctly rendered.

C03 placement is intended, not a discovered code defect. patchPullRequestListEntries copies updatedAt and sorts changed rows descending at [pullRequests.ts:213](https://github.com/pingdotgg/t3code/blob/bbb5f62a6ff0346b82fd9779187098bec572088a/packages/client-runtime/src/state/pullRequests.ts#L213). The server also sorts descending at [PullRequestService.ts:1562](https://github.com/pingdotgg/t3code/blob/bbb5f62a6ff0346b82fd9779187098bec572088a/apps/server/src/pullRequest/PullRequestService.ts#L1562); merged client pages are sorted by recency at [pullRequestList.logic.ts:574](https://github.com/pingdotgg/t3code/blob/bbb5f62a6ff0346b82fd9779187098bec572088a/apps/web/src/components/pullRequest/pullRequestList.logic.ts#L574). The final C03 difference is25135 Diff0 pixels/6121 Diff16 pixels, confined to list placement. It does not isolate causal patch behavior: baseline post-action normal all99/id69 and all198/id81 omit5; head all99/id73 and all198/id85 already return5 first. Both builds reread expanded198 data. The index accurately disclaims causal isolation and zero-data-reread/host-count proof.

| Round-2 row | Status | Reason within this capture round |
|---|---|---|
| A04 | SUFFICIENT | Ordinary reopen is visible; identical settled row/detail; recorded detail reads2→1, with head read spanning reply. |
| A05 light | SUFFICIENT | Merge/confirmation→Merged, matching fixture identity/light profile, no merge action afterward; exact endpoint match. |
| B14 | SUFFICIENT | Real description link Open hover→internal detail close→Closed rehover; exact endpoint match; description-only scope disclosed. |
| B15 | SUFFICIENT | Actual composer chip follows the same close/rehover sequence; one RGB1 pixel, zero above16; limited profile disclosed. |
| B12 | SUFFICIENT | Stopped thread linked row/tab settle Closed via modern snapshot; exact match, eventual-only claim. |
| B19 | SUFFICIENT | Same recording includes linked sidebar badge/footer0open; no claim of immediate reply patch or other badges. |
| C02 | SUFFICIENT | Four-row cursor response retained, head-row close and final Closed detail; exact match; expanded198 reread admitted. |
| C03 | INSUFFICIENT | Cursor/action/placement captured, but required initial99-row exclusion response missing. Sorting is intended. |
| A02 | SUFFICIENT | Convert-to-draft chapter visibly reaches Draft/Ready for review before next action. |
| B05 | SUFFICIENT | Shift-ready chapter returns ready/open/Merge; final endpoint exact match. |
| B07 | SUFFICIENT | Shift-reopen chapter opens the closed row; no filter-exclusion claim. |
| A03 Android | INSUFFICIENT | Uncontrolled native comparison; unrelated OS/CTA/age differences, diagnostic only. |
| D01 Android | INSUFFICIENT | Diagnostic detail-close/list-removal sequence, no list long-press/context-menu action. |

Validation coverage: opened every accepted before/after still, all20 full diff images and10 labelled sheets. Checked SHA column labels, A05 light versus other web dark, whole-frame dimensions,16 web terminal/probe environments1440x1000/DPR2/Chromium156/Linux/16px, and Android1080x2424 portrait/dark. Recomputed all pixels and both diff masks: zero mismatches with uploaded processing. Every within-side settle pair is pixel-identical. Six web pairs have0 changed pixels; B15 has1/max1; C03 has25135/max241. Native detail2166/max235 and list604/max235 match disclosure. All recorded PR requests in16 accepted web logs have terminal Exit responses. Independently read recorded action/read/cursor payloads and relevant source behavior.

Opened all18 accepted recordings as continuous1-second frame contact sequences across their full durations, with enlarged action/hover/endpoint frames. Written index narration matches visible action order, including the three draft/ready/reopen chapters and native confirmation/navigation. This is sampled visual review, not a claim of examining every encoded frame or measuring compositor paint. The clips have provenance banners; narration is accompanying index text, not audio. Native cloned4-second tails are disclosed and are not measured live holds. Superseded failed attempts were not counted as accepted proof.

Independent rebuild: isolated detached head/fresh database/read-only fixture clone, exact Chromium156/Linux1440x1000/DPR2/16px/dark rendering environment and uploaded seed/probe/finish recipe. Rebuilt A02/B05/B07 final PR2 endpoint without replaying host mutations. Two own frames750ms apart are identical. Unmasked delta against uploaded head:158/1137920 pixels at threshold0,28 above16/255,max140,bbox(185,195,298,833).103 pixels are list avatar rasterization;55 are fresh database Settled0 versus uploaded Settled2. PR content/geometry matches. See replay-description.md, metrics.json, replay.png, diff0.png, diff16.png, replay.mp4 and browser log. Nonzero data-state differences are retained, not masked.

Cleanup: own preview tab closed, own dev-runner session75422 stopped, ports14385/6345 no longer listening; own source checkout remained clean. No emulator/lease started. Own worktree and scratch are removed after artifact upload; lead checkout remains untouched. Startup credential line is excluded from uploaded dev-redacted.log. Downloaded source evidence remains at its original uploaded URLs.

FRICTION:
- tool | Browser event export exceeded64KB; recovered in12KB chunks. | Use uploaded chunked exporter from the outset.
- tool | Direct127.0.0.1 preview navigation returnedERR_CONNECTION_REFUSED; environment-port navigation worked. | Use environment-port target for this preview runtime.
- brief | Required C03 initial99-row response is absent from accepted raw traces. | Start the probe before the initial list load and retain both complete page responses.

VERDICT: INSUFFICIENT C03, A03, D01