# Read-only proof audit — pass 2

PR [#24440](https://github.com/pingdotgg/t3code/pull/24440), issue [#7212](https://git.cafe/pingdotgg/t3code/issues/7212), pinned head [69c6d53c30](https://github.com/pingdotgg/t3code/commit/69c6d53c30d974da83f21ed5d9142dc7afd5f224). 2026-10-25 UTC. Evidence audit only; no implementation changes or new runtime capture.

## Findings and row statuses

**One remaining evidence gate: N3 native chat hit testing after Close.** This is a missing proof finding, not a confirmed product defect. The new connected recording proves connected chat input. It cannot establish that an Electron native contents surface stops intercepting input after Close. The retained completion report's Close/composer sequence is explicitly a stream sequence; its native clip covers Retry, healthy input and resize, without native Close/chat. Supply a retained native recording that visibly closes the failed float and types/selects in the exposed composer, or make that one supplemental recording in an authorized capture task. Do not repeat the established native recovery/re-float matrix.

| Row | Status | Audit result |
|---|---|---|
| P1 | RETAINED | Prior accepted protocol/status proof; no new gate or regression test run in this pass. |
| N1 | RETAINED | Prior exact independent native failed-frame reconstruction remains accepted. No rebuild repeated. |
| N2 | PASS, lineage resolved | Linked retained report and independently decoded/inspected native390dark clip establish refused Retry hold, then healthy same URL, Count1/native proof input and aspect restoration. The distinct new failure timestamp is retained transcript evidence, not a newly executed protocol observation. |
| N3 | PARTIAL / BLOCKING PROOF GAP | Native Close/tab retention/panel recovery/re-float retained from pass1; explicit native chat hit testing remains absent. |
| N4 | RETAINED | Accepted native continuation/persistence remains in force under the brief. No expanded persistence matrix. |
| S1 | PASS | Primary wide first frames show connected Agent has control, Take control, refused URL/reason, hint and Close; Retry absent. Actual takeover changes to viewer controls. |
| S2 | PASS | Primary wide approximately0.4–16.4s: takeover, fresh refused Retry, settled failure, same-URL recovery, real Count1 and typed S2 proof, then Release returning Agent. |
| S3 | PASS | Primary wide approximately22–38s: fail, Close, select/replace composer text, retained Browser tab selection, Refresh, Float, settled connected Agent endpoint. Logged source remains exact tab2. |
| S4 | PASS | Primary wide approximately52.3–53.8s visibly moves failed shell; log/stills agree x1108/y568 → x956/y348, unchanged320 square, delta(−152,−220). Composer subsequently accepts appended text. Supplementary wide-drag clip independently shows movement too. |
| S5 | PASS | Primary narrow approximately5–6.2s moves failed shell x58/y404 → x12/y204, delta(−46,−200), unchanged320 square. Approximately28.4–29.5s NW resize reaches x92/y284/w240/h240. Final frames show Release→Agent hint/Close/Take control→viewer Retry. Long address/reason and controls remain readable/reachable in original-resolution stills. |

**S1–S5's previous definite gaps are cleared.** The old nominal zero-movement drag is superseded by visible displacement with real mouse input and matching measured bounds. No ghosted or stale overlay endpoint was accepted as proof. Expected short navigation blank frames and Connecting during remount are transient, not claimed settled endpoints. The native recovery briefly shows blank white content before Healthy preview; acceptance uses its later Count1/input endpoint.

## Media and provenance checked

Sources: [pass1 matrix](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/d4911d7cd69fcf26/REPORT.md), [new capture report](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/4a39e38051ce0399/REPORT.md), [new archive](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/8dc1bb3ee5948487/all-captures-and-sanitized-drivers.tar.gz), [retained native completion report](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/862520fa8ae26cc6/REPORT.md), and [native390dark recording](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/c63286ecbda1aa08/native-390-dark.mp4).

Local origin is GitCafe. Read current PR metadata through cafe only: source fix/issue-7212, target main, draft, observed base [d53873b6a4](https://github.com/pingdotgg/t3code/commit/d53873b6a4799e87afd63c8d288ef0089f5a3748), head matches requested full SHA. Checkout HEAD matches and tracked status is clean before/after. Changed render/server/contracts implementation files checked against [b06c58abb6](https://github.com/pingdotgg/t3code/commit/b06c58abb634eaab1be2c30eb4bd84bf7dea5115) have no delta. Existing PR description still documents older incomplete proof; it was not edited.

All42 manifest entries (38 PNGs,4 MP4s) independently match recorded byte sizes and SHA256. All4 new recordings decode end-to-end with ffmpeg exit0. Inspected every new PNG through labeled sheets, original-resolution critical endpoints, all4 videos through1fps whole-app frame sheets, and10fps detail sequences for primary takeover/refusal, healthy input, wide drag, narrow resize/ownership, plus the complete4.533s retained native clip. This is sampled visual inspection with denser transition inspection, not a claim that a human inspected every encoded frame. Diagnostic crops were used only to magnify transitions; delivered whole-app media remain the evidence.

Directly fetched primary video URLs match archive bytes:

- [wide-S1-S4.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/adc0c8c6c0252766/wide-S1-S4.mp4): SHA256 `54b75eccfca0e70435e2166d1583da8dbe50122a7c41e6714420ccde560943ab`,1440×900,54.966667s.
- [narrow-S5.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/eed969d60deffdeb/narrow-S5.mp4): SHA256 `cebda3977c56a6e159a570d65e761db5a2bd168f2412aa6a3ef3b7d53e0e5ac2`,390×844,30.233333s.

Sanitized actions show real mouse down/move/up (24steps/900ms), elementFromPoint identifying exposed grip, geometry reads, and alert observer. No inspected driver uses a component setter, mocked failure or dispatched DOM pointer events to obtain the accepted endpoints. Product theme persistence precedes narrow flow. Synthetic fixture is a real HTTP server with Count/input/Fail target, recovered at59999; long fixture links to genuinely refused59999. Logs identify browser:tab_2_6a5e2fe3-8e52-42f1-89a3-2baacbcfce47 before and after retained recovery. The visible provider-start failure in surrounding chat is disclosed setup history, not a browser failure fixture.

Fresh S2 observer records old alert removed06:33:29.105Z and new refusal06:33:29.181Z. Capture report places listener start06:33:31.635Z; video holds the fresh refusal until subsequent Retry/recovery. Timing therefore avoids pass1's refusal/recovery race. I did not independently restart the listener. Media hashes verify integrity within supplied provenance, not cryptographic attestation of which source built a runtime.

Accepted narrow minimum endpoint:

![After — narrow dark connected viewer at240 square, readable long refusal and controls](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/c9d94de49bee8d67/S5-narrow-min-viewer.png)

## FRICTION

- No browser or dev server started: this remit is read-only with capturer cleaned. The retained report's request that pass2 rebuild its failed fresh bootstrap cannot be fulfilled under these constraints. Its `Error: missing milestone [role=alert]` remains an unresolved replay limitation. I did not invent a reconstruction result or redo established native Diff0=0 proof. This does not create a new fixture gate beyond pass1's minimal matrix.
- One exploratory source glob matched no files; corrected to actual preview component directory. No runtime action followed.
- Supplemental wide-ownership-recovery video contains setup failures, panel-selected internal Chromium error content, and waits; those frames are not substituted for primary settled S proof. Its final Float settles in later stills/primary clip.
- Native N2 clip is only4.533s at15fps and has no recording pointer. Its held refusal, native Count1/input and minimum resize are visible; the new refused-load event timing comes from retained report lineage. No fresh native input-event transcript was executed here.
- No new baseline reconstruction, whole-scene identity comparison, tests, unrelated provider/transport/theme products, or native mobile checks. Existing native exact-frame proof is retained, not re-certified.
- Required app-local PR registration/listing succeeded, watchingfalse. No public comment, edit, push, watch or merge. Own scratch removed after report upload; worktree retained unchanged.

VERDICT: INSUFFICIENT — S1–S5 pass and N2 lineage is resolved; N3's explicit native chat hit-testing recording is still missing. This is the sole remaining minimal-matrix proof gate identified in this pass, not a claim of an implementation bug.