# Fresh proof audit pass 2 — final recheck

Scope: [#22024](https://git.cafe/pingdotgg/t3code/issues/22024). Reviewed head [d70a9233b7](https://git.cafe/pingdotgg/t3code/commit/d70a9233b7de0e48997529dcb5347a92ebe8de8f), base [70e5e384da](https://git.cafe/pingdotgg/t3code/commit/70e5e384dab96a1d1e4a253ee5a9415ed5d8ce30), product [56809818ac](https://git.cafe/pingdotgg/t3code/commit/56809818acb0f3d35bd224faa086cd246b160923). Read the original pass-2 audit, corrected evidence, correction note, and corrected pass-1 acceptance matrix. This is a proof recheck, not a fresh full product/CI audit or merge authorization.

The prior common blocker is resolved. The original audit inferred that the fixture first existed in the 23:37 consolidation commit. That inference is contradicted by the restored original commits and the existing reflog. All eight original commit objects are current ancestors; their original committer times and reflog commit entries run from 22:50:25 through 23:29:21. They precede the applicable capture starts as the corrected per-row table states. These are not new commits given old dates. Consolidation temporarily hid their ancestry; restoring it satisfies the requirement to have committed fixture script and synthetic data before capture. It does not prove which script bytes each running process loaded.

I do not turn that remaining process uncertainty into a finding that no committed fixture existed. Nor do I claim complete historical process attestation. The corrected matrix required committed fixtures, shared inputs, real supervisor states, and observable behavior; it did not require a signed per-process revision record. The retained state traces, paired recordings, historical revisions and disclosed deviations are sufficient for that practical proof scope. This is an evidence judgment, not a waiver of a newly identified product bug.

## Verified corrections

- Independently verified all eight ancestor relationships, original committer times and all 29 non-null historical fixture file SHA-256 values against Git. The reflog independently retains the original commit sequence before consolidation. An absent early init.mjs is correctly null: seed script/data/proxy already existed; stable environment-ID initialization arrived later. Narrow light A2 is not attributed to that later setup.
- HEAD is exactly the requested SHA; working tree is clean. Its full tree is `5773252def0da17ca7e071f9d56448b558639e6f`, identical to the previous CI-fixed head [e210cea113](https://git.cafe/pingdotgg/t3code/commit/e210cea113). Excluding the fixture directory, HEAD equals the product commit exactly. The complete base/head diff remains the two product substitutions plus four fixture files.
- Inspected the fixture relocation/import changes: namespace Node imports, formatting, and one additional parent segment for browser module imports. No changed scenario semantics found. Scoped lint/knip/test pass claims remain prior validation, not tests independently rerun here.
- Independently downloaded all 16 new public standalone diff images and matched local bytes. Reconstructed every Diff0 and DiffT image exactly from the original stills: unregistered RGB absolute difference amplified ×8 and max-channel threshold strictly >16. All eight four-column tables exist. Threshold counts match the prior audit exactly.
- Corrected wide light A2 narration explicitly separates operation offsets, trace time and MP4 PTS. Inspected the labelled decoded-frame sheet and PTS manifest: click rings 6.948133s / 30.035367s, resets 7.378000s / 30.764633s, warning absent by 32.973800s. It no longer claims the second visible click happened at the logged 25.979s operation offset. Exact dispatch/paint latency remains unproved.
- Reuse of the prior committed-fixture rebuild is valid as supplemental evidence: unchanged product, unchanged scenario semantics, and retained images/metrics. Inspected its comparison. The reported 1,829 raw / 16 threshold pixels remain attributable to hover fill, caret and faint compositing differences, with no substantive layout discrepancy. This is a prior rebuild result, not a new run or a replacement for all paired recordings.

## FRICTION

- Exact loaded proxy revisions are not independently established for every historical process. Commit availability and loaded-process identity are different claims; the evidence now says so. No blanket proxy parity is granted.
- Narrow light A2 predates stable environment-ID/setup additions. Narrow A2 baseline natural recovery is not established in the retained clips; older proxy/preflight behavior limits that claim. Head action acceptance and recovery were sufficient in the original behavior audit, and the corrected A2 requirement does not require baseline natural recovery.
- Sparse recordings and unfrozen clocks/random do not establish smooth motion or exact timing. Wide stills are downsampled. Pixel counts quantify difference, not correctness; narrow A2 includes a substantive expected stack-height change.
- Desktop remains explicitly UNVERIFIED under the accepted hardstop. Responsive browser is not native mobile or physical touch. Relay/tunnel and additional provider execution remain unclaimed. No extra unchanged-control capture requirement is added.
- This recheck reused the original audit's inspection of all sixteen recordings and its real-client rebuild, rather than claiming a new independent playback/rebuild of every flow. No browser/server was started, so no owned server cleanup was needed. No product edits, commits, pushes, PRs, comments or watchers. Used the pr-audit and [ghdrop skill](/home/ubuntu/.codex/skills/ghdrop/SKILL.md).

## VERDICT: SUFFICIENT

Sufficient for the agreed eight affected web proof rows and handoff to Julius with the stated limitations. No remaining concrete proof blocker found within that scope; no product fix or mandatory recapture indicated.

| Row | Verdict | Retained qualification |
|---|---|---|
| A1 · 1440 light | SUFFICIENT | Retry before old deadline; two-action composition and recovery |
| A1 · 1440 dark | SUFFICIENT | Focus/Enter activation; Tab traversal not claimed |
| A1 · 390 light | SUFFICIENT | Responsive pointer interaction; not native/touch |
| A1 · 390 dark | SUFFICIENT | Hover contributes incidental pixels |
| A2 · 1440 light | SUFFICIENT | Corrected visible PTS distinct from operation time |
| A2 · 1440 dark | SUFFICIENT | Reset may temporarily hide retry; expected initial presentation |
| A2 · 390 light | SUFFICIENT | Older fixture setup; baseline recovery not established |
| A2 · 390 dark | SUFFICIENT | Baseline clip ends early; head acceptance/recovery established |

If a maintainer separately requires exact process-revision attestation for every row, that stronger claim is still INSUFFICIENT historically. It cannot be reconstructed from commit times. The minimal replacement set for that stronger per-row standard is the same eight affected base/head pairs (A1 and A2 at W4), which may be chapters in four paired recordings. Start fresh owned processes from a pinned committed fixture; record script/data/proxy hashes and PID/start/revision before capture; use that same version in both isolated checkouts; retain state/deadline traces, A1 early retry and delayed Disconnect fit, A2 held subsequent attempt/accepted request/recovery, and narrow expanded-stack access. No unaffected controls, extra desktop host, provider matrix, or relay/tunnel recordings are required. This conditional set is not a new blocking requirement of this verdict.

Reviewed by gpt-6.1-sol through the Codex harness.