# Fresh proof-audit recheck

Read-only audit of [#25379](https://git.cafe/pingdotgg/t3code/issues/25379), focused on the correction and real integration supplement. Read the prior pass-2 report, local index/provenance and live-proof transcript; inspected both live frame sheets, all three supplied stills, and independently decoded samples from all three actual MP4s (including their endings). No repo edits, children, servers, browsers, forge writes, merges or watches.

## Findings

- Real integration gap is closed for active Codex goal plus active Tasks: before.png and the start of before.mp4 show the empty cap above Tasks inspect, 0/3; after.png and the start of after.mp4 show `Pursuing goal · 22k tokens` above the same Tasks state. Both recordings reveal the goal objective on hover. The same title, prompt, running response and task state support the transcript's same-thread account. Earlier provider errors are historical timeline entries, while the new turn visibly runs. The exact component-swap procedure and DOM measurements are transcript provenance, not independently observable provider events in these videos. Stills visually agree with 12px versus 24px; source uses h-3 versus h-6.
- These are complementary live integration recordings, not matched pixel comparisons: working clocks differ (14s versus 39s in stills). Prior pass-2's independently checked harness matrix supplies matched rendering proof; this pass did not rerun that matrix. Sampled ending frames remain expanded, so the transcript's pointer-away ending should not be read as proof of return-to-collapsed behavior.
- Completed-goal still and actual observation video show `Goal complete · 23k tokens · 2m 22s` while a subsequent turn is Working, above Tasks inspect 0/3. This supports completed-goal labeling during another working turn. It does not establish completed-task retention: these tasks are still incomplete and were requested again. Hover timeout is disclosed; no expansion guarantee is needed or certified here.
- A5 correction is substantively present in index.md and provenance.md: the longer recording is explicitly limited to tooltip dismissal, and second-Escape collapse/focus restoration is withdrawn. **Correction is not fully cleaned up:** provenance.md line 36 still states `A settled rich-title tooltip consumes the first Escape; a second collapses the stack.` Remove or narrow that stale FRICTION bullet. Index also retains an old provenance link and old A1-unverified paragraph, though its supplement explicitly supersedes the latter. Do not present the old immutable provenance upload as the corrected record.
- Source identity checked: current [1449d0a663](https://git.cafe/pingdotgg/t3code/commit/1449d0a6635498656543a0d0e8ee98a53d2aeba6) has parents [62f38ad75b](https://git.cafe/pingdotgg/t3code/commit/62f38ad75bd152a22c7c8730750db3fd124d389e) and [72dde22b6f](https://git.cafe/pingdotgg/t3code/commit/72dde22b6f092491fac4ddc480bc609bb8092bfd). Both affected chat component files have zero diff against the captured revision and against working HEAD. Branch diff against merged main remains exactly those two files (28 additions, 6 deletions). ChatComposer/selector code is outside that diff. Completed-task retention remains a separate unchanged-path unknown.
- No listeners were present on capture ports 5733/13773. Working status remained the pre-existing untracked app.json.

Before: ![Live baseline: empty cap above active Tasks](/tmp/issue-25379-live-proof/before.png)

After: ![Live revised components: labelled active goal above Tasks](/tmp/issue-25379-live-proof/after.png)

## FRICTION

- One stale A5 second-Escape assertion remains in local provenance FRICTION despite the explicit withdrawal elsewhere; evidence wording needs that small cleanup.
- Video decoding verifies visible UI, not the hidden source-swap operation, exact capture timestamp, or provider tool payload. Those remain supplied provenance.
- Completed-task retention, full update hook, hardware touch, native clients and remote modes remain outside this focused recheck. No new implementation regression found.

VERDICT SUFFICIENT — for the narrowed rendering proof plus real active-goal/Tasks integration and completed-goal labeling during a later working turn. The evidence correction is partial, not fully done; remove the stale assertion before calling the record fully corrected. No A5 second-Escape or completed-task-retention certification.