# Proof audit pass 2, final re-audit: #22997 (mirror #23000)

Head 9c3c3acbc6 (9c3c3acbc65b6cd756052c03ce9a41a905e40b90); merge-base 680152e27b (680152e27b2a961686389ca69a1557cbb2c55f08). Worktree HEAD and merge-base match; worktree clean.

## Findings
1. Blocking: 12/70 rows still have no evidence: C1, C2, A1, I1, V2–V9. The report admits this.
2. Blocking: P1, H1 and E1 still contradict the matrix expectation on both versions.
3. Blocking: R1 still has no command-palette or keybinding entry point. V1 is recorded in 4 grids but without the required collapse/expand sequence (the report says so).
4. W1 is unresolved (I opened repair-newrows-1440-dark before and after):
   - Segment 1 (`  - x`, ArrowLeft, Q): the frames match the narration. Both sides stay a literal paragraph (before shows "- Q|x", after shows "- x" with Q still pending because of skew).
   - In the leading-Tab segments, after is narrated as `ordered` (0:20.4) and `task` (0:25.3). Before is narrated as paragraph across the same span (0:19.6–0:58.7). The after frame at 0:21.3 shows "1. x".
   - The W1 expectation is that indented paragraphs are excluded by the normalizer. The capture does not show whether Tab moved focus or indented the text, so this before/after divergence is neither explained nor shown to be correct.
   - My status for W1: captured, outcome unverified. This is a possible regression.
5. Resolved since round 1:
   - The 12 previously incomplete grids (D1, S4–S6, G1–G5, I2, R1, R2) now have recordings, per the repair appendix. I did not open them.
   - K1 now has 4 grids. Its narrated before/after block states are identical (quote, bullet, task), which matches "no surprising conversion".
   - Timing claims are now labelled approximate.
6. The new repair rows (W1, K1, V1) are recording-only, with no still or diff pairs. The matrix marks them as recording rows, so that is acceptable.
7. The published final index (7251b40bba4095f1) is byte-identical to /tmp/pr-22997-capture-report-final.md. The inline block's counts (58/70 rows, 232/280 cells) are consistent with it.
8. Rebuild: I reused the round-1 independent F1 rebuild (1440 light, stopped thread, head):
   - Composer region: 15 px (caret only).
   - Whole frame: 11,484 px >16/255. The difference comes from the toast, the sidebar age and one header glyph.
   - No new rebuild was run.

## Row table
| Row | Status |
|---|---|
| M1, F1 | OK-verified (recordings opened in round 1) |
| L1 | Present; stills and diffs; inline narration consistent |
| M2–M5, F2–F5, L2–L5, T1–T5 (incl. F variants), N1–N7, B1–B3, O1, P2, H2, D2, S1–S3 | Present (L5 390-dark S has a 2:19 stall) |
| D1, S4–S6, G1–G5, I2, R2 | Present after repair (not opened) |
| R1 | Partial: Settings only, no palette/keybinding |
| W1 | Captured, divergence unexplained (finding 4) |
| K1 | Present after repair; narrated states consistent |
| V1 | Partial: collapse/expand missing |
| E1, P1, H1 | Contradicted |
| C1, C2, A1, I1, V2–V9 | Missing |

## FRICTION
- About 12 minutes of budget, so I opened only one repair recording pair (frames at 4.9/4.2 s and 22.0/21.3 s).
- Classifying an ordered list versus literal "1. x" text from pixels is ambiguous, and the narration doesn't say which W1 segment is meant.
- I used the bundled imageio ffmpeg and PIL with `python3 -I`.
- I started no servers or browsers. Downloads are in /tmp/pr-22997-evidence/audit2-final. No forge writes. Disk 5.8 GB.

VERDICT: INSUFFICIENT C1,C2,A1,I1,V2,V3,V4,V5,V6,V7,V8,V9 missing; P1,H1,E1 contradicted; R1,V1 incomplete; W1 unexplained before/after divergence