# Read-only proof audit: pass-2 correction closure

Finding first: no remaining blocking finding. The COMPLETE corrected index closes the sole pass-2 blocker: both sides of C3/C4 now really show matching first-post-J Stop 2 states; old stills/diffs/counts are explicitly withdrawn; older action times are qualified as tool wall offsets, not video PTS. No remaining incorrect active C3/C4 claim found across the introduction, corrected tables, amplified-diff summary, retained C3/C4 sections, row matrix, or limitations. Immutable historical indexes/report are explicitly superseded or labeled historical, not silently rewritten.

Reviewed index: https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/1d7d4616c5415cc4/unified-corrected-index.md

## Independent artifact verification

Downloaded the public four source videos, four corrected PNGs, four corrected gain-8 diffs, and stats JSON into /tmp/18300-closure-audit. All four video downloads are byte-identical to their original app-owned attachment files and have the stated attachment IDs/SHA-256. Independently ffprobe-decoded timestamps and ffmpeg frame-index extraction; all extracted RGB arrays exactly equal published replacement PNGs. Visually inspected all four full frames plus neighboring frame 15–18 crops: selected states are Stop 2 between Stop 1 and subsequent K returning to Stop 1. No replacement is merely an approximate seek.

| Pair | Before frame / PTS | After frame / PTS | Raw T0 / T16 of 1,316,640 |
|---|---|---|---|
| C3 | 16 / 1.427433 | 17 / 14.589900 | 157,882 / 3,793 |
| C4 | 16 / 1.427400 | 17 / 1.513233 | 45,406 / 3,177 |

Each frame is 780×1688. Independently recomputed signed abs RGB, saturated 8× gain, and max(raw channel delta)>16 before amplification: all four published diff arrays and both counts exactly match. Repeated references in all three table locations select the same corrected assets/counts. Narrow behavior already worked before; replacement index correctly claims retention rather than a newly fixed narrow regression. Host metadata/compression differences remain disclosed and unmasked.

## Row acceptance

| Row | Acceptance / disclosed boundary |
|---|---|
| 1 | Retained C1–C4 initial keys accepted; separate body-focused first mount still unverified. |
| 2 | Desktop stepping/reengagement accepted from prior audit; corrected narrow Stop 2 pairs accepted independently. |
| 3 | C1/C2/C5 chat handoff and C3/C4 sheet-close handoff retained and accepted. |
| 4 | Pointer/editable before+after and final keyboard-only transfer accepted; before keyboard-only remains unverified. |
| 5 | Active header/chrome ownership accepted; arbitrary standalone chrome permutations remain supplementary. |
| 6 | Both modifier reels accepted; log numbers fall under explicit tool-wall-time caveat. |
| 7 | Real competing menus and recovery accepted; old base routing difference disclosed. |
| 8 | Real palette ownership accepted. |
| 9 | Corrected narrow sheet ancestor comparison accepted; historical revisions retained honestly. |
| 10 | Hide/reopen and distinct closing-window comparison accepted from prior audit. |
| 11 | Real C5 pending→arrival complete, accepted from prior audit. |
| 12 | C5 spinner and C6 null accepted; discarded focus attempt is excluded. |
| 13 | Unchanged no-guide run/error/permission branches; no additional required proof. |
| 14 | C6 zero/one reservation and three-stop recovery accepted. |
| 15 | No-banner keyboard proof accepted; unchanged rewrite-banner gallery remains unverified and supplementary. |
| 16 | Ask context/composer accepted; no inside-guide file/editor fixture exists, unchanged callback boundary disclosed. |
| 17 | Actual model picker accepted; unchanged terminal exclusion remains unverified and supplementary. |
| 18 | Actual IME/defaultPrevented/isComposing/legacy-229 remains unverified; no focused new 229 coverage claimed. Supplementary limitation, not corrected affected proof. |

## Provenance and rebuild acceptance

Read-only local check: clean worktree at ce1aa4035b7512ccfa9fbb6e94ee840019b71121; its tree and fee1f231f67bd18feac0814c9cd34944cd48a9c3 tree both 66ea8dfd138400798ea233c7b884b8829b781d7a. Current merge history introduces no tree change. Base remains 51273020db3245f4ec29514a419d29b5dcb012dc. Prior production-equivalence audit remains applicable: core recordings at 877; closing compares prior candidate to e8; fee differs from e8 only by test removal. These distinct comparisons are accurately preserved.

Accept and reuse /tmp/18300-proof-audit-pass2.md independent rebuild: committed seed/data/README, stored guide exact, real Chrome154/Linux/1440×900/DPR2, trusted Stop1→Stop2 and empty composer. Compared with original C1 after: 19,816 T0 / 801 T16 of 5,184,000; mean per-channel delta 0.0207275/255, cursor/compression disclosed. Observed reproducibility, not pixel identity. Explicit user authorization for committed reproducibility seed/data governs; generic evidence-asset rules do not disqualify the fixture. No new runtime, server, browser or fixture write is needed for a correction that changes no source/runtime/fixture. Prior C1/C2/C5/C6 math acceptance is reused, not falsely represented as newly recalculated here.

## FRICTION

Web fetch returned Internal Error; curl retrieved the public index successfully. Python urllib downloads returned HTTP403; curl retrieved all artifacts successfully. Neither affected final verification. Sparse native PTS/seek ambiguity was resolved by exact frame-index extraction. Historical uploads remain immutable but are explicitly superseded. Correction-run historical checkout caveat is not confused with this clean current tree. No production/test/fixture edit, checkout switch, server/browser, test run, push, comment, forge mutation or public upload in this closure pass. Audit files only in /tmp.

VERDICT: SUFFICIENT. Exact remaining blocking gaps: none. Disclosed supplementary/unverified boundaries above remain; this verdict closes the affected proof-package correction and does not certify those boundaries as tested.