# Final evidence audit — [#23950](https://git.cafe/pingdotgg/t3code/issues/23950)

Read-only audit started 2026-10-24 13:27:42 UTC; completed approximately 13:31 UTC, within the strict 10-minute limit. No source/state edits, browser, server, rebuild, push, merge, PR, or public comment. Only audit artifacts under /tmp were written and uploaded.

Combined evidence is sufficient for the requested visual rendering proof at capture head [0593c8bc89](https://git.cafe/pingdotgg/t3code/commit/0593c8bc89712e399a758aaf4c3fb2c07a88a0cc), against base [a46f80ee28](https://git.cafe/pingdotgg/t3code/commit/a46f80ee28d9b498bb8b407c8b60bd4e93bbe2e3). Use the [repair index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/cb2aee29c0587c23/repair-index.md), preserving its 20 valid original pairs and replacing only its specified 28 pairs. Do not use superseded original failures as verified evidence.

| Row | Status for combined set | Proof / remaining qualification |
|---|---|---|
| S | VALID | Desktop light now settles exactly; desktop dark has stable (1) shelf counts at endpoints and sampled collapsed states. All three shelf labels/counts and hide/show visible. Retain original narrow light/dark pairs. |
| R0 | VALID plain-header rendering | Desktop dark layout is matched; repaired child reads Done. Retain original other three combinations. Those show Idle: completed-child business behavior remains UNVERIFIED there; they validly cover the same plain-header visual branch. |
| R1 | VALID | All four combinations, both states: visible 1 failed accessory, Failed child when expanded, count suffix when collapsed, separator and chevron legible. |
| W | VALID | All four combinations, both states: Run/Done/Fail phase chevrons and accessories visible; Done 2/2 agents provides plural coverage; Run/Fail provide singular. Narrow full-frame before/after composer matches. Outer workflow toggle remains an orientation reference. |
| D1 | VALID | Repaired desktop-light endpoints settle exactly with emphasized Settled during drag. Retain original dark desktop and narrow combinations. Head replacement recording includes unrelated D2 attempts; use only its D1 segments. |
| D2 | VALID | Repaired collapsed accent in all four combinations; visible accent Settled (1), down chevron after / up before, slim dragged Snoozed source, selected Settled-row retention. Retain original valid expanded-accent grid. Held-drag rendering only, no drop persistence claim. |

## Independent checks

- Read pass2-audit.md, matrix.md, evidence.md, repair-index.md, repair-analysis.json, repair-uploads.md/json and relevant inspection/analysis/upload scripts. Downloaded supplied remote repair-index.md; byte-identical to local copy.
- Opened all 14 repair-inspection sheets, the full-frame repair-diff-inspection sheet and brand-inspection image. Also opened full-resolution narrow W and R1 endpoints and a separately enlarged before/after brand crop.
- Independently recomputed every repaired pair using signed RGB subtraction: all 28 exact/threshold pixel counts match repair-analysis.json; all 28 Diff0 arrays equal abs(RGB difference) x16 clamped, and all 28 DiffT arrays equal magenta where max channel difference >16. No masks. All 56 repaired primary/settle comparisons are RGB-identical.
- Independently recomputed all 20 retained valid pairs' Diff0/DiffT arrays and their 40 settling comparisons. All match exactly. Combined total: 48 endpoint pairs, 96 exact primary/settle comparisons. This verifies pixel equality; the 500ms interval is supported by capture narration/driver, not encoded cryptographically in still images.
- Opened all five supplied recording sample sheets; independently extracted focused 1/4/7s samples (clamped to duration for short clips) from all 35 repair recordings, plus final-frame samples. Opened all nine resulting focus sheets and final-frame montage. Shelf/R/W return states are visible in final samples. D2 samples show held accent and cancellation/retention as applicable. This is sampled inspection, not frame-by-frame or animation-smoothness certification; animation was disabled for reproducibility. Retained recordings rely on the prior pass2 audit, which inspected all 40 originals.
- Downloaded and SHA256-compared six representative remote repair assets: S desktop-light base collapsed; R1 narrow-dark head expanded; W narrow-light head expanded; D2 desktop-dark base collapsed; D2 narrow-light head collapsed; repair-analysis.json. All match local bytes. Other remote repair assets were not individually downloaded/hash-certified in this pass.

## Desktop brand raster delta

Eight repaired desktop shelf/drag pairs contain 266 changed brand pixels in light or 267 in dark, confined to text at exported coordinates x[46,96), y[18,29). These pixels are preserved in full-frame diffs and included in reported counts. Enlarged inspection shows the same T3 Code glyphs, placement and surrounding geometry with edge/intensity variation consistent with rasterization/antialiasing. The exact rendering cause is UNVERIFIED; do not claim experimentally proven antialiasing or chevron-only full-frame equality. This small isolated raster variation does not invalidate the shelf/header orientation comparison. It is materially different from the original mismatched sidebar width/composer problems.

## SHA and narration limits

The capture index names the requested head/base consistently. Images alone cannot authenticate a Git revision; capture provenance relies on retained capture records and the prior reconstruction audit. At audit start the checkout had already advanced externally to [e2e2af57b3](https://git.cafe/pingdotgg/t3code/commit/e2e2af57b3f9c802d804f68731543f3fe1d571b4), a clean merge commit timestamped 13:27:40 UTC whose first parent is the capture head. Thus the index's historical “restored clean at 0593...” statement must not be presented as the current checkout SHA. Relevant header/caller files have no changes from capture head to that merge; composer and other files do change. No new-head rebuild was performed, and this report does not certify full visual equality at the merge SHA.

These projection fixtures establish rendering, not backend orchestration or native-mobile/Electron-specific behavior. Desktop means desktop-width web rendering; narrow means responsive web, as the original matrix specifies. No broad source re-review was performed.

| FRICTION category | issue | prevention |
|---|---|---|
| Provenance | Checkout advanced after capture; static report still names old current head | Label capture SHA separately from current checkout/PR head; never relabel old screenshots as new-head captures |
| Rendering | Small isolated desktop brand raster variation | Retain unmasked diffs and exact counts; describe it as consistent with rasterization, without inventing a proven cause |
| Coverage wording | Retained R0 Idle fixtures could be narrated as completed-status proof | Claim plain-header rendering; explicitly leave completed business behavior unverified in those three combinations |
| Recording inspection | Fixed 7s sample exceeds some short clips; reused temporary frames can look like fresh samples | Probe actual duration, decode fresh bytes, clamp timestamps and inspect the final frame; state sampling limits |

VERDICT: SUFFICIENT