Proof audit, pass 2 — [#29263](https://git.cafe/pingdotgg/t3code/pull/29263)

Finding: **D Light fails the independent reproduction requirement.** At the exact requested head, the saved fixture and driver produce a materially different full-device image: **375,511 of 2,617,920 pixels differ (14.34%)**, maximum RGB-channel delta 228; **157,226 pixels remain above 8/255**. This is geometry and flat-color variation, not anti-aliasing. Delivered input: `/tmp/lead-29263-r1/evidence-capture-2b/after-light-home-1793718019410881422/D.png`. Rebuilt input, preserved in the accompanying archive: `after-light-home-1793720992107092915/D.png`.

The rebuilt Search field and root menu are **79 px lower**: Search y=88 → 167; Environment y=267 → 346; Project y=393 → 472. The native status-bar window also differs (height 63 → 137 px). The menu crop x=392..1049, y=216..505 matches pixel-for-pixel after translating the rebuild downward by 79 px. Thus this is a rendering-environment reproducibility finding; it does not establish a regression in the PR's menu contents or sizing. The cause of the inset difference is unestablished. The recipe needs to reproduce the delivered inset/palette, or explain and substantiate the mismatch, before this gate passes.

![D Light: delivered after image, independent rebuild, and amplified exact pixel diff](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/6f1cbfc11367f329/rebuild-D-comparison-2.png)

[Pixel counts and input filenames](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/139e8023cc81966a/rebuild-D-result-2.json).

No additional contradiction was established in the delivered settled pairs, diffs, or narrated interactions. The D Light delivered pair itself passes visual inspection; its INSUFFICIENT status below is specifically the independent reproduction gate.

| Row | Light | Dark | Visually verified |
|---|---|---|---|
| A | SUFFICIENT | SUFFICIENT | Steer-first short draft; full after menu, both subtitles/check, reopen/dismiss/IME behavior, Queue action and queued result. Intermittent baseline clipping retained. |
| B | SUFFICIENT | SUFFICIENT | 720×1280, 360×640 dp, 22 pt; three-line draft, Queue-first order, width clamp, descriptions and Steer action. Dark third-open baseline clipping shown; Light baseline remains full. |
| C | SUFFICIENT | SUFFICIENT | 11 pt, left placement, both attachment icons/labels, reopen, Choose Files and return from native picker. |
| D | **INSUFFICIENT — independent rebuild** | SUFFICIENT | Untitled Environment/Project root, first/repeated opens and outside dismissal. Baseline clipped retries retained. Delivered root pairs themselves are usable. |
| D1 | SUFFICIENT | SUFFICIENT | Environment parent/back row, All environments, Proof environment, arrow return, two Android Back presses with IME consumption, changed selection/check and reset. |
| E | SUFFICIENT | SUFFICIENT | Model filters caption and both root rows; legacy toggle, checked reopen and Provider drill. Baseline clipping retained; after rows fully rendered. |
| E1 | SUFFICIENT | SUFFICIENT | Back Provider, All providers, Favorites, Codex and Claude; check moves to Codex, back restores root title, hardware Back and dismissal shown. |
| F | SUFFICIENT | SUFFICIENT | Exact 13-project fixture/order, four-line long label, overflow scroll to 12, checks on 01 and 12, back/root and dismissal. Delivered fresh-Metro after run reproduces all eight theme/checkpoint states with zero geometry drift. |

Coverage: all 16 retained row/theme combinations in the supplied matrix; 38 checkpoint tables, comprising all before/after captures and all 76 diff images. Every selected recording was decoded and visually inspected at opening/reopening, narrated actions and final states: 26 raw MP4s, including the independent F runs. The PR body's older nine still sheets and eight GIF attachments were also opened and inspected. D/D1 Dark use only capture 1; its superseded partial header was not treated as the combined result. No required row was omitted. G and optional H fixtures remain outside the fresh minimum.

Provenance: table headings and manifest commits match [47f2c8d29e](https://git.cafe/pingdotgg/t3code/commit/47f2c8d29e981c7723467efe137ee19e91df924e) → [3b020ab39b](https://git.cafe/pingdotgg/t3code/commit/3b020ab39b72d0a7be90f85515b41d22e5e7b87d). Normal captures are 1080×2424; B is 720×1280. Explicit Light/Dark, app text sizes, IME bounds and menu shapes match the stated fixtures. The exact changed file only replaces the inline Android Animated.View/FadeIn wrapper with View. Native popup, placement, action data and dispatch are unchanged by the diff.

Diff verification: independently recomputed all 76 masks; all are pixel-exact matches to their declared rendering. Capture 2a uses maximum-channel delta >0 and >16/255. Capture 2b and the selected capture 1 pairs use >0 and >8/255. Visually lit regions are explained by menu clipping/reflow plus the disclosed live provider prose/status, timestamps, Working timer/counters, caret phase, and tiny system-navigation differences. No additional unexplained region was established in the delivered before/after comparisons. Written step narration matches the visible action sequence; command timestamps are approximate recording offsets, not frame-exact synchronization. Recordings establish qualitative behavior, not a measured 120 ms transition duration. The old independent F mismatch remains historically unexplained; the new supplied F independent run has the same bounds and line breaks, with only 262 caret pixels or zero pixel difference across its eight states.

Independent rebuild: created a clean detached checkout at the exact head, ran vp install, installed the cached Android development client through mobile:install, seeded the saved 13-project fixture into a fresh isolated server home with T3CODE_DEV_SEED=0, and ran a separate Metro --clear on port 8095. Pixel 9/API36, 1080×2424, density 2.625, system font scale 1.0, Gboard [0,1541,1080,883], T3 Code Light, app16pt, Material navigation, Android animations off, native fingerprint 971743703b097b8870b39ff842d130f50b115eeb. The original driver captured D/D1, then its hard-coded Search tap y=130 missed after the changed inset. A second run changed only that tap to y=210 and completed the full home sequence; both runs reproduce the layout mismatch. No production source was edited. Rebuilt D also has an identical later screenshot partner.

Read-only scope: no edits to the reviewed checkout, no commits, pushes or comments, no PR mutations. CI and external review status were not audited. No repo checks were run for this evidence-only review. Completed within the 45-minute budget; no row or rebuild was abandoned at the deadline.

Cleanup: own Metro/server process groups stopped using captured spawn PIDs 2256624 and 2256361 after confirming each group leader and checkout cwd. Ports 8095 and 18196 no longer listen. Device close returned UNAUTHORIZED: Lease is not active, including after metadata refresh; session list then reported zero active sessions. Own detached checkout and scratch are removed after the credential-redacted evidence archive and this report are uploaded. Lead checkout, original evidence and scratch root are preserved.

FRICTION:
- tool | initial public ghdrop artifact downloads returned HTTP 403 | supplied local originals allowed the audit; an artifact endpoint readable by scripted downloads would avoid the detour.
- runbook | nominally matching pooled Pixel 9 rendering environment produced a different system inset and palette, and the driver's Search y=130 missed | record/reproduce the effective window insets and use a semantic Search locator.
- infra | device close reported Lease is not active; refresh did not change it, and session list was empty | clearer expired-lease cleanup confirmation would avoid ambiguity.

VERDICT: INSUFFICIENT D Light (independent rebuild differs materially)