Findings first

No blocking visual evidence finding for S1, S2, I1, S3 or I2. No blank/loading capture, ghosted text or unrelated before/after change was found. The narrow supplemental table intentionally overflows; the recording reaches its rightmost column and returns. User-message fading is the existing collapsed-message presentation, not loading.

Nonblocking artifact limitations:
- The remotely delivered uploads.json has 144 entries; local final uploads.json has 146. Its missing entries are png-ihdr.json and uploads.json itself. Both are independently linked in the delivered index and fetch successfully. All other 145 delivered indexed files match local bytes by SHA256, including every image, recording, fixture, source provenance, measurement and diff. This bookkeeping snapshot is stale; the visual proof is not.
- Fresh reload returns the mobile composer to a compact resting state. A fresh head S1/S2 390x844 dark reproduction differs in 38,680/329,160 full-page pixels, 7,457 above threshold16, max channel delta229; bounds [12,640,378,828]. The table/paragraph/fenced region [16,110,374,427] differs in zero pixels, and every pixel above y640 is identical. The initial recapture before reload was fully pixel-identical (zero changed pixels). Thus full-page fresh-reload identity is not claimed. The composer/scroll-to-end state is outside the affected behavior and is unchanged within delivered before/after pairs.
- Desktop preview twice reported an exported1280x800 screenshot while JS reported1440x900 DPR2. Correctly sized Electron export remains unproved; authorized Chromium fallback is the matrix evidence.
- Private runtime handoff's backend PID is stale. Current listener ownership was checked against this worktree; no server was started or stopped. Concrete process details remain private in local handoff context.

VERDICT: SUFFICIENT — S1, S2, I1, S3, I2, within the supplied matrix's Chromium web scope.

| Row | Sufficiency | Independent checks |
| --- | --- | --- |
| S1 | Sufficient | Four matched before/after pairs:1440x900 and390x844, light and dark, DPR1. TH, mixed TD and code-only TD12→10.5px; cells12px. No new clipping. |
| S2 | Sufficient | Same four pairs: paragraph prose14px/code12px, fenced code13px unchanged, visually unchanged and outside diffs. |
| I1 | Sufficient | Both revisions' wide-light and narrow-dark recordings decoded and sampled every2s, plus all20 delivered review frames. Collapse→expand→collapse confirmed, actual toggle present. Expanded wide wrapping remains usable; narrow horizontal overflow preserved. Lanes728px and358px straddle576px threshold. Head code10.5px at all recorded states. |
| S3 | Sufficient | Wide-light, narrow-dark left/right pairs inspected. Path/file chips are A.chat-markdown-file-link, not CODE,10.32px unchanged. Plain external label12px unchanged; actual code inside external link12→10.5px. |
| I2 | Sufficient | Both narrow recordings visibly reach hidden Control/file-chip/code-link cells and return in both states. Collapsed scroll0→186→0; expanded before0→739→0, after0→726→0. Controls remain usable. |

Source/environment and diff integrity

Baseline [eea9895](https://git.cafe/pingdotgg/t3code/commit/eea989536849e5b5049987ee24318b5a77326542), head [1cff6be](https://git.cafe/pingdotgg/t3code/commit/1cff6be879da1dcd1153d8d4d2b78a429a5238cf). Git confirms only five CSS insertions differ; clean head remains checked out. Recorded baseline/head CSS blobs814cfa484dcd61efa59031d555109b0c27f08a17/29d7027309309537180d17203913b9147f877599 agree. Baseline CSS swap reconstructs the entire baseline source because this is the only changed file. Historical capture execution itself cannot be independently witnessed after the fact.

Every delivered primary still and both diff variants were visually inspected in lossless comparison sheets. All18 settled duplicates were opened/decoded and verified pixel-identical to the inspected primary stills. All20 delivered video-frame stills and both diagnostic exports were inspected. All four videos were decoded independently, not accepted from ffprobe alone. Nine Diff0 images were independently recomputed as saturated absolute RGB delta×8 at threshold0; nine DiffT images as magenta max-channel delta>16/255, all pixel-exact. Column redistribution, wide wrapping and faint composer blur differences explain visible before/after differences; no paragraph/fenced changes.

Matched metadata verifies viewport/client/visual viewport, DPR1, effective theme, scroll, fixture text, nonstreaming assistant/user roots, lane, source hashes and unchanged font family/chrome. PNG raster1440x900/390x844 agrees. Chromium148 Linux, system-default theme with emulated media, reduced motion and loaded fonts. Fixtures/scripts remain outside source. The supplied seed script was read/run (it only prints fixture text); existing normal-composer seeded data was re-rendered by full reload, without submitting another agent turn or writing projections.

Reproduction artifacts: /tmp/issue-24166-audit/reproduced-S1S2-390-dark.png, reproduced-S1S2-390-dark.json, reproduction-delta.json, fresh-reproduction-DiffT.png. Independent file/diff checks: remote-integrity.json, independent-validation.json. No source edits, commits, pushes, public comments, PR activity, spawned agents, or new servers. Runtime/browser sessions retained for lead cleanup.

Limits: matrix-scoped responsive web proof; no native mobile, shell/IPC, other providers/callers or remote connection-mode exercise. Video review sampled frames rather than every frame; settled stills supply legible typography proof. Fresh full-page composer state differs as disclosed above.

FRICTION category|happened|avoid
export|Preview CSS1440x900 DPR2 exported1280x800 twice|Correlate JS metrics and PNG raster before evidence; use one authorized fallback for both revisions
bookkeeping|Uploaded uploads.json predates final two uploads|Generate final inventory after uploads, or label it as a snapshot
reproduction|Reload reset the mobile composer and created unrelated lower-page delta|Record composer focus/resting state and distinguish whole-page from affected-region deltas
runtime|Backend PID in private handoff no longer owns listener|Resolve current listener and verify cwd before cleanup; never kill from stale handoff