# Final proof audit pass 2 — round 4

**32/32 matrix cells sufficient.** The supplemental D source-preservation evidence closes audit3's only outstanding finding. No new production defect or blocking proof gap was found. Parent owns final integration and publication; this verdict applies to the exact audited candidate, before any newer-main integration.

Target [#21140](https://git.cafe/pingdotgg/t3code/pulls/21140), contributor [SwiggitySwerve/t3code](https://github.com/SwiggitySwerve/t3code), branch `fix/web-seeded-model-no-handoff`, original head [093c34f](https://github.com/SwiggitySwerve/t3code/commit/093c34f843ecb338362636a54b93b241f9a8ecb1), base `main`. Candidate [a514925](https://git.cafe/pingdotgg/t3code/commit/a51492518429f805be2ed3f6d323d02382d98023), baseline [eea9895](https://git.cafe/pingdotgg/t3code/commit/eea989536849e5b5049987ee24318b5a77326542). Local branch `takeover/pr-21140` remains clean at candidate.

## Overall matrix

S = SUFFICIENT. Columns enumerate revision × viewport × theme; all 32 cells are individually accepted. Baseline cells are behavioral controls, not claims that baseline has the fixed behavior.

| Row | Before 1440 light | Before 1440 dark | Before 390 light | Before 390 dark | After 1440 light | After 1440 dark | After 390 light | After 390 dark |
|---|---|---|---|---|---|---|---|---|
| A — seeded selection/navigation | S | S | S | S | S | S | S | S |
| B — explicit selection/handoff | S | S | S | S | S | S | S | S |
| C — traits/empty options | S | S | S | S | S | S | S | S |
| D — links/carry/default/source return | S | S | S | S | S | S | S | S |

A/B/C and D links/destinations carry forward from [audit3](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/3541e328c31edb8c/AUDIT.md), with its qualifications. The [supplement](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b80b5a596d5fc322/SUPPLEMENT.md) supplies D's previously missing source return. Existing independent A replay remains sufficient: A-seed differs at eight whole-frame pixels with maximum channel delta1; the open A-seed picker matches exactly. No repeat source audit, A/B/C capture, or new runtime reproduction was needed.

## D finding resolved

Read the entire supplement driver, its snapshot/equality helpers, full raw state ledger, all eight accepted indices, and capture/processing/validation scripts. `seed()` is called exactly once in the executed D chapter, after the initial reusable destination is established. Each carry executes the real New thread button, then `goto()` clicks the actual sidebar row and waits for source identity. Each return calls read-only `selection()` before the stable still and open picker. There is no harness source write, reseed, reset, or explicit choice between leaving the source and these returned checkpoints.

The real explicit A pick occurs between the stale and explicit experiments. `resetDrafts()` clears destination draft-thread entries before the default-C experiment's **before** snapshot; project settings also change before that snapshot. Project settings clear only after the final returned source capture. Neither operation is inserted into a leave/return interval. Helper definitions left over from earlier capture work are not invoked by this D chapter.

Independently recomputed all 24 equalities from original before/after state objects, rather than trusting the supplied boolean assertions. Compared source environment/thread identity, all selection-bearing draft fields (provider, model map, options, explicit flag), and the source thread's own selection. Independently derived the draft selection subset from each full draft and verified it matches the ledger subset. Prompt narration is excluded deliberately; full drafts remain available. No selection data is hidden by that exclusion.

Across all eight configurations, stale source remains non-explicit A/high and own thread remains B/low. Baseline displays A/High; candidate displays B/Low Fast both before and after carry. After a real explicit picker choice, source remains explicit A with model-only draft selection (options absent), both after ordinary carry and after default-C fresh carry. Displayed A/Low and the selected picker row remain consistent. The fresh destination shows C/Luna. Exact reusable destination identities match and the fresh destination starts from an empty draft-thread map. D links and clipboard destinations retain their earlier accepted proof.

## Visual and artifact inspection

Opened and visually inspected all 40 final sheets: eight still sheets, eight video endpoint sheets, sixteen cross-revision diff sheets, eight preservation diff sheets. This includes every supplied >0 and >16 diff presentation. No essential blank text, clipped viewport, wrong theme, or obscured source selection was found. Responsive 390px composers collapse on return; their open pickers retain readable model identity. Before/return preservation masks show that responsive collapse as expected, not zero-pixel preservation. Desktop differences include sidebar markers, hover/relative-time decoration and disclosed notifications. Cross-revision stale selection changes include model/trait labels, selected row and picker height; explicit/default source models remain A.

Independently extracted 48 additional full-viewport video samples: six keypoints in each of the eight accepted recordings, at checkpoint time minus0.65s. These cover stale/explicit destinations, stale/explicit returned pickers, default destination picker and default returned picker. Opened all eight resulting sample sheets. Also inspected the supplied timestamped endpoint sheets for all16 checkpoints per recording. This is targeted sampling of full recordings, not continuous viewing of every frame. Full native extracted samples remain locally under audit4; sample sheets and timestamps are in the support archive.

The slow-request toast is present in some **stable** desktop stills and samples; these endpoints must not be described as toast-free. It occupies the upper-right region and does not cover the source composer selection or picker. It remains in full-frame diffs. It does not invalidate the claimed source-preservation proof. All accepted picker checkpoints visibly have the picker open; the earlier picker-dismissal failures were replaced by complete recordings, and no failed clip is needed for this verdict.

Independently checked exact revision metadata in all eight indices and ledger agreement, all128 native PNG dimensions (1440×1000 or390×844), stable flags, empty error arrays and matching H264/20fps video dimensions/durations. Chromium148.0.7778.96, DPR1, 100% scaling, reduced motion, fixed time and viewport geometry are recorded. Source SHA metadata establishes recorded provenance; pixels alone cannot cryptographically establish which source produced them. Retained hashes identify the exact inspected sheets and videos.

Independently recomputed **112 native whole-frame pairs**:64 cross-revision endpoint pairs and48 before/return preservation pairs. All raw absolute RGB images, amplified maximum-channel-delta>0 masks, >16 masks, counts, maxima and bounding boxes match the submitted artifacts exactly. No cropping, alignment, sidebar exclusion or pixel masking was applied.

## Scope and handoff

No source edits, checkout/detach, runtime launch, browser automation, tests, repo-wide checks, live userdata access, subagents, push, comments, watches or merges in this audit. No server or browser was started, so there are no owned runtime PIDs to stop. The supplement startup script uses isolated `/tmp/ev-21140` homes with `--no-seed`, `T3CODE_DEV_SEED=0`, and automatic bootstrap disabled. Synthetic projection replay retains audit3's limitation: it is responsive web proof, not native-mobile/Electron-shell or backend event-log coherence. U1's context-meter predicate constraint remains accepted from audit3 without repeating source review; U2/U3 remain scope qualifications, not new gaps in this matrix.

The PR is registered with this thread and not watched. Publication and newer-main integration remain with the parent. Audited by gpt-6.1-sol through the Codex harness.

[Supplement archive, scripts, stills and full videos index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b80b5a596d5fc322/SUPPLEMENT.md) · [Raw state ledger](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/80a32838046a9c46/raw-state-ledger.json).

![Before baseline A and after candidate B: source return preserves the underlying selection](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/83548df6f6ec41a5/source-return-before-after.png)

[Independent audit support: verifier, results, sheet/video hashes and 48-sample sheets](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/9d88f3bb1f70a1fc/audit4-support.tar.gz). Uploaded with the [ghdrop skill](/home/ubuntu/.codex/skills/ghdrop/SKILL.md).

FRICTION none | No new audit friction. | No prevention needed.

VERDICT: SUFFICIENT