# Evidence-only audit2 for [#24577](https://git.cafe/pingdotgg/t3code/pulls/24577)

**SUFFICIENT for the requested evidence scope.** All 34 rows are supported by inspected matched recordings and sidecars, with explicit limitations below. No new product regression found. This is evidence sufficiency, not a merge-readiness or universal browser-compatibility verdict. Completed 2026-10-25 01:17 UTC.

Exact comparison: BEFORE [c4fa4629](https://git.cafe/pingdotgg/t3code/commit/c4fa4629eb645b95aae2beef56c70f705b736049), AFTER [69b0eb2c](https://git.cafe/pingdotgg/t3code/commit/69b0eb2c4c9b20210a40c1fad35977ba056aa946). Model gpt-6.1-sol, Codex harness.

All 34 new requested recording rows were independently inspected on both sides: 68 core videos plus all 26 supplementary videos supporting the same rows. No static pair was redone. The audit opened every decoded contact-sheet page, correlating visible sequences and target overlays with complete DOM/network/input sidecars, not merely inspecting encoding metadata. There are 615 pages / 6,984 sampled temporal frames; event checkpoints and regular frames cover all recorded steps, with denser first-paint mutation/gate samples. This is finite temporal sampling, not inspection of every encoded frame. `inspected.json` records per-clip inspection timestamps and SHA256; `final-inspection.json` independently verifies all current final video hashes and exact revision IDs. Diagnostic failed attempts are excluded.

Independent sidecar checks: 68 core clips / 394 contexts / 1,688 DOM states, plus 26 supplements / 98 contexts; zero oracle errors. The checks correlate prior target href to intercepted native request, profile/touch identity, phase-specific platform and label/icon, disabled-JS/no-fetch behavior, centered focus visibility and full resize checkpoint order. Decoded visual review is separate from these checks. No new product regression was observed.

- I01: all seven profiles, hero and lower primary, 390×844 and 1440×900, fresh touch/pointer and real Tab/Shift+Tab/Enter contexts. Before mobile routes to the desktop installer fixture; after iPhone/iPad routes to App Store and Android to Google Play. Desktop destinations remain fixture DMG/EXE/AppImage; unknown remains Releases. Actual native navigation requests agree with prior DOM.
- I02: real focus, held down/up active states and hover on fine-pointer profiles. Touch-only hover is N/A. Core framing gaps are explicitly retained as limitations; matched supplements show fully framed target outlines and both targets at all breakpoint triplets, including 419/420/421 in both directions, plus actual footer, decorative git hint and hero return. The macOS core AFTER final caption claims a hero return that is not visible; its supplement supplies the actual return. These are evidence presentation gaps resolved by supplement, not observed product regressions.
- T01/T02: three mobile profiles, controlled exact-homepage-module gates and natural ungated throttled navigation, hero and lower viewports. Real lower keyboard activation supplements cover all phases and widths. Before script execution both revisions show default macOS/Apple/Releases. Before revision then detects desktop platform and waits for release fixture; after revision synchronously assigns the mobile store with zero release requests. The AFTER “pre-release” checkpoint is a matched milestone, not an actual pending release request. Existing webfont reflow and fallback changes are visible. No no-flash/no-CLS or universally safe pre-script mobile activation claim is supported.
- T03: all seven profiles with page JavaScript disabled before navigation, both targets, both widths and pointer/touch plus keyboard activation. Default macOS/Apple/Releases remains, platform attribute absent, no release fetch. Actual footer store links remain explicit stores; footer Download remains /download.

Requested behaviors are covered by those 34 rows; pass1 guidance extras are not grounds to expand indefinitely. The 13 supplemental pairs close concrete framing/timing gaps in the requested rows. Native Safari on an actual iPhone is explicitly unavailable and remains unverified. These Chromium UA/touch emulations do not establish Safari/WebKit, hardware behavior, normal-motion smoothness or GPU performance. Intercepted 204 navigation proves native target selection, not store landing/install/delivery. Before fixtures are deterministic release assets, not live release verification. A transient icon-only button while HTML streams and existing font reflow are observations, not new regressions.

No source changes, commits, pushes, merge, PR comments or body edits. No T3 server or live userdata access. Capture checkout/display/server was left alone while owned by capture. All own subprocesses were tracked at spawn and stopped; their exact PIDs no longer exist and separate port 53543 is closed. The checkout remains clean at the exact AFTER revision. Scratch remains for parent cleanup.

Friction: tool — finite video sampling and core framing/overlay limitations; brief — lower primary versus actual footer terminology and proposed guidance extras; runbook — exact-module gating and capture-completion handoff; infra — native iPhone Safari unavailable. No forge/policy friction observed.


Independent reproduction: **PASS**, I01-iphone-hero AFTER only, headed Chromium 145 under owned Xvfb, own foreground marketing server and ffmpeg recording. Four fresh contexts (390×844 / 1440×900 × touch / real Tab, Shift+Tab, Tab, Enter) produced four native App Store GET attempts and zero release requests. Both keyboard cases had fully framed focus-visible outlines. BEFORE was independently audited from the matched capture, not recaptured. All eight reproduction sheets (85 frames at 250ms spacing, 21.2-second video) were opened. Post-subcase screenshot calls visibly rescale the emulated page; those intervals are capture-tool behavior and do not establish product resize/motion behavior. The independent video has no target overlay; the script and timestamped sidecar identify its four contexts and inputs. Browser/recorder/server/driver/Xvfb PIDs were captured, then verified absent; no process killed by pattern.

Capture provenance: [final capture index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/32b1c67357810ca7/evidence-index.md) · [delivery manifest](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/77397206893289c4/delivery-manifest.json). All 94 video hashes and 94 sidecar hashes match that immutable manifest (188/188); all exact before/after SHAs match. Capture reports 432 native navigation requests, 492 contexts, zero validation/decode errors. The audit independently checked its own recorded counts and oracle results, not merely the capture summary. All capture video links and sidecars remain in that index; no duplicate capture upload was needed.

Audit scripts, per-row verdicts, hash ledger, sidecar checks and cleanup proof:

[audit-scripts-results.zip](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/5b18feaa60b213c4/audit-scripts-results.zip)
[row-verdicts.md](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/3252851d8a0e58cf/row-verdicts.md)
[final-inspection.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/82f6090882f67c1f/final-inspection.json)
[delivery-crosscheck.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/28012c0367eddbe7/delivery-crosscheck.json)
[own-cleanup-verification.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/def6efdc2b31027b/own-cleanup-verification.json)
[reproduction.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/3bd714e73ecd4882/reproduction.json)

Independent reproduction and scripts (ordinary video link):

[reproduction-I01-iphone-hero-after.mp4](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e72bd32d4509f1fe/reproduction-I01-iphone-hero-after.mp4)
[reproduction-evidence.zip](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b73dfb91a2fcfc9b/reproduction-evidence.zip)
[reproduce.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/ee943fbb8cdeadb8/reproduce.py)
[reproduce.mjs](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/3c19d56d533c4958/reproduce.mjs)

All inspected temporal contact sheets and sampled timestamps, grouped within upload size limits:

[sheets-core-before.zip](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/f246d951b16a2ad6/sheets-core-before.zip)
[sheets-core-after.zip](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/9d8b26a61c0b51e7/sheets-core-after.zip)
[sheets-supplement-before.zip](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/ea1b6a33f1ae7d78/sheets-supplement-before.zip)
[sheets-supplement-after.zip](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/32461a2a5a876113/sheets-supplement-after.zip)


Uploads used the [ghdrop skill](/home/ubuntu/.codex/skills/ghdrop/SKILL.md), `ghdrop upload -m`; no forge comments or body edits. Final local verdict: `/tmp/pr24577-audit2/verdict.md`; row details: `/tmp/pr24577-audit2/row-verdicts.md`; hash inspection: `/tmp/pr24577-audit2/final-inspection.json`; independent result: `/tmp/pr24577-audit2/reproduction.json`; cleanup: `/tmp/pr24577-audit2/own-cleanup-verification.json`.