# Final-head proof audit, pass 2

**SUFFICIENT for the approved settled-rendering/routing matrix.** No required row is missing or invalid. This is an evidence verdict, not a merge authorization or a claim about deliberately unverified interactions.

Replacement [#24577](https://github.com/pingdotgg/t3code/pull/24577), branch `takeover/pr-18816`, replaces [#18816](https://github.com/pingdotgg/t3code/pull/18816). Original contributor [saphid/t3code](https://github.com/saphid/t3code), branch `t3code/sweep/marketing-mobile-downloads`, revision [707797ab](https://github.com/saphid/t3code/commit/707797ab66291552c8aeca4c8eea7932e21c6ff0), is provenance only.

Before: [c4fa4629](https://github.com/pingdotgg/t3code/commit/c4fa4629eb645b95aae2beef56c70f705b736049). After: [69b0eb2c](https://github.com/pingdotgg/t3code/commit/69b0eb2c4c9b20210a40c1fad35977ba056aa946). The latter is a two-parent merge of reviewed [8e2e67e3](https://github.com/pingdotgg/t3code/commit/8e2e67e31da19a11255a7eaf49702176489f52de) and the supplied baseline. Its marketing tree is identical to the reviewed revision, but this audit uses only the fresh final-head capture set and a new final-head recapture; historical screenshots/audit are not substituted.

[Delivered evidence index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a514c823e1485117/evidence-index.md) contains each full-size Before/After/Diff0/DiffT/sections image and both DOM/network logs. I independently opened **all 80 delivered PNGs** through the image viewer: 32 full-page images, 32 diffs, 16 section sheets. I also independently downloaded and SHA256-compared **all 165 public files** against local delivery, validated **54 transcripts**, recomputed every pixel of both diff variants, and verified **208 source-file hashes** against git objects.

The complete source diff is four marketing files (88 additions / 14 deletions): homepage routing/CSS, platform helper, helper tests, and package test task. Only the homepage has a changed render path. The planning matrix's historical three-file/count wording omits the nonvisual package task; this does not omit a rendering fixture. The final index correctly describes the published replacement. No current “unpublished head” objection remains.

## Per-row proof

Every static row below has the exact two SHAs above, both CTA hrefs checked independently, full page plus readable hero/git/lower/footer sections, both correctly generated diffs, and matching original-viewport geometry guards. Preferences are dark except L01/L02 (light); actual theme is dark in every row. Viewport is the load/DOM viewport; PNG height is full-document CSS raster height, not device-pixel height. Delta counts are changed pixels at >0 / >16 maximum absolute RGB-channel delta.

| Row | Profile; load viewport; effective DPR | PNG dimensions | Observed before → after, both CTAs | Diff0 / DiffT pixels | Result |
|---|---|---|---|---|---|
| M01 | iphone; 390×844; 3 | 390×4894 | macOS/Apple/DMG → iOS/Apple/App Store; command hint retained | 6100 / 5860 | SUFFICIENT |
| M02 | iphone; 1440×900; 3 | 1440×4553 | macOS/Apple/DMG → iOS/Apple/App Store; wide lower group re-centers | 7513 / 6980 | SUFFICIENT |
| M03 | android; 390×844; 3 | 390×4893 | Linux/Linux icon/AppImage → Android/no OS icon/Play; Ctrl retained | 3406 / 3224 | SUFFICIENT |
| M04 | android; 1440×900; 3 | 1440×4552 | Linux/Linux icon/AppImage → Android/no OS icon/Play; wide lower group re-centers | 4580 / 4106 | SUFFICIENT |
| M05 | ipad-desktop; 390×844; 2 | 390×4894 | Mac UA + touch 5: macOS/DMG → iOS/App Store; Apple/command retained | 6100 / 5860 | SUFFICIENT |
| M06 | ipad-desktop; 1440×900; 2 | 1440×4553 | Same iPad desktop branch, wide; 3 residual pixels accounted below | 7516 / 6980 | SUFFICIENT |
| C01 | mac; 1440×900; 1 | 1440×4552 | macOS/Apple/DMG unchanged; 3 residual pixels | 3 / 0 | SUFFICIENT |
| C02 | mac; 390×844; 1 | 390×4893 | macOS/Apple/DMG unchanged; narrow control | 0 / 0 | SUFFICIENT |
| C03 | windows; 1440×900; 1 | 1440×4552 | Windows/Windows icon/EXE/Ctrl unchanged; 3 residual pixels | 3 / 0 | SUFFICIENT |
| C04 | linux; 1440×900; 1 | 1440×4552 | Linux/Linux icon/AppImage/Ctrl unchanged | 0 / 0 | SUFFICIENT |
| U01 | unknown; 390×844; 1 | 390×4893 | No platform attribute; macOS/Apple/Releases/command unchanged; no release request | 0 / 0 | SUFFICIENT |
| U02 | unknown; 1440×900; 1 | 1440×4552 | Same unknown fallback, wide | 0 / 0 | SUFFICIENT |
| L01 | iphone; 390×844; 3 | 390×4894 | Light preference, fixed dark; identical to M01 on each side | 6100 / 5860 | SUFFICIENT |
| L02 | android; 1440×900; 3 | 1440×4552 | Light preference, fixed dark; identical to M04 on each side | 4580 / 4106 | SUFFICIENT |
| F01 | windows; 390×844; 1 | 390×4893 | Missing assets: Windows/icon/Ctrl; Releases fallback retained | 0 / 0 | SUFFICIENT |
| F02 | linux; 1440×900; 1 | 1440×4552 | Rejected fetch: Linux/icon/Ctrl; Releases fallback retained; 3 residual pixels | 3 / 0 | SUFFICIENT |

| Collapsed row | Independently checked real-page transcripts at both SHAs | Result |
|---|---|---|
| B01 | Explicit iPad and iPod UAs, touch 5: mac/DMG → ios/App Store; release requests 1 → 0; Apple and command text retained | SUFFICIENT |
| B02 | Mac touch 0 and touch 1 remain mac/DMG; release requests 1 on both sides | SUFFICIENT |
| B03 | Narrow Windows/Linux success selects EXE/AppImage respectively, with corresponding icon/label/Ctrl | SUFFICIENT |
| B04 | iPhone/Android rejected endpoint: before requests and falls back to Releases; after selects store without requesting endpoint | SUFFICIENT |
| B05 | macOS, Windows, Linux session reload: correct two hrefs retained and zero additional release requests on both SHAs | SUFFICIENT |

The five B groups represent 11 distinct inputs and 22 transcripts; with 32 static transcripts this accounts for all 54. No helper-unit-test output was treated as browser evidence.

## Environment, provenance, geometry and network

The driver runs real Astro homepage servers from the supplied commits with installed repository dependencies. Its only network interception is the exact latest-release endpoint, using fixed release JSON, empty assets or an intentional abort. No CSS/source edits, navigator injection, synthetic component harness or app-data access occur. I read the actual source call path and constants rather than accepting the driver's oracle as proof. Every transcript's document request belongs to its declared server origin; the only failed network requests are the four expected release aborts (F02 both sides and B04 before). No store or installer request was activated.

Both hero and lower hrefs independently equal App Store/Play on mobile after, arm64 DMG/x64 EXE/x86_64 AppImage fixtures on desktop success, and Releases on unknown/missing/rejected paths. After mobile and unknown issue no release requests; normal desktop paths issue one, with cache reload adding none. Footer and nav Download remain `/download`; actual footer iOS/Android links remain the explicit stores. The lower alternate is visually unchanged except its expected horizontal displacement when its wide group re-centers.

Headed Chrome for Testing 145.0.7632.6, Playwright 1.58.2, Linux x86_64, Xvfb, software compositing, hidden scrollbars, reduced motion. Effective DPR is 3 for phone profiles, 2 for desktop-iPad, 1 for desktop/unknown. All exact UAs, effective touch points, UA-CH, coarse pointer, media preference, scale 1, original inner viewport and no horizontal overflow agree with fixtures. DM Sans and JetBrains Mono are loaded; both labels compute to 15px and the hint to 10px. Default font scaling/zoom are retained with natural Chromium autosizing. Safari UA profiles are **Chromium emulation**, not native Safari.

Each full-document frame is a CDP scale=1 CSS raster. Loading uses 390×844 or 1440×900, then height expands once to scrollHeight without a width change. The source driver asserts exact original-versus-expanded hero, lower and hint rectangles; the independent recapture executes the same guard. Each delivered PNG hash matches both recorded byte-identical frames, separated by 500ms. Expanded viewport dimensions and effective DPR agree with logs and PNG sizes. This is not a claim that every unrecorded element was separately DOM-measured; full-page zero-threshold diffs cover the rendered surface.

## Every visible change / residuals

Both iOS labels become “Download for iOS”; Apple remains. Both iOS buttons narrow from **217.1875 to 189.703125 CSS px**. Both Android labels become “Download for Android”; Linux icons disappear without replacement; widths narrow from **203.625 to 200.796875 CSS px**. Hero buttons remain centered (within 0.02 CSS px in the recorded coordinates). Narrow lower buttons remain centered; wide lower action groups re-center, moving the adjacent alternate horizontally. No vertical displacement of surrounding content is visible between a matched pair. Text, decoration, endorsements, feature sections, git hint pixels and actual footer match apart from the explicitly counted changes.

Command/return characters remain the same DOM strings on both sides, but their Linux raster glyphs look boxlike. I do **not** claim those glyphs are legible or that this PR fixes their appearance. They are unchanged pixels, not a new regression.

I independently regenerated all Diff0 and DiffT arrays without scaling/alignment and obtained exact equality with the delivered diff PNGs. Outside the two CTA action regions (including the adjacent wide alternate), only **M06, C01, C03 and F02** have residual changes: exactly three pixels per pair at **(1301,724), (1302,725), (1302,727)**, max channel delta **1/255**, none >16/255. These lie at the transformed product-image corner. Compositing/raster variability is a plausible inference, not a proven source cause. All other rows have zero outside-region changed pixels. Thus the fresh set has four residual pairs; this precise count supersedes the brief's “six wide frames” shorthand. No unexplained above-threshold change remains.

L01 equals M01 pixel-for-pixel before and after; L02 equals M04 pixel-for-pixel before and after. The light-preference coverage demonstrates a fixed-dark site, not a light design. C02/C04/U01/U02/F01 are wholly pixel-identical controls; C01/C03/F02 have only the three one-level residual pixels above.

## Independent affected recapture

Fresh scratch: `/tmp/pr-18816-final-audit`. I downloaded the public runner, browser driver, processor and fixtures, verified their bytes, then executed those downloaded files with `CAPTURE_ROWS=M05 CAPTURE_REPO=/home/ubuntu/wt/pr-18816 python /tmp/pr-18816-final-audit/run.py`. Existing installed Playwright dependencies were reused through a scratch symlink. The runner asserted exact clean final HEAD, archived the baseline to scratch, and started two real marketing servers plus owned Xvfb. It used isolated Astro preferences and only permitted ignored generated installer/cache files in source. No implementation edits occurred.

| Independent M05 image | Raster | Pixel delta versus delivered | Maximum channel delta | PNG SHA256 (also equal to delivered) |
|---|---|---|---|---|
| before | 390×4894 | **0** | 0 | `9f389886ac7bba8aff19e49a6af988c0f2331185f16284cbdcff997914e0929f` |
| after | 390×4894 | **0** | 0 | `e45bb126a876916c7f1fa24ddcad5e45a3c2c18335d941442af7ada10333d4f4` |

[Independent before](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/dbc5fc01c3f7b8f5/M05-before.png) · [Independent after](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/1996c336c2e41f90/M05-after.png) · [Independent before DOM/network](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/4ada35f369b1d7cf/M05-before.json) · [Independent after DOM/network](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/7c3e520140bea665/M05-after.json)

![Independent M05 sections: before on left, final head after on right](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/ac3e165049e073f0/M05-sections.png)

## Limits, cleanup and audit artifacts

T01/T02 transients, first-frame/no-flash/early-tap behavior; I01 actual pointer/touch/keyboard activation and destination navigation; I02 hover/focus/active/resize/scroll motion; T03 JavaScript disabled; native Safari/WebKit and real hardware remain deliberately **UNVERIFIED**. No additional static row or recording is required for the narrower approved settled-routing claim. Installer URLs are intentionally nonexistent proof fixtures, never download-success evidence.

Both independent servers and Xvfb exited; captured PIDs no longer exist and their loopback origins refuse connections. Browser closed in the driver's finally block. Git status is clean at the exact final head. No T3 server or `~/.t3` access, git history/source edits, push, forge read/write command, comment or watch was performed. The replacement was registered only in T3's thread PR links; no forge state changed. No old-SHA images were relabeled as final-head proof.

[independent-checks.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/fabd618457004432/independent-checks.json) · [independent-public-readback.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a55f5391fe3d9ab7/independent-public-readback.json) · [audit-checks.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/c2330592ac021bef/audit-checks.py) · [verify-public.py](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/75a9105836a3c282/verify-public.py) · [environment.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b69965ec9d9f8d47/environment.json) · [run.log](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/3a333c598f55d777/run.log) · [final-state.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/246f2bdafa9e812a/final-state.json)

FRICTION | capture-environment | unchanged command/return glyphs appear boxlike on Linux | avoid claiming glyph legibility; font changes would require separate matched evidence
FRICTION | rendering-limit | CSS-raster expanded-document captures preserve effective DPR but are not native-DPR hardware screenshots | retain explicit raster and geometry disclosures; do not infer native Safari proof
FRICTION | stale-metadata | historical matrix says three files and calls macOS the longest label although Windows is wider | use final four-file diff and measured widths; no missing render row results
FRICTION | workflow | final merge preserves marketing tree but changes HEAD | recapture exact final SHA, as delivered here; do not carry historical proof forward

Audited by gpt-6.1-sol through the Codex harness using the pr-audit and ghdrop skills.

VERDICT: SUFFICIENT