# Fresh current-head proof audit r6

2026-10-22 UTC, read-only follow-up for [#23462](https://git.cafe/pingdotgg/t3code/pulls/23462). **VERDICT: INSUFFICIENT for the complete requested matrix.** This pass closes the missing current-head independent capture/comparison, with the exact geometry/scroll limits below. It does not close interaction rows or establish a new production defect. No source fix is requested.

## Revision and provenance

Independently verified clean detached checkout `/home/ubuntu/wt/pr-23462` at [49e0abba89](https://git.cafe/pingdotgg/t3code/commit/49e0abba8922c4edb7b210b981f880471c8ae8cc), initially and finally. Merge-base with local origin/main remains [3b95e7831b](https://git.cafe/pingdotgg/t3code/commit/3b95e7831b11b17627dff16843bc087bfca397ec). Read the actual complete delta from prior capture revision [366ec74cf8](https://git.cafe/pingdotgg/t3code/commit/366ec74cf818a8b0ccd032d115e2f6d353f61245): exactly seven additions in the Python fixture, four docstring lines and three guard lines. The database-adjacent disposable marker must be a file with the expected content before sqlite3.connect executes. No seed execution or marker creation occurred here. Production render code is byte-identical between those revisions; `git diff --exit-code` over apps/packages passes, and the only whole-tree delta is the fixture.

The primary owns the current-head server/restart at localhost:7456. This auditor reused it without process or database actions. Checkout identity is verified; screenshot provenance also relies on that owner-supplied server binding, rather than a cryptographic SHA in the pixels. BEFORE remains the stopped base above. No old artifacts are relabelled as current head: original valid Wide core, twelve new canonical recordings, twenty-three new pairs and r5 independent image remain captured at 366ec. Their earlier arithmetic/interaction findings remain attributed to proof2/r5. Under the requested strict revision policy, unchanged production code does not turn these into 49e interaction recordings.

Read current PR envelope and proof2 first, then matrix-r4, evidence-r3, proof-r5, Wide-reference-rebuild and independent-setup. The PR body already discloses fixture-only follow-up and superseded capture revision. No body/comment/check/watch operation was performed. Required app registration was confirmed linked and not watched.

## Fresh reconstruction

Used [test-t3-app](/home/ubuntu/t3code/.agents/skills/test-t3-app/SKILL.md) and [ghdrop](/home/ubuntu/.codex/skills/ghdrop/SKILL.md), T3 preview tools only. preview_status initially reported unavailable; preview_open(reuseExistingTab=false, profileId=default) created a fresh owned tab, and existing cookie authenticated without pairing. All subsequent browser calls targeted that tab.

Set CSS viewport1440×900, light appearance, real Wide preference through import of useSettings/updateClientSettings. Clicked the existing Load earlier turns button four times, checking existence after each; absent at completion. Final DOM has80 data-minimap-strip nodes. Native rail click at280,480 mounted paragraph41. Waited for18 consecutive requestAnimationFrame samples with scrollTop variation<0.05, capped300 frames with failure if unsettled. Setup normalized the paragraph toward y240 through the real scroll container, then native hover280,480 and another18 stable frames; fonts loaded. No arbitrary timeout, CSS geometry/content changes, synthetic input, app-ref manipulation or selection creation.

Measured DPR2, fine=true/coarse=false, root16px, fonts loaded, sidebar256, details closed, no toast, empty selection, BODY focus, Chrome154/Linux, reduced motion=false. Timeline viewport1184/content1088 yields g48/h36. Rail x268/y170/w36/h612. Preview41 x300/y424.8671875/w320/h110 and pointer-events:none. Composer GPT-6-Astra/Medium. Whole-surface original preview_snapshot(save=true) PNG is1280×800; copied byte-for-byte, no secondary scaling/crop.

**Exact mismatch retained:** paragraph x308/y240.0703125/w1080/h68.25, versus supplied target y240.1953125: current is0.125 CSS px higher. Absolute scrollTop30139 versus delivered reference33285, delta−3146; equal to r5's30139. This is a qualified visible-state reconstruction, not exact target geometry/virtual-cache identity. Source-estimated unmeasured row heights are a plausible cause of absolute-scroll history differences, not a proved explanation. No forced scroll33285 or layout shift conceals them. The downscaled PNG matches r5 outside the age label despite the reported DOM y difference; output pixels do not prove identical subpixel DOM geometry.

![Current-head AFTER Wide light, paragraph41 preview and loaded80](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a1b50f73db5584a3/current-head-after-Wide.png)

## Exact pixel comparison

Both reference originals inspected visually, along with current image and diff outputs. Compare RGB int16 absolute channel differences; saturating×8 output includes every nonzero channel. Strict mask is magenta on black iff max channel delta>16. No registration, tolerance, exclusions, replacement, masking, cropping or rescaling. All1,024,000 pixels counted. Threshold16 is sensitivity, not an antialiasing classifier.

| Reference, both at366ec | Changed>0 | Changed>16 | Maximum | Changed bbox, half-open |
|---|---:|---:|---:|---|
| Delivered Wide reference |50,129|20,785|205|[187,63,1238,800)|
| r5 independent Wide |91|61|129|[187,102,200,111)|

Current versus r5 differs only within the sidebar age text,26m→37m. All pixels outside that bbox are exactly equal; this claim uses the complete difference array, not an excluded-region comparison. Current versus delivered includes sidebar7m→37m plus user bubble text/edges, reference icon, Result/list/context lines, preview/glass edges and composer/shadow/bottom content. These are not established as AA-only. Same visible content/wrapping does not imply identical rasterization, and cause of the broader reference differences is not resolved here.

| Exhaustive PNG y-band | Delivered changed0/>16 | r5 changed0/>16 |
|---|---:|---:|
|0..147|5888 /3634|91 /61|
|147..310|119 /54|0 /0|
|310..340|1743 /1391|0 /0|
|340..480|9271 /4044|0 /0|
|480..600|14234 /11662|0 /0|
|600..800|18874 /0|0 /0|

Bands partition the full surface. [Manifest, measurements, hashes and counts](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/337240328d85b4f7/current-head-manifest.json).

![Current versus delivered366ec absolute RGB difference ×8](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/24e08c5cb1a4e46b/current-vs-delivered366ec-diff0-x8.png)

![Current versus delivered366ec strict max-channel delta greater than16](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/805363a6f02dd711/current-vs-delivered366ec-diff16.png)

![Current versus r5 absolute RGB difference ×8](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/6ff3f003b5bf0b6a/current-vs-independent-r5-diff0-x8.png)

![Current versus r5 strict max-channel delta greater than16](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/894b6aad37f4dee6/current-vs-independent-r5-diff16.png)

## Remaining matrix, preserved from r5

| Row | Precise verdict |
|---|---|
| Original literal Wide light/dark | Valid core selection/link evidence preserved at366ec, superseded revision disclosed. This still does not retest those interactions at49e. |
| D light/dark | Settled first/last disabled and opposite-neighbor clicks supported by prior recordings; successful message-owned cross-rail selection remains absent (empty ranges), outside-window held release/re-entry/recovery unproved. |
| N20/N8 | Labelled normalized real-app geometry supplements retained; observed neighbor behavior and fade-to-zero supported previously. Cross-selection/recovery and tightly measured fade timing remain unproved. Sample intervals175.9/214.3ms do not establish exact150ms. |
| Z h0 | Actual persisted Full acquisition measured h36, not requested h0. Valid Full h36 selection/link/wheel additions remain prior-head evidence. Old CSS h0 supplement is not actual-Full h0 proof. |
| C2 | Rail height8, hit width40, blank fallback/no first assistant and short final assistant retained. Text y237..260 does not overlap preview410..480; both cross-drags empty. Overlapped-text drag remains unproved. BEFORE toast/sidebar variation retained. |
| C0/C1 and M fine | Prior absence/narrow fine-pointer controls retained with their stated scope. |
| M coarse | Missing actual fine=false/coarse=true and computed display:none, both revisions/themes. No coarse emulation or native mobile verification performed. |
| Independent current-head capture | Delivered with comparisons; supersedes current-head recapture gap only. Exact y/absolute-scroll mismatch and nonzero reference pixels prevent exact-identity claim. |

Instrumentation caveats persist: all six canonical D/C2 Space calls used key=" " and failed; do not count Space success. Passive `button[aria-label^="Jump"] > span` preview flag is non-specific and can match strip spans; cannot prove presence/removal/timing. Rapid keyboard actions are not individually settled assertions. DOM focus setup is not native Tab proof. Native drag success alone is not selection or no-jump proof; input automation may scroll targets into view. No fresh exhaustive frame review or23-pair re-audit claimed; r5 arithmetic validation remains attributed to r5, older67-pair validation to proof2.

## FRICTION and handoff

| Friction | Outcome |
|---|---|
| Preview initially unavailable | Fresh preview_open succeeded; default cookie worked. No alternate browser or pairing action. |
| Bounded document output truncation | Required proof-r5 and evidence-r3 reread separately; current body revision disclaimer read. Old body artifact table not re-audited exhaustively. |
| Initial loaded-count selector counted container button | Corrected to data-minimap-strip, final80; load-button existence checked each step. |
| Current y differs0.125px, absolute scroll differs3146 from delivered | Disclosed, not forced away; exact target identity remains qualified. |
| Live age label | All91/61 current-vs-r5 changed pixels retained; no AA-only claim. |
| Strict current revision policy | Earlier media stays366ec despite identical production code; still does not close current-head interaction matrix. |
| Scope and ownership | No code/source edits, branches, comments/body, watch/CI polling, seeds, direct database writes, server starts/stops or process actions. Only /tmp evidence/report writes and authorized public uploads. Clean exact head confirmed. |

**No more browser verification needed for this audit.** Primary may perform cleanup through its owned sessions; auditor has not stopped anything. Web-only pass; desktop/mobile/provider paths were not independently exercised. No request for source fixes. Final complete-matrix verdict remains **INSUFFICIENT: D cross-selection/outside, N20/N8 cross-selection/recovery/timing, Z h0, C2 overlap, M coarse**, plus strict prior-head interaction provenance limitation. GPT-6.1 Sol / Codex harness in T3 Code.