# Independent proof audit PASS2 / r5

Read-only web evidence audit of [#23462](https://git.cafe/pingdotgg/t3code/pulls/23462), 2026-10-22 UTC. **VERDICT: INSUFFICIENT for the complete requested proof matrix.** Valid original literal Wide light/dark selection/link proof is retained. This is an evidence verdict, not a request for source changes or a newly established implementation defect.

## Findings first

1. **D, N20, N8, Z, C2 and M-coarse remain unclosed.** Settled D endpoint click coverage is now supported on both revisions and themes. Cross-rail native drag calls still produce empty selected text; they do not prove successful message-owned selection across the rail. Outside-window held release/re-entry remains unavailable. N20/N8 now have normalized anchor/geometry supplements and observed neighbor behavior, but not successful cross-text selection or tight fade timing. Actual Full was acquired but measures h36, so it cannot close the requested h0 control. C2 has no text/preview overlap in this fixture/layout. Coarse-pointer hiding was not exercised. These last two are acquisition/classification limits; no h0 or overlapped-text passing assertion follows.
2. **Independent reconstruction and pixel comparison completed, with a disclosed absolute-scroll mismatch.** I opened a fresh default-profile tab with reuseExistingTab=false, used tabId on every call, authenticated with the existing cookie, set the real Wide preference through dynamic import/updateClientSettings, loaded10→30→50→70→80 turns while the button existed, navigated natively at280,480, normalized paragraph41, hovered and waited18 stable animation frames. CSS1440×900, DPR2, light, root16px, fonts loaded, fine=true/coarse=false, reduced motion=false, sidebar256, details closed, no toast or geometry override. Paragraph x308/y240.1953125/w1080/h68.25 and g48/h36 match the reference. Preview41 is visibly open and pointer-inert. **scrollTop30139 differs from reference33285.** LegendList's source estimates unmeasured rows; different measured-height history is plausible, but exact cache/scroll equivalence was not proved. I do not label this an exact-scroll reproduction or pixel-invariant match.
3. **The independent comparison is not AA-only.** Final canonical whole-surface1280×800 PNG has **50,173 changed pixels at zero threshold and20,824 at strict maximum-RGB delta>16, of1,024,000**; maximum delta205. Changed bounding box is[187,63,1238,800), with no alignment, masking, crop or resizing. Sidebar age7m→26m is a real content difference. Other changes affect user-bubble text, link icon, Result/list/context text, preview/glass/shadow and composer regions. Neither their exact cause nor an AA-only explanation is established. Every changed pixel is preserved in the two full-surface outputs below; region counts partition the entire image. The visible paragraph/preview/text wrapping match, but this is not a zero-unrelated-change guarantee.
4. **All23 new pair arithmetic checks pass, including the strict threshold masks.** I independently recomputed both the saturating absolute-RGB×8 images and magenta-on-black strict max-channel>16 masks for every pair, compared the output pixels exactly, and checked both changed counts and totals. Zero failures. This validates arithmetic, not synchronization or behavioral sufficiency. ActualFull-released AFTER is visibly stale/empty in its snapshot while encoded frames retain content. C2 BEFORE has an unmatched provider-update toast and extra sidebar entry/status differences. These remain disclosed; none is dismissed as AA.
5. **Two capture-instrumentation limitations need explicit qualification.** All six canonical D/C2 Space calls used key=" " and returned isError=true; native Space success is not demonstrated by these new recordings. Other key calls succeed, but rapid D arrow/Home/End sequences are not individually settled behavioral assertions. Also the passive selector `button[aria-label^="Jump"] > span` matches ordinary strip spans before the preview: its boolean preview flag is true even without a preview, and state preview text is empty. That field cannot establish preview presence/removal or exact preview timing. Opacity samples, strip widths, selected text/owners, native click events and visible encoded frames remain independently usable. No source change is requested to repair this audit instrumentation.
6. **Full h36, C2 non-overlap, and passive drag outcomes are supported, with attribution.** Capture routine Z explicitly calls the actual updateClientSettings({chatWidth:'full'}) path and removes geometry CSS; measured viewport1184/content1088 gives g48 and h36 on both. The source's normal desktop timeline padding and stable gutter explain why actual Full need not yield h0. Captured Full paragraph→link drags produce the same nonempty message41-owned range on both revisions, trusted synthetic reference-link hits, and wheel+120; these are valid additions at h36. C2's state records show two strips, height8 and width40 after/352 while active before; whole-surface last-preview images and encoded frames place short assistant text well above the preview. Reported paragraph y237..259.75 versus preview410..480 does not overlap, consistent with the images and assistant row235..307.75. C2 cross-rail ranges are empty. No manufactured text, synthetic selection/events or CSS hiding closes these gaps.
7. **Revision/source checks pass and source stays clean.** Independently checked AFTER [366ec74cf8](https://git.cafe/pingdotgg/t3code/commit/366ec74cf818a8b0ccd032d115e2f6d353f61245), merge-base [3b95e7831b](https://git.cafe/pingdotgg/t3code/commit/3b95e7831b11b17627dff16843bc087bfca397ec), and nested ignored BEFORE checkout at that base; both git statuses clean initially and finally. Full four-file diff and committed Python/JSON seed read; seed not executed. Implementation fixes hit-strip width, makes preview pointer-inert and removes obsolete expanded-width/preview exemptions. No new implementation defect is established. No repository edits, branches, pushes, comments, body changes, watch, CI polling, server actions, seeding or database writes. Required app PR registration was the only PR metadata operation; confirmed linked and not watched.

## Row table

Desktop new recordings are1440×900 CSS, DPR2, fine pointer, loaded80 long turns or exactly2 for C2. All timestamps below are action-call offsets from recording acknowledgement, not exact compositor deadlines. Complete per-action arguments/times and errors are in [recordings.json](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/aba254f3dafc85fa/recordings.json). I visually inspected sampled encoded contact frames from all12 canonical recordings and the relevant full-size stills; no exhaustive frame-by-frame claim.

| Row | Matched recordings and supported additions | Exact remaining requirement / status |
|---|---|---|
| Original literal Wide | Original WPD light/dark selection/link core remains as recorded in the preserved body and earlier proof. New still below is supplemental. | **Retained; not replaced.** No new claim of complete temporal/pixel invariance. |
| D light | [Before](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/9d88e1c7f9be2925/before-light-D.mp4) / [After](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/9f65b2ff9682a948/after-light-D.mp4). First/disabled/next at16.750/19.558/20.399s BEFORE,11.471/14.449/15.711s AFTER. Last/disabled/previous21.963/24.749/25.814s BEFORE,17.441/20.271/21.336s AFTER. States confirm disabled scroll unchanged,1→2 and80→79; middle41→40→41. | **Incomplete:** successful message-owned cross-rail drag and outside-window release/re-entry/recovery. Empty native cross-drag ranges at0.372/1.958s BEFORE,0.384/2.291s AFTER. Space failed; do not count it. |
| D dark | [Before](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/30e7c9abb0c2c09f/before-dark-D.mp4) / [After](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/490d28d7d45ec8f4/after-dark-D.mp4). First/disabled/next4.005/6.682/7.694s BEFORE,3.267/5.954/6.881s AFTER; last/disabled/previous9.244/12.011/12.847s BEFORE,8.537/11.189/12.207s AFTER. Encoded samples show paragraph1/2/80/79 destinations. | **Incomplete:** same cross-selection/outside-release gaps. Drag0.321/2.006s BEFORE,0.297/1.736s AFTER yields empty range; ordinary inside-page endpoint scroll equality cannot establish outside recovery. Space failed. |
| N20 | [Before](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e28be1ebe3934631/before-light-N20.mp4) / [After](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/6aefcd39b6d69c7e/after-light-N20.mp4). Labelled CSS real-app supplement g32/h20, nonpersistent, same mounted anchor. Hover/footprint0.306/1.925s BEFORE,0.336/1.910s AFTER. Neighbor5.489/6.354s moves41→42; rail7.435/8.254s returns41. BEFORE active352, AFTER20; AFTER opacity reaches0. | **Incomplete:** successful cross-text selection/held recovery; exact fade interval unresolved. AFTER samples1→0 separated175.9ms, too sparse for exact150ms. CSS root overflows physical parent; geometry supplement, not ordinary Wide layout. No independent N link-hit sequence. |
| N8 | [Before](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/3eec58fa8ee39623/before-light-N8.mp4) / [Corrected After](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/45ff3aa0e1dccd1d/after-light-N8-corrected.mp4). g20/h8; neighbor pointer-events none. Hover/footprint0.298/1.836s BEFORE,0.321/1.937s AFTER; neighbor5.424/6.405s leaves scroll unchanged; rail6.673/7.667s remains usable. | **Incomplete:** same cross-selection/recovery/timing gaps. AFTER sampled1→0 spans214.3ms with sparse frames. Initial AFTER h20 attempt explicitly superseded; never N8 proof. |
| Z / actual Full | [Corrected Before](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/4702a6e305d40024/before-light-actualFull-corrected.mp4) / [After](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/122c9b35db0c177b/after-light-actualFull.mp4). Native message-owned range0.372/0.366s; double-click2.343/1.897s; trusted link3.150/2.744s; wheel3.457/3.056s BEFORE/AFTER. h36 throughout idle; keyboard preview and pointer/blur supplemental. | **h0 control missing.** Actual Full is a valid h36 addition, not zero-gutter/pointer-inert proof. Old CSS h0 supplement stays separately labelled. First misclick attempt historical. ActualFull-released still is stale; use recording/logs. No source fix requested. |
| C2 | [Before](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/707b469db0801ac7/before-light-C2.mp4) / [Corrected After](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e21a7ea2ffb067c2/after-light-C2-corrected.mp4). Blank fallback first/no assistant, short assistant last; first/last hover0.302/2.089s BEFORE,0.259/2.041s AFTER;8px rail height, h40. | **Overlap text-drag unproved/unreachable in this layout:** text237..260 vs last preview410..480. Both drag directions empty at3.221/5.168s BEFORE,3.264/4.809s AFTER. Space failed. BEFORE toast/status/sidebar differences remain. |
| M-coarse | No matched coarse acquisition. Existing390×844 fine-pointer controls retained. | **Missing:** actual fine=false/coarse=true browser and computed display:none, both revisions/light+dark. Resize/fake matchMedia/CSS hiding would not satisfy it. Web-only scope; no native mobile/desktop required. |
| Independent recapture | Fresh own AFTER Wide/light whole-surface screenshot, exact two diff outputs, setup and manifest below. Same visible paragraph41/preview/geometry/environment. | **Acquired and compared, qualified match:** absolute scroll30139 vs33285,50173/20824 changed pixels. Exact-scroll/pixel identity not established. The old authentication-blocked recapture is superseded; no zero-diff/AA-only claim. |

Passive logs are cumulative per page; I isolated relevant state-time windows before interpreting samples. `isTrusted` on scroll/focus alone is not native-input provenance because setup scroll/focus also produces trusted events. Native action steps establish initiating input. Full paragraph/link range owners both identify synthetic-minimap-long-run-041-assistant; D/N/C2 end-state selections are empty. Focus setup uses DOM focus(), so these are not Tab-navigation proofs. No independently held down/outside/up primitive exists in the authorized catalog.

## Independent images and exhaustive difference accounting

Delivered AFTER reference (preserved):

![Delivered AFTER Wide reference](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/cfcc8e765fe9ba0e/after-light-Wide-reference.png)

Fresh independently rebuilt AFTER:

![Independent AFTER Wide light paragraph41 preview](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/71ea409033631c16/independent-after-Wide.png)

Absolute per-channel RGB difference, saturating×8; every nonzero channel retained:

![Independent versus delivered absolute RGB difference amplified eight times](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/6aa6d16e051159ef/independent-diff0-x8.png)

Strict max-channel>16 magenta mask on black:

![Independent versus delivered strict maximum RGB delta greater than sixteen](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b1ad829c677a89e2/independent-diff16.png)

| Exhaustive disjoint PNG y-band | Changed0 | Changed>16 | Observed affected region |
|---|---:|---:|---|
|0..147|5932|3673|User bubble text/edges plus sidebar age; age alone80/65. |
|147..310|119|54|Reference-link icon. |
|310..340|1743|1391|Result text. |
|340..480|9271|4044|List/context text and preview/glass edges. |
|480..600|14234|11662|Long context text on several lines. |
|600..800|18874|0|Composer glass/shadows, edges and bottom content. |
|Total|50173|20824|1024000 pixels, no exclusions. |

Bands are half-open and partition the whole PNG. Regions and full masks report all differences without deciding their cause. Several unchanged text lines sit alongside changed lines; that does not establish a uniform alignment error or AA-only origin. Snapshot compositor/raster variability is possible but unproven here. Threshold16 is a sensitivity cutoff. Original tool PNG1280×800 is downscaled from CSS1440×900/DPR2; neither source was rescaled by this audit.

[Independent exact setup](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/c13cbd9134e064b1/independent-setup.md) · [Acquisition/environment manifest](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/938fcd7883e73ee8/independent-manifest.json) · [Exhaustive region counts](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e4f2e9a42576cc71/independent-regions.json) · [23-pair independent arithmetic result](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/ab0b7b00b222b524/pair-arithmetic.json).

## Source, package and audit limits

Read FIRST current PR JSON and proof2, then source-confirmed matrix-r4 and capture progress. Rebuilt the reference before extensive package audit. Subsequently read final [capture index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/8a780a64c05a2919/issue16447-evidence-r3.md), raw59 shots/140 states/285 steps, capture expressions, passive logs,23 pairs and15 recording entries (12 canonical,3 historical). All canonical contact sheets inspected visually; relevant full-size C2/reference/diff images inspected. New arithmetic independently checked, older67-pair verification remains attributed to proof2. No fresh test/lint/typecheck execution was needed for read-only evidence work. No exhaustive public-URL or archive-content hash audit claimed; uploaded independent artifacts are the CLI-returned public locations.

[Source-confirmed matrix](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/ca3412dc1a3836ce/issue16447-matrix-r4.md) · [Capture bundle](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/dba355770d6ec467/issue16447-r3-evidence.tar.gz) · [Raw manifest](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/169402d97e607219/capture-manifest-raw.json) · [Exact capture routines](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/65475bcbef6a62e4/capture-routines.json).

Committed [seed](https://git.cafe/pingdotgg/t3code/src/commit/366ec74cf818a8b0ccd032d115e2f6d353f61245/scripts/fixtures/issue-16447.py) and [fixture](https://git.cafe/pingdotgg/t3code/src/commit/366ec74cf818a8b0ccd032d115e2f6d353f61245/scripts/fixtures/issue-16447.json) supply reproducible synthetic80/2/1/0-user projections, not event/backend proof. No seed execution here. [TimelineMinimap.tsx](https://git.cafe/pingdotgg/t3code/src/commit/366ec74cf818a8b0ccd032d115e2f6d353f61245/apps/web/src/components/chat/TimelineMinimap.tsx) confirms fixed rail, display-only preview, fine-only CSS and button-free release logic; [MessagesTimeline.logic.ts](https://git.cafe/pingdotgg/t3code/src/commit/366ec74cf818a8b0ccd032d115e2f6d353f61245/apps/web/src/components/chat/MessagesTimeline.logic.ts) confirms g/h, persistence and neighbor thresholds. This audit does not infer a new bug from unexercised boundaries.

## FRICTION

| Friction | Outcome |
|---|---|
| Preview initially unavailable | preview_open fresh default tab succeeded; existing cookie worked. No pairing CLI, token or alternate browser needed. |
| Independent absolute virtual scroll differs | Visible paragraph rect matches exactly; difference retained and qualified, not forced away. |
| Provisional image50165/20818 vs final50173/20824 | Second canonical snapshot explicitly followed18 stable frames; eight extra zero/six threshold pixels are live age-label change. Final outputs use final image only. |
| Generic nav measurement1048 | Selector found another nav; corrected Threads aria-label measurement256 stored separately. No layout change. |
| Initial batched output truncation and unmatched optional rg path | Relevant required documents reread; focused source reads completed. No inference from omitted output. |
| Capture report arrived during audit | Raw package inspected first, then final35KB report inspected; no waiting loop or duplicate capture work. |
| Space errors / preview flag selector | Failed keys and non-specific preview flag disclosed; not promoted to behavioral proof. |
| Toasts/stale snapshots/coarse/outside limitations | Preserved and qualified; no concealment, synthetic proof or demanded source change. |
| Server ownership | No server action. No more browser verification needed; primary may stop servers through its owned sessions when appropriate. |

VERDICT: **INSUFFICIENT — D cross-selection/outside release, N20/N8 cross-selection/recovery/timing, Z h0, C2 overlapped-text drag, M-coarse.** Independent capture/comparison now delivered with the exact-scroll and nonzero-pixel limits above. Original Wide core remains valid and preserved. Source remains clean.