# Independent evidence proof pass 2

PR [#23519](https://git.cafe/pingdotgg/t3code/pulls/23519), pinned [a9dcdf37fd](https://git.cafe/pingdotgg/t3code/commit/a9dcdf37fd17842bb87a068fa4c94f535cbfc140). Requested scope: evidence only. No product implementation, commit, push, forge comment/body edit, watch or PR polling. Initial metadata matched the requested pin. Initial checkout is product [44215dadd1](https://git.cafe/pingdotgg/t3code/commit/44215dadd198b563f2e2722afb955774ed5dbb9b), with capture-owned HTML/inert-page changes; left untouched while capture active.

**VERDICT INSUFFICIENT.** P6 is a confirmed committed-fixture failure, independently code-reproduced by review A and now witnessed through the real HTTP/WS decoder. New core M6/M13 evidence improves specific chapters; it does not satisfy every chapter in the final matrix.

## Source and data attribution

Read the immutable product diff against [43879d9702](https://git.cafe/pingdotgg/t3code/commit/43879d970204eb28092d098deb4c842a22a9ab75), the fixture sources, previous matrix and independent review A round 3. No new product defect established; this report is evidence acceptance, not a substitute code-review verdict.

Independently hashed all four changed product files (AppSidebarLayout, Sidebar, LegacySidebar, routeReveal): source-attribution.json matches the git objects, and all four are byte-identical between 44215 and a9dc. The entire commit delta between those pins is replay.js (21 insertions/5 deletions). New captures therefore run **product44215 plus external fixture a9dc**, not an entire a9dc build. Bootstrap hash 7d8b1e58dcf495b2cbf0b4c04142e2f3230336ed093b45d857784bdc4324621c; replay hash 2ad001e40b866139485c8d9f508c457f638f7acb8dbcf90b62c3855c795f1adc. Both independently match the immutable a9dc objects.

Extracted the six committed fixture files to this audit's scratch directory and independently invoked seed.py with `python3 -B`, once for plain/legacy and once for all. All eleven SQL files, fixture.json and digest files match capture/plain-seed and capture/seed byte-for-byte. Plain/legacy canonical digest 5253deeb1eb360369afd538920bfb89145f764a843a2c9b43df0f498f3cfdadf; all digest 01dc400cf2d1dc7ad412ffdba6ee87849a963fe3c4abab72c241515120f3cb58. These inputs were then used for the independent real-app rebuild below. Full file hashes: independent-seed-hashes.json.

The synthetic-transcript parity journal shows equal A1(empty) and B8(three messages) projections in both plain homes. The two similar navigation assistant messages come from standalone and completed-run seed records; they are intentional equal inputs. Do not blame that duplication on product behavior.

## Fresh visual and arithmetic inspection

Sampled every final mapped public MP4, read its available action/protocol journals, and inspected all four corrected public M6 clear sheets and all eight corrected public M13 A/B-active sheets, plus sampled original MP4 chapters against timestamped journals. Rejected-stale-sheets remains private and excluded. All inspected dark sheets actually render dark. Configs are WL1440×900/light, WD1440×900/dark, NL390×844/light, ND390×844/dark, DPR2. Frozen epoch1792627200000. Extracted frames retain full native resolution (2880×1800 wide, 780×1688 narrow); contact sheets are resized inspection aids only.

Independently recomputed all twelve finalized full-resolution pairs: every frame SHA/size, raw absolute delta, diff0=clamp(8D), diffT=white iff max RGB D>8, changed counts/percent/bounds matches frame-diff-manifest.json. No masks/crops/tolerance pass. Full numerical results: final-arithmetic.json (the earlier arithmetic-source.json predates the corrected narrow B frames). M6 >8 counts WL73216(1.412346%), WD135490(2.613619%), NL70337(5.342159%), ND138837(10.544796%). M13 A-active WL31106(0.600039%), WD62626(1.208063%), NL34783(2.641800%), ND68196(5.179548%). M13 B-active WL17515(0.337867%), WD888(0.017130%), NL6377(0.484339%), ND10187(0.773712%).

Visible differences: M6 main pane is A1 on both sides. Baseline retains B scope/list; head switches to All, mixed A/B/Spacer list scrolls to A1. Settled footer leaves the visible region, scope glyph changes, active row appears. Wide main pane layout/content stays the same; residual transcript/composer/right-card edges and pointer/focus/compression differ. Narrow open drawer covers main heading; separate closed-drawer A1 chapter supplies actual route evidence. NL has minor upper-header/edge encoding differences; ND includes search caret/focus and dark-gradient compression noise. No claim that every nonzero pixel is the intended fix.

M13 A-active: baseline A8…A2 + Show more becomes A8…A1 once, no A overflow. Replacing the 28px control with a 32px row shifts B header/list/control by 4 CSS px. B remains inactive but retains its completed status; hover/active treatments differ. Main pane remains A1, empty transcript. B-active pairs have identical seven-row previews and both overflow controls; WL has visible encoding noise around links/main pane and composer, not a content mismatch. WD noise is much smaller. Corrected final NL/ND B frames expose actual separator/wrapping differences at the right transcript edge, with 6377/10187 pixels above8. Final visible-change-accounting explicitly leaves that residual unattributed. Equal DB transcript records do not prove identical rendered layout; these pairs do not receive strict pixel-equivalence acceptance.

## Chapter ledger

Retain R1–R12 with original recording/build/fixture attribution and limits from the final-diff-proof-matrix. This pass does not relabel historic images as newly captured final-fixture evidence. No blanket recapture requested.

| Row | Current disposition | Exact remaining gaps |
|---|---|---|
| M1 | Partial; retain R1 history/palette/link chapters | Matching scope, selection clearing, measured fully offscreen controls, ordinary click, same-route manual-scroll/menu/data no-snap across required pairs/configs. |
| M2 | Partial; retain R2 same-route guards/direct load/drawer | Repeated Settings history, matching-scope mount scroll, palette Settings, different final route while unmounted, implementation toggle and measured quiet chapters. |
| M3 | Partial; retain R3 shelves; new P2 shelf fields independently equal | Complete matched shelf configs/indexing and expanded settled target older than initial tail, uniqueness/count/control after reset. |
| M4 | Partial; retain R4 arrival; new late M6 retry supports pending search-clear arrival | Completed-same-A mount-owned absent/release (visible and hidden/manual-B), legacy delayed beyond-preview, archived/orphan/missing resolved route controls, later real data quiet. P6 cannot certify resubscription. |
| M5 | Partial; retain R5 child/fork endpoints | Same-root manual-B grandchild/sibling with actual route/storage, remaining matched baseline/configs. |
| M6 | New matched core B-query→A1 route→clear chapter supported all4; All owed-clear, delayed retry, manual-scroll/retype-clear no-snap are head-side support all4 | Null/new-thread→valid route and transient-null Settings controls remain missing. All/matching controls lack paired baseline recordings; matching-A distinct scope not shown. Unrelated real menu/data rerender quiet not supplied by retyping alone. Final recipe and recording journals attribute compared Back/Forward inputs to hardware Control+[ / Control+]. |
| M7 | Partial; retain R7 ordinary preview6 union; new preview8 head control all4 | Matched preview8 baseline, null removal, measured offscreen palette/link, ordinary click and manual-scroll/menu quiet. |
| M8 | Partial; retain R8 active-only/expanded | Measured wholly-offscreen collapsed A header/row Back/Forward/Back, inactive B collapse and matching narrow baseline. |
| M9 | Partial; retain R9 Settings/drawer | Repeated/palette Settings, different route while unmounted, quiet/manual scroll and complete matching drawer proof. |
| M10 | Partial; retain R10 traversal | Actual hints/jump7, range selection/count, recalculated previous/next boundary and matched other configs. |
| M11 | Partial/unverified; R11 only corroboration | Start B8 with A1 uniquely highest hidden and A2 distinct lower, resolved summary migration/restoration, More/less, header aggregate and matched pairs. Protocol status fields alone do not prove this visual sequence. |
| M12 | Partial/unverified; retain R12 correct ND group endpoint | Matched actual A/B-only group plus separate Spacer, union/collapse/hints/order/scroll sequence. |
| M13 | New preview7 Back/Forward boundary supported all4, with corrected stable A/B endpoint pairs. Preview6+8 head controls all4 corroborate expected boundary | Hint/jump8→A1 and order change on leaving still missing. Narrow closed/open chapter shows B8, not closed A1. Control6/8 lack matched baseline; no complete-set acceptance. |

## Protocol ledger

| Row | Disposition | Proof and limits |
|---|---|---|
| P1 | Supported normalized disabled passthrough for both startup paths/builds | Baseline all60 HTTP/WS and final all60 HTTP/WS snapshots retain same normalized original/delivered snapshot facts. Earlier final plain16 and later17 with sentinel have separate identities; do not assert those are equal dataset pairs. |
| P2 | Partial | Independently compared full decoded records sorted by ID: settled31, pinned17, snoozed17, working17 equal HTTP/WS at same options. Counts include sentinel. Final genuine WS lineage on/off witnesses now exist: off56/on60, on-only A9–A12. Independently compared IDs, lineage, latestRun and runtime against HTTP with zero mismatches. WS compact records omit title/projectId/pinned/snoozed/settled fields present in HTTP, so full-record parity for those lineage chapters is not established. Shelf chapters independently carry those fields. Snapshot project identity currently retains IDs only; do not claim full repository metadata equivalence. |
| P3 | Supported observed real-event continuity on current uninterrupted stream | HTTP sentinel persisted16 before release→A1 inserted17→real persisted17 delivered18, decoded final title accepted. WS persisted18 before→A1 inserted19→real visited19 delivered20→sentinel persisted20 delivered21, decoded title accepted. Persisted journal correctly names orchestration_events (V1 shell path), not empty V2 table. This is distinct from P6 and does not prove held-row resubscribe transfer. |
| P4 | Partial: transformed delayed decoder fields supported by B-active reruns | A1 absent held, working HTTP release decodes running runtime/latestRun; status WS release decodes failed runtime/latestRun. Original A1-active release journals omit runtime/latestRun after opening detail; use B-active reruns for retained transformed-field proof. Working badge is visible in original release recording. Original WS status recording briefly shows Failed at A1 around PTS1.283s, then detail hydration removes it; B-active decoded rerun retains failed fields. This supports transient classification, not sustained M11 migration/restoration/header proof. |
| P5 | Supported recorded clock preflight and attributed pre-module injection | Epoch/now/newDate/callDate agree; explicit/parse/UTC946684800000, instance true; readyState loading and native performance advances ~50ms. Config journals also retain clock and actual palette. Module count2 alone cannot prove execution order; exact injection bytes/order must remain attributed. |
| P6 | **FAIL at a9dc, both HTTP and WS seeds** | Hold→same-socket foreground resubscribe→release sends A1 to retired request4/3 instead of current20. Current decoded A1 remains absent. Object envelope was actually emitted; no invented array/NDJSON browser matrix. |

P6 confirms review A round3, not the earlier fixed post-release cursor bug. Original request lifetime is retained in delayedStream/held closure; metadata-only current snapshot supplies no replacement row. Fixing product code is not implicated. Required regression: pending row transfers once to current shell stream, reducer accepts it and a subsequent real source event, next resume cursor remains correct, separately for HTTP/WS seeds. User explicitly requires committed fixtures, so this is within acceptance scope.

## Independent rebuild / final handoff

Final [handoff](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b757fe763a086c5b/final-handoff.md) completed11:11:01UTC. Final [artifact manifest](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/11f1f690b49cbc12/artifact-manifest.json) independently hashes clean with zero mismatches. The capture-index.md initially present is the old07:27–09:27 recovered index, not round2 final authority. Final [frame manifest](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/251e7604296a2930/frame-diff-manifest.json) and corrected narrow B sheets were inspected again after final handoff.

After the capturer's cleanup proof, independently confirmed clean checkout/free6229+14269 ports. Prepared `/tmp/pr23519-proof-pass2/home` by SQLite read-only backup from the permitted synthetic head-plain snapshot, copied environment-id/settings/keybindings, then applied the independently regenerated eleven committed SQL files once in README order while stopped. No live DB or credentials directory used. Independent preparation script SHA2568dcdd77399c8ccbff0ca71a8e8f71e79f29f62225a2e5b460794ac7f6e191c8c; regenerated-input hash ledger SHA256cdc8abd8e87a658348ae7f9d1bf8823dc440506b4e84ae28142139e53b99c4bd. Built bootstrap/replay injection directly from extracted immutable committed files (SHA25669fa40cabacad470122a554bb14f26de369b8b939b2172ecfbb9a0b35c06331e), inserted before module graph in temporary HTML. This deliberately omitted transport observer wrappers; it is a product/replay rebuild, not a byte-identical instrumentation bundle. Product files stayed44215; their a9dc equivalence is established above.

Ran the normal dev runner with explicit isolated home, paired one fresh native Browser incognito tab, reset only capture shell/thread/vcs-refs IndexedDB stores and exact delivered NL config. Actual390×844/DPR2/light/preview7/expandedA+B/frozen epoch matched. Started A1, opened drawer, ordinary-clicked B8, verified B8 heading, reopened drawer, recorded hardware Control+[ back to A1. Rebuilt settled endpoint has A8…A1 exactly once, active A1 at CSS top418, no A overflow, B8…B2 and its nonempty overflow retained. Native recording plus sanitized narrated action journal retained as `rebuild-M13-NL.mp4` and `rebuild-journal.json`.

**Behavioral endpoint reproduced; strict delivered-frame reproducibility remains incomplete.** Full780×1688 screenshot compared to delivered head-M13-stable-NL-A-active.png: raw773483(58.746734%) pixels differ; threshold>8 gives28554(2.168702%), bounds[0,0,755,1664]. See final-arithmetic.json. This compares lossless screenshot against decoded lossy MP4, so background/gradient/text-edge noise is expected, but cannot explain all changes: delivered B8 retains hover highlight whereas rebuild pointer stayed on drawer toggle; rebuild exposes the empty-conversation placeholder suffix at right and an additional composer/card outline below, absent from delivered frame. Rebuild A1 pill is more clearly rounded. Sidebar row content/order/positions match; these visible content/interaction-state differences are not waved away as AA or compression. Same committed data and endpoint route do not prove every transient view state. Lead should supply exact original pointer/renderer-ready/provider/composer state or accept a narrowly scoped residual; no product fix inferred.

Stopped only captured runner PID1342058 via SIGINT, session exit130. Closed only own Browser tab. Restored original HTML SHA256930433af15929fc4328bca8715dd2702760745a7a3743f69b456f0ae4021c2fa; removed only own inert page. Git status empty, both owned ports free. Lead-owned worktree/scratch and audit synthetic home retained. No product suite recapture.

## FRICTION

- Native recording/journal timing cannot be guessed: rejected earlier sheets demonstrated lag. Stable corrected endpoints were visually checked, and original lossy MP4 PTS are retained. Audit contact-sheet checkpoint offsets are inspection aids, not replacement provenance.
- Committed fixture README plus independently regenerated SQL/settings are usable; no committed chapter runner exists. Final handoff supplies caller/reset/storage/injection mapping; M13 exact renderer/pointer/composer state still not independently identical.
- No raw wire buffers, auth/session state, pairing links, DB, private server logs or secrets are included in this report or its upload.

VERDICT INSUFFICIENT — confirmed committed-fixture P6 failure; exact M1–M13 gaps above; P2 lineage full-field parity limit; narrow B render residual; independently rebuilt M13 behavior with unexplained nonAA endpoint deltas. Preserve supported chapters and repair only missing/invalid ones.

## Public audit evidence

- [Independent preparation script](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/43bb15daa34039d0/prepare-rebuild.py) and [regenerated-input hashes](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/ddba59d9ff82868b/independent-seed-hashes.json)
- [Final arithmetic and manifest verification](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/c85b59ea1cec2852/final-arithmetic.json)
- [Independent final lineage parity comparison](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0c197cc1f3eea9e2/final-lineage-parity.json)
- [Independent rebuild narrated action journal](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/76a4b3353738f2cd/rebuild-journal.json) and [native recording](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/5bbc65eaa1fa81d4/rebuild-M13-NL.mp4)
- [Full-resolution gain8 rebuild difference](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/5cde7efd3493b1a1/rebuild-diff0-gain8.png)

![Delivered M13 NL A1 endpoint on left; independent rebuild on right, with visible hover and exposed main-pane differences](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/e5f19c96399d93d9/rebuild-comparison.png)