# Proof audit PASS 2

No blocking proof findings. Every mandatory affected matrix row is supported. The independent fresh-home R1 rebuild reproduced the delivered after behavior. The checksum packaging defect found during this pass was repaired and independently rechecked; it leaves no missing row.

Original [#14496](https://git.cafe/pingdotgg/t3code/pulls/14496); replacement [#24616](https://github.com/pingdotgg/t3code/pull/24616). Before [5a969c17e7](https://github.com/pingdotgg/t3code/commit/5a969c17e76301e45e53e1fd9aad2f5b4599d865), full SHA `5a969c17e76301e45e53e1fd9aad2f5b4599d865`. After [a7c675caa1](https://github.com/pingdotgg/t3code/commit/a7c675caa1f2ea52842c77ebc6fe6cc874d6ec01), full SHA `a7c675caa1f2ea52842c77ebc6fe6cc874d6ec01`. Local HEAD, merge-base, transcript/manifest SHAs and replacement PR head agree. Clean source worktree. Five-file content diff, 266 insertions and three deletions: worker, orchestrator, internal contract and their two existing test files. No adapter/client/transport/migration source changes.

## Row-by-row result

| Matrix row | Evidence inspected | Result |
|---|---|---|
| F1 / R1 | Paired actual-server question captures; all RPC/subscription records, casts, four SQLite exports, projections, outbox and receipts | PASS. Five attempts; no warnings during retries. Before zero warnings, after one committed error received online and on reconnect. Saved Blue answer, resolved request, completed node, running run and session stay intact. |
| F2 | Optional success/transient/ambiguous controls; source review | Correctly omitted as optional. Conservative unconfirmed-delivery wording is appropriate; evidence does not claim proven non-delivery generally. |
| F3 / H1 | Paired executor pure interruption, typed, defect and mixed-cause segments | PASS. Pure interruption propagates without reporting; non-pure terminal causes report after only. Retry remains silent. |
| F4 / R2 | Paired actual-server native Bash approval with accept decision | PASS. Five attempts; before zero warnings, after one. Saved accept decision and request/node/run/session remain intact; reporting emits only the error event and no policy/run action. |
| F5 / H1 | After-only actual isolated runtime services with synthetic detached/null-run request; missing request, missing node, deleted and missing threads | PASS. Null-run event omits event.runId and has payload.runId=null with node/provider correlation. Absent/deleted context emits no event. Direct dispatch reports existing no-domain-events errors, explicitly disclosed; missing thread retains lookup error. |
| F6 / H1 | Paired reporting-command failure; after-only replay; real-app event IDs | PASS. Reporting failure logs a warning and preserves original failure. Command ID derives from response command; item ID derives from request. Replay returns same event/sequence with one item and receipt. |
| F7 | Unchanged validation paths traced in orchestrator/runtime service | PASS as source-reviewed coverage, not claimed runtime coverage. Missing/resolved/not-resumable/feedback validation precedes delivery; message capability uses message dispatch. |
| F8 / H1 + R1 | Complete internal schema and omitted required-field cases; internal routing; ordinary typed error event; public schema trace | PASS. Required-field checks reject incomplete internal command; new command remains absent from public dispatch input. Direct schema/routing cases labeled after-only. |
| S1 | Before/after stills, stability pairs and both full-frame diffs | PASS as optional support. No missing mandatory browser row. |

## Actual visual inspection and provenance

Downloaded all 18 files indexed by the original upload index, including the actual five public MP4s, five full transcripts, before/after PNGs and both diffs. Their bytes match canonical local artifacts. Fully decoded all five videos with ffmpeg and viewed chronological frame samples from every downloaded video, in addition to the delivered opening/middle/final sheets and H1 after-executor frame. Fresh samples and download hashes are retained in pass2. Video durations: R1 before 12.25s, R1 after 12.75s, R2 before 12.25s, R2 after 12.70s, H1 21.95s. Initial labeled startup holds give way to live request/event output; final holds contain complete results. H1 visibly shows paired executor and explicitly after-only direct-command cases. Scripts and ffmpeg logs show actual x11grab from owned Xvfb/xterm while the drivers run, not post-hoc log playback. Casts exactly reproduce the full real-app transcript records.

R1/R2 satisfy the task's real-app definition: actual isolated `vp run dev --no-seed --home-dir ...`, authenticated typed Effect RpcClient/WsRpcGroup/Socket JSON WebSocket calls, normal thread launch and answer dispatch, unchanged ClaudeAdapterV2/SDK spawning a disclosed deterministic stdio process. The external overlay only replaces the selected allowlisted `runtime.respondToRuntimeRequest(input)` boundary with Effect failure before forwarding. It does not replace the server, RPC, adapter, worker, persistence or projections. Provider protocol logs show native AskUserQuestion/Bash requests. No paid live LLM or OAuth was required or claimed. H1 is correctly identified as harness evidence, including its private-schema exposure loader.

Compared every tracked archive file to its Git object: 28,261 regular files per revision match. The tracked `.claude/skills` symlink also matches its Git target string; an absent target is unrelated to runtime source. Symlinked dependencies and generated caches are outside tracked-source comparisons. Actual adapter source is unchanged.

All four delivered captures have before-response, committed, preterminal and settled exports. Independent verification correlated response-command IDs, deterministic reporting-command IDs, event metadata/payload, outbox attempt_count=5/status=failed, online event and reconnect projection. The only event after committed response is the new error after; before has none. No reporting-command request/node/run mutation occurs. Observed backoff gaps are approximately 104–107, 205–207, 406–410, 805–822ms; these are actual wall-time attempts, with monotonic cast timing supplementing UTC records.

## Stills and claims

Actually viewed before, after, Diff0 and DiffT. Recomputed both diffs and confirmed 5,544 changed pixels at threshold zero and 5,035 at max-channel >16 out of 1,296,000. Both 500ms stability pairs are pixel-identical. No clock freezing, image masking or crop appears in the browser capture script. The listed visible changes are accurate: warning icon/title/two-line body and timestamp, Thinking shifted down, incidental elapsed Working label 5m32s to 5m37s. Remaining pixels are identical. No unintended visual difference was found.

The current evidence sheet and PR evidence claims are supported within their stated scope. Other providers, desktop/mobile, relay/tunnel, live paid LLM and optional control cases are expressly not independently exercised. Current forge state reports 28 successful checks at the final head. Historical local test counts and private review outcomes are lead-verified per the parent handoff; this pass did not recreate those historical runs or independently read private review outputs. They are distinct from the independently audited captures. The original author's description is retained as attributed historical context, not proof of test files present on the takeover revision. This is a proof-sufficiency verdict, not a new comprehensive merge audit.

## Independent rebuild

Preserved delivered `after/home` as `after/home-captured` and `after/question` as `after/question-captured`; also copied mutable manifest/log/credential/target files into private pass2 scratch before launching. Ran the supplied launcher on the verified final source with a newly created home, awaited backend port19002, then ran `python scripts/record.py after question`. Fresh transcript span: 2026-10-25T00:29:49.783Z–00:29:56.233Z. This rebuild is transcript/cast/state evidence, not a new MP4, as permitted by the brief.

Rebuilt semantic result equals delivered R1 after: pending live question -> committed Blue answer/resolved request/completed request node -> five failed, non-forwarded attempts -> one transport_error notice -> online receipt and reconnect persistence. Request, nodes, running run and provider session stay equal from commit through settlement. Rebuild attempt gaps: 107,207,406,806ms. Generated IDs/times/home contents differ; saved answer, statuses, warning shape, count and correlation do not. The paired before capture has zero error items; both delivered after and rebuilt after have exactly one. Full rebuild transcript, cast, four DB exports, projections, manifest and verdict are included in the separate audit bundle.

Stopped the captured launcher process group with the supplied stop script. Checked captured provider PIDs absent and ports19001/19002/5933/5934 without listeners. No live T3 userdata, source edit, push, PR comment, watch or delegation. Only required app PR-link metadata was registered. No browser was started in this pass.

## Corrected packaging verification

Initially found stale self-checksum in SHA256SUMS.json; all 67 other archived payload files already matched. Parent repaired only external packaging/public-bundle metadata. Downloaded the corrected public archives and compared bytes with local deliver archives. Corrected state archive changes only SHA256SUMS.json, excludes its own entry, and passes all 67 payload hashes. Corrected scripts archive changes only scripts/package.py, still includes all 20 scripts, and matches the local script. No recorded artifact changed during repair.

[Corrected state bundle](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b33b801204ae9090/ev-14496-transcripts-state.tar.gz) · [Corrected scripts bundle](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/f025da02dcdedf79/ev-14496-scripts.tar.gz) · [Updated evidence sheet](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/3cce177a392cadf8/evidence-sheet.md) · [Updated index](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/1650ea4ca25078c6/upload-index.md).

Known original/rebuild dev credentials are absent from all downloaded artifacts and all archive payloads. No credential/auth/env/live database files are packaged. Provider fixture contains no credential or network request. This is evidence-based credential inspection, not a claim that arbitrary secret detection is mathematically exhaustive.

FRICTION: tool — urllib download returned403; curl succeeded. runbook — repeated packaging included stale manifest self-entry; parent repaired, independent verification passed. infra — none blocking. brief|forge|policy|model|other — none.

VERDICT: SUFFICIENT