# Proof audit pass 1

Audited [#22775](https://git.cafe/pingdotgg/t3code/pulls/22775) in the requested checkout, exact head [80d27e96db](https://git.cafe/pingdotgg/t3code/commit/80d27e96db888d3ab0dbf12314f136f4e8b3e9b4) against [21da3061aa](https://git.cafe/pingdotgg/t3code/commit/21da3061aa006fc852ea4bd84c31ec084a3afa26). Read complete six-file diff, implement.md, implement-followup.md, review-2-claude.md, review-2-codex.md, relevant fixture/worker source and baseline/final evidence logs. Both round-2 code reviews conclude MERGE with no confirmed finding. This audit adds no production-code finding.

[Published matrix](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/10444e208cf49342/matrix.md). Local matrix: /tmp/issue-16815-evidence/matrix.md.

Evidence required: five paired scenario captures R1 prearm, R2 already armed, R3 Stop-before-arm, R4 manual Resume/held queue (including active rejection), R5 post-reset success. Capture both requested full SHAs, inputs, committed event order and observable projections; preserve failures on BEFORE. Include candidate identity and shell failed/usage_limit on the positive pre-reset paths to cover both changed SQL selectors. Include stopped-source marker for R3, queueHeld false after manual continuation but still queued for R4, and no added run from stale automatic/manual requests for R5. Repeat automatic command identity for idempotence in R1/R2.

R6 active supersession, R7 ordinary/user-first work, R8 reset/null/invalid boundary, R9 failure/root identity and R10 status/guard cases are test rows; no additional runtime capture or UI screenshot is needed. Matrix lists exact changed branches, fixture payload templates and before/after predictions per row. Existing committed recovery/client tests are reusable; supplied final logs report 99 passing tests. Independent parity diagnostics cover 28 histories but their temporary script was removed; do not call that diagnostic committed reproducible coverage.

Important distinction: supplied tests exercise EventSink/Orchestrator/SQL with synthetic provider history. They are isolated event harness evidence, not dev-server or Claude adapter evidence. Original baseline establishes the two completed-wake failures, while follow-up baseline used intermediate 34a6a9635a and cannot satisfy merge-base BEFORE capture. Neither report contains an actual isolated dev-server transcript at either requested revision. Real Claude quota exhaustion remains unverified; it is not needed for the explicitly synthetic server-state proof.

Recommended smallest next step: a committed test/development-only seed-and-capture runner and versioned fixture recipe, reusing existing internal services and synthetic provider setup, with explicit disposable home guard. No direct projection SQL and no production transport/MCP/permission surface. Seed with server stopped, close database, then observe and dispatch via the real running isolated server's existing transport; exclusive seeding between stages must remain clearly labeled. Log exact payloads/IDs/timestamps and await persisted milestones. In-process service execution alone must stay labeled harness, not dev server. Do not install/create checkouts to acquire BEFORE under this scope; report it missing if unavailable safely.

Root disk checked first and last: 12 GB free, no cleanup needed. Production checkout clean and unchanged at requested head. No installs, worktrees, browsers, server starts, tests rerun, pushes, comments, forge mutation or unrelated work. Only audit artifacts written under /tmp; matrix uploaded using configured ghdrop as requested. PR app registration performed and verified. No gh command or GitHub API read was used.

VERDICT: INSUFFICIENT — missing R1–R5 paired actual dev-server captures at the requested SHAs, plus committed reproducible runtime fixture runner/data.