# Proof auditor pass 2: PASS

Target: [#24949](https://git.cafe/pingdotgg/t3code/pulls/24949).
Exact audited HEAD: de68dbb1e55355950669fbb33a67aa0f6e2a8c8e.
Frozen before / current origin-main merge-base: 9296c20e82cef89371e2a4fb433f31c314384261.
Checkout: takeover/pr-24949, clean. No target source differences from 7a8a916e4d to audited HEAD in checkpointing, orchestration-v2, VCS, or ProjectService.

Evidence audited: matrix.md, transcript.txt, audit-notes.md, external proof.ts, both isolated SQLite databases opened read-only, and both throwaway Git repositories under /tmp/ev-24949. No live user data was accessed. No repository changes, server launches, browser use, delegation, or forge writes were performed in this audit.

## Findings

- Both successful databases independently contain exactly three threads, two completed runs per thread, and three ready checkpoints per thread. All nine original Git object IDs in each run resolve to commit objects.
- Recomputed scope namespaces using the current source SHA-256/hex/base64url algorithm map each initial ref to the claimed thread. Before retains D's three refs; after removes D's three refs. Archived A and retained K each keep all three refs with identical object IDs. Final snapshot files match current repository for-each-ref output exactly.
- Both databases confirm committed D deletion and A archive. Before has no checkpoint.cleanup effect. After has a succeeded checkpoint.cleanup effect with attempt_count 4.
- Transcript shows the after effect pending at attempt 0 after deletion commits, pending after Git exit 128 at attempt 2, then the same effect running at attempt 3 after restart with the owned lock retained. It returns pending at attempt 3 with another Git exit 128, then succeeds at attempt 4 after lock removal. This demonstrates persisted retry across a real process restart.
- Source matches the proof: deletion plans persist thread/scope ownership and candidate repositories into a durable outbox request; execution uses namespace deletion; Git lock failures propagate; checkpoint.cleanup is replay-safe and exempt from the finite attempt budget. Cleanup is queued last to avoid obstructing other deletion cleanup.
- External harness performs normal server RPC operations and read-only SQLite observation. Its only manual Git setup is repository initialization and an initial empty commit. It does not inject checkpoint refs or events. The manually created .lock file is an exclusive, owned failure fixture, subsequently removed. Committed focused tests do manually seed refs/events for their separate unit/integration coverage; those are not the running-server proof.
- All captured server PIDs (1569866, 1570921, 1571904, 1572683) are absent from /proc at audit time.

## Limits and excluded attempt

Provider turns use the existing local ACP mock fixture, not commercial-model inference. They make no repository file edits. The real server creates ready checkpoint commits over unchanged contents; this proves ref ownership, removal, persistence and retry, not provider edit quality or diff fidelity. This limitation is clearly stated in the published matrix and audit notes.

The first attempt is explicitly included as failed in the transcript and preserved separately. Its isolated database contains one completed run, two ready checkpoints, no deletion and no cleanup effect. The corrected external harness supplies the message.dispatch attribution fields required by the existing schema and starts a fresh successful home. No evidence suggests a hidden product defect in the excluded attempt.

The running-server proof exercises root-scope V2 refs, one repository, one archive and a graceful process stop/restart with retry pending. Legacy refs, orphan ordinals, retry beyond the default budget and lost in-flight claims are covered by the committed focused test source, not claimed as additional live observations. Final scoped check results belong to the parent and were not rerun or certified by this proof-only audit.

Verdict: PASS for truthfulness and sufficiency of the stated before/after runtime proof at the exact audited HEAD. No blocking proof gaps found. Historical launch SHA/cleanliness is supported by the captured transcript and harness, rather than independently re-executed during this read-only audit.

[Published matrix](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/0d94fb4e81b15c52/matrix.md) · [Published transcript](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/214f9f34ad2eea1f/transcript.txt)