# Proof audit pass 2

Read-only proof-completeness audit for [#23840](https://git.cafe/pingdotgg/t3code/pulls/23840), final head [1fba4795e3](https://git.cafe/pingdotgg/t3code/commit/1fba4795e3342b9a4db956f25391bde81e0b0fab), merge-base [ef92441313](https://git.cafe/pingdotgg/t3code/commit/ef92441313da327139df75daa8f723c0f42348f6).

Applied the original [pass-1 checklist](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/369dbf4c67045a69/proof-audit.md) to [the complete transcript](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b3d471b9f4d92a3b/runtime-transcript.md), [head records/native frames](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/da6d2390ec3ba165/runtime-head.json), [merge-base records/native frames](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/3eed49e7152ff070/runtime-before.json), [controlled late-delivery runtime](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/b80f2210eb7634a9/runtime-gate2.json), [processing boundaries](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/7d80e1486e63eb11/processing-boundaries.json), [exact tests and 38 proof assertions](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/04764f3bca3d6dca/tests.txt), and [the driver](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/07b576f73b4335d2/driver.tar.gz). All ten artifacts were fetched back with curl and their SHA256s matched. Python urllib's default request was rejected with HTTP 403; curl succeeded. This is a proof audit, not another code review or merge-readiness approval.

## Fixture checklist: 8/8, with the original failure-coverage exception

1. **Direct native child after settled root — PASS, real runtime.** Final-head root ended 21:38:30.258Z < child ended 21:39:15Z < automatic continuation started 21:39:15.928Z. Durable agent/server-origin input contains the exact unique result; parent answers it correctly. One external prompt, one server-created input, two root runs, no parent wait tool or manual follow-up. Same driver/prompt/model bytes at merge-base: root ends 21:48:03.712Z, child ends 21:48:43Z, root remains at one run through a later separate-thread provider observation boundary.
2. **Original foreground root active — PASS, real runtime.** Child terminal 21:39:15.963Z occurs inside original root run 1, ending 21:39:51.125Z. Root answer includes the child result. No server wake or extra run after processing/settling.
3. **Later active root — PASS, real runtime.** Root 1 ends; explicit root 2 starts 21:39:11.034Z; old child completes 21:40:22Z inside root 2; root 2 acknowledges it and ends 21:40:56.783Z. Exactly two externally requested runs, no competing or queued redundant continuation.
4. **Nested child — PASS, real runtime.** Grandchild completes 21:39:25.329Z with direct child still running. An intermediate snapshot shows root at one run. Stronger FIFO boundary: resume run 3 is requested/delivered at 21:39:54.849Z–21:39:56.864Z, after grandchild completion but before direct-child completion at 21:40:09Z. Persisted root history has no second run until direct completion. Only the direct result causes a wake.
5. **Resumed child / result preservation — PASS, real runtime plus focused normalization.** Same child produces FIRST, native followup_task resumes it, then SECOND. Each distinct result appears once in a durable continuation; three completed root runs, correct SECOND answer, no old-result replay. The actual CLI did not emit a trailing empty final here; that exact normalization remains labeled focused-test coverage in green.
6. **Interrupted parent / late-result suppression — PASS, real runtime with disclosed transport control.** Ordinary cancellation stops root/child without wake. A stronger fresh-home fixture holds genuine CLI completion bytes, interrupts parent, awaits persisted interrupted root at 21:58:16.662Z, then releases the real successful child turn/completed at 21:58:16.674218Z and its original activity frames. The three held/released raw-line SHA256 pairs match in [the transport trace](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/da7bac6d826e8fa9/gate-transport-events.json); [the driver log](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/1c1cca34a572e25a/gated-cancellation-driver.txt) records the persisted milestone. Child work actually completed earlier, at 21:57:54Z; delivery was delayed deliberately. Server native logs show the successful frame arriving after root interruption. No fabricated event or production edit. Root stays interrupted, child projection stays interrupted with result preserved, no wake through a later eligible FIFO continuation. Only a subsequent explicit fresh user input starts recovery.
7. **Failed or interrupted child — PASS with precise split.** Real native grandchild interrupts its direct parent after root settled: root ends 21:43:39.290Z, direct child becomes interrupted 21:44:04.589Z, grandchild completes 21:44:06Z. No root wake through later queue processing. A genuine provider failed-child turn was not induced. As pass 1 explicitly permits, failed-child wake coverage is **focused-replay-only**: an external temporary copy of the final regression feeds a failed child after completed root and asserts exactly one message_text offer, failed outcome and preserved partial detail. Exact generated driver/transformation and passing output are attached. This does not claim real Codex consumed a failed-child wake. Neither a shell failure nor initial wrapper-start failures are misclassified as failed-child evidence.
8. **Recovery after stop — PASS, real runtime.** Explicit fresh turns after ordinary cancellation and after controlled late delivery both return their exact recovery markers. Interrupted work never spontaneously restarts. Controlled fixture has exactly two root runs: interrupted original and explicit recovery.

## Supporting checklist

- **Red/green output — PASS.** Exact three-file command from apps/server. Final head: exit 0, 104 passed. Merge-base production with unchanged final regression driver: exit 1, 4 failed / 100 passed. Two explicit expected-1/actual-0 assertions; two existing 120-second waits time out because required continuations are absent. Red timeouts are not used as passing no-wake evidence. Original merge-base tests separately pass 100. Production tree matches merge-base; only final regression test overlay differs.
- **Provenance — PASS.** Exact head/base, real Codex 0.163.0, gpt-6.1-sol, low effort, absent service tier, full access/never-approval policy, OS, isolated homes, native IDs/parent relationships, prompts/markers, spawn/completion events, persisted run lifecycle, message text/origin and resulting answers are recorded. Principal before/head use identical driver/prompt SHA256s. Unmodified transport is used outside the separately disclosed cancellation-ordering fixture.
- **Observation and duplicates — PASS.** Persisted terminal milestones precede later eligible continuation requests consumed by the shared FIFO worker. Final head snapshot 21:45:43.834928Z and controlled snapshot 21:59:29.895993Z have no active children or pending/running effects. Expected run/input counts remain exact. Initial intermediate sequence=null came from querying the retired empty V2 event table; projections were valid and later exports correctly use orchestration_events. No proof rests on a short sleep or early run-count sample.
- **Visual/surface matrix — PASS, 0/0 fixtures.** No rendering paths changed. Shared-server runtime proves the behavioral outcome; no screenshots or additional relay/tunnel matrix is needed.
- **Isolation and cleanup — PASS.** Fresh /tmp homes, seeding disabled, no real T3 database read/copied/opened. Public exports omit actual credentials, pairing URLs, injected instructions and reasoning. Source restored byte-for-byte with empty git status; temporary worktree removed. All captured server/watch/native CLI PIDs exited and port released ([PID verification](https://gh-file-drop-api-prod-mi5fy3sowv63ufte.pinglabs.workers.dev/f/a59859d413ed25df/process-cleanup.json)). Evidence directory is removed after the authorized body append.

38 artifact assertions passed. Exploratory stop/self-interrupt restrictions and the first control-socket collision are disclosed and excluded from required proof; a unique socket and fresh home resolved the wrapper setup. No commit, push, merge, review, comment, assignment, browser or PR watch was performed. This audit is read-only; publication is the separately authorized body append.

Corrections to earlier PR claims: final scoped count is 104 rather than 103; earlier runtime stays attributed to its older implementation. Final-head runtime and guard coverage now have attached evidence. Failed-child and empty-final limitations remain explicit.

Model and harness: gpt-6.1-sol through Codex, evidence collection and read-only proof audit pass 2.

VERDICT: SUFFICIENT