Fix Codex request_timeout salvage diagnostics

[probepark]

Fixes #6301

What

Add the complete Codex timeout replay matrix from the issue and log the guard inputs when complete-tool salvage is refused.

Why

All 24 requested replay variants pass on the current dev baseline, so there is no RED on dev and changing salvage behavior would be unsafe. The refusal-path diagnostics make the production guard decision observable without changing timeout, retry, model, preset, or salvage behavior.

Acceptance

AC	location file/function	local test
Complete todo_write replay variants cover arguments.done/output_item.done, trailing item kind, and delta correlation	packages/ai/test/openai-codex-stream.test.ts / createCompleteToolTimeoutReplayEvents and replay matrix	bun test packages/ai/test/openai-codex-stream.test.ts
Refusal guard inputs are logged without changing the return path	packages/ai/src/providers/openai-codex-responses.ts / trySalvageCodexFinalizedToolCalls	Full Codex stream test file; existing fail-closed tests remain green
Changelog fragment validates	packages/ai/changelog.d/codex-request-timeout-tool-salvage-r2.md	bun scripts/changelog-fragments.ts check

Testing

• RED: no RED on dev. All 24 new replay variants pass on dev; there is no pre-fix failure to paste.
• bun test packages/ai/test/openai-codex-stream.test.ts -> 182 pass, 0 fail, 740 expect() calls.
• bun --cwd=packages/ai run check:types -> exit 0.
• bun scripts/changelog-fragments.ts check -> exit 0; 75 fragments across 5 packages validated.
• bunx biome check packages/ai/src/providers/openai-codex-responses.ts packages/ai/test/openai-codex-stream.test.ts -> exit 0.
• bun run ci:test:smoke -> exit 0; smoke-test: ok.
• Native addon build completed successfully before the CLI smoke test.

Risk classification

• low-risk
• regression-risk
• high-risk

Assumptions

• The current dev behavior is the intended baseline because every requested replay variant already returns toolUse.
• Existing partial-argument, non-empty-text, and non-transient-error fail-closed behavior must remain unchanged.

Approval

Merge approval is expected to stay red on this agent-authored PR until a maintainer approves the exact head (human-review path, #6037); the contract check itself should be green.

The PR targets dev; no merge or force-push is performed by this change. The branch was refreshed by merging current dev (no force-push) so exact-head CI contains the event base.

Agent

• First attempt (gjc ACP) ended with prompt_failed (-32603, "Agent run failed after execution started") right after a failed todo_write, with no commits; this PR was produced by the fallback agent (gpt-6.1-sol).

[snowykr]

Verdict

APPROVED

Summary

This PR adds refusal-path Codex timeout diagnostics and a 24-case replay matrix while leaving salvage eligibility and behavior unchanged. Review of head c3c19c8f2be52d3df4786b9c1d946cdcfac2c080 found no actionable defects.

Findings / Required Changes

No blocking or actionable findings.

Non-blocking Observations

• The added replay matrix verifies salvage outcomes, but does not directly assert the new debug refusal record. A focused debug-enabled assertion could protect its fields.
• The logged item type originates in the provider response. The diagnostic is opt-in and no material trust-boundary impact was established; allowlisting the value would be optional log-hygiene hardening.

CI / Verification

GitHub Dev CI run 37165385130 is completed successfully for the reviewed head. The affected-path test job for packages/ai/test/openai-codex-stream.test.ts passed. GitHub reports 18 successful and 25 skipped check-runs, with no failed conclusions, on the exact head. Tests were not run as part of this read-only review.

Axis Coverage

Axis	Verdict	Coverage
A1 — Intent / Policy / Contract	APPROVED	Diagnostic-only change matches the issue's conditional request; the existing salvage contract and changelog conventions are preserved.
A2 — Architecture / Correctness / Failure	APPROVED	The added refusal-path log reads existing state and does not alter salvage predicates, result state, retries, or error recovery.
A3 — Security / Privacy / Trust	APPROVED	No actionable boundary crossing found; the new log is debug-gated and does not expose tool arguments or identifiers.
A4 — Verification / Tests / CI	APPROVED	Replay permutations assert observable tool-use results; the affected test shard passed on the reviewed head. The diagnostic itself lacks a direct assertion (optional coverage).
A5 — Context / Compatibility / Platform	APPROVED	No public interface, packaging, platform, or consumer contract changes; existing debug logger abstractions are reused.

Limitations

This was a read-only review; tests were not independently rerun, so the PR's claim that all 24 variants pass on the current dev baseline was not independently reproduced.

[Yeachan-Heo]

Merged into dev as 61f2e172.

• snowykr approved the exact head c3c19c8.
• CI on that head: 18 passed, 0 failed.
• Base 33a29331 is 4 commits behind dev (feat(session): add fenced durable task-owner provisioning API #6303 and merge commits); none touch this PR's 3 files.

—
[repo owner's gaebal-gajae (clawdbot) 🦞]