Make C09 recovery finite and verify the real Codex CLI lifecycle

[KeyffMS]

Completed repair

Final reviewed head: 1629eb1d3200b4ac36d78caf16624c2b65a4e6f7, based on main 44f95258079d6b3ab5b79c178815bbc085edbe02. The final diff contains exactly 13 intended files. Temporary publication files/workflow are absent and will not enter main's squashed change.

• Canonical checkpoint next_action and prompt both describe C09_FINITE_RECOVERY, not the stale bootstrap MAP_INSTRUCTIONS action.
• Exactly first → second → finish. Every command performs real read-only checkpoint/Git reconciliation and canonical-file reads. Only first/second produce bounded compaction stimuli; finish requires the final structured result.
• C09 owns its 8192 body-after-prefix threshold; the hidden 200-token compatibility override is removed. The live 900-second timeout is unchanged.
• Require ordered real command receipts, two automatic PreCompact → PostCompact → SessionStart(compact) cycles, reconciliation after the second cycle, completed positive output and unchanged source/planning Git and file bytes, including ignored state. Partial counters, timeout or incomplete telemetry cannot pass.

Verification completed before merge

Final PR CI #116: SUCCESS on the final head. All 8 jobs passed: six Ubuntu/macOS/Windows × Python 3.11/3.14 jobs, distribution/release, and actual pinned Codex CLI conformance. Earlier CI #115 also passed all eight jobs with the same executable implementation.

• Product suite: 67 tests.
• Distribution/harness suite: 176 tests, OK (one test intentionally deferred to the separate CLI job).
• Real CLI job: 4 tests, no skips; actual Codex 0.153.4 performed 3 real tools, 2 real compactions, 2 compact recovery-context deliveries and 1 completed turn against a deterministic loopback Responses peer.
• The real CLI test initially caught a simulator error: ordinary Responses Lite requests can omit tools. It now uses canonical request_kind metadata, with executable missing/conflicting/no-tools regression cases.
• Compilation, evidence template hash/structure validation, candidate release checks, deterministic archive build and whitespace checks passed.

The conformance peer chooses responses/token usage. This verifies actual CLI integration, not autonomous model completion, and does not create or promote live capability evidence.

Hosted CI prerequisite, scope and remaining live qualification

The separate disposable hosted Ubuntu conformance VM temporarily enables unprivileged user namespaces and removes the host AppArmor restriction on creating them, matching pinned upstream Codex .github/actions/setup-ci/action.yml. Original sysctls are saved and restored in an always-run cleanup. Distro bubblewrap is preflighted before the CLI. Codex remains non-root/read-only, approval never, model-tool network disabled. The self-hosted runner's security policy is not changed.

C10/C13 runtimes, product .agents/.codex payload, live sandbox/approvals/auth ownership, artifact packaging and full release requirements are unchanged. The audit records exact pinned upstream source and the observed test failures/corrections. No self-hosted Codex qualification was dispatched.

After merge and post-merge CI, launch a NEW PlanAnvil Codex qualification, branch main, mode full, as requested. The full model-backed C01–C16 run remains the final live verification; no 100% prediction of its outcome is claimed.