[abruption]

Problem and change

The macOS contract still needs launcher-inheritance and exec-in-place evidence before a refusal-only filter can be considered. Add two unpublished synthetic experiments: Rust Command with default CLOEXEC versus explicit inheritance, and a C actor that execs at the same PID while retaining its lock. Compare the actual observations with expected tables, reject malformed/incomplete output, and verify normal exit and lock release.

Run these fixtures in a dedicated macOS 15 ARM/Intel CI matrix with Rust 1.90.0 and 1.95.0. Results report the measured scope; they do not implement the filter, validate production refusal, or select a recovery policy. Refs #127, which remains open.

Also link the completed Windows evaluation to its newly created W1–W7 follow-up issues #146–#152. #122 remains closed as the initial evaluation; its unresolved evidence continues in those issues.

Validation

• Local existing package checks passed: npm ci/check/lint/test (263 passed, 3 Linux-only skips)/smoke:package.
• Local same-PID experiment compiled with clang and observed unchanged PID/start time, changed executable path/argv, a held lock before/after exec, and a free lock after normal EOF exit. All fixture children exited normally.
• Rust is unavailable locally. Research CI run 39084774350 built and linted the fixture with Rust 1.90.0 and 1.95.0 on macOS 15 ARM/Intel. All four combinations passed all four inheritance observations and both final free-lock probes; error/incomplete counts were zero. Compiler metadata and source/binary hashes are in the logs. The same-PID exec experiment passed its seven checks on both OS architectures with normal child exits. These are synthetic measurements, not native Codex or product-policy validation.
• Read-only independent CLI reviews completed on the same frozen source input (SHA-256 0ea0290f80a7159b8879feaac3213ed644dae41c32c1c716152da1cdfb61adbf): Claude Code 2.1.302 returned claude-opus-5-5; Codex CLI 0.163.0 used isolated default model selection, actual model unconfirmed. Both were source reviews without execution. Subsequent corrections record early failures, actual FD flags, final free probes, cleanup errors and diagnostic metadata, and narrow stale evidence claims. The earlier reviews do not certify those later changes; the integrating session validated them and current-head CI/review remain separate gates.
• Focused Python helpers passed 11 tests without skips. The final C fixture rerun passed all seven expected observations with no pending process. Workflow YAML parsing and staged whitespace checks passed.

Remaining gates

Mach fileports, hidden shared descriptors, permission-limited observers/targets, more lsof/filesystem variants, native Codex behavior and production revalidation remain unverified. FWASLOCKED stays a refusal-side clue, not proof of the current holder. No production source, package content, platform support or release changes are included.

At head 6082a34a664078491df4904f19e38c69f3c5d0aa, all 43 GitHub checks passed. Automatic Code Review and Security Review both completed for that commit with no findings. The PR is open for human review; no merge or publication was performed.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-18T01:01:30.823563Z	6082a34	PR opened
🔒 Security Review	✅ Completed 2026-10-18T01:02:04.598140Z	6082a34	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.