[abruption]

Summary

Adds an explicit, local preference for dismissing the automatic runtime notice for one stable release (#294):

• update --dismiss-runtime-notice VERSION and update --clear-runtime-notice-dismissal are offline preference operations, mutually exclusive with checking and all remote/SSH options. They never install, fetch, approve or mark a release current.
• Only the exact normalized version is hidden. Later release notices remain eligible. Explicit local checks retain actual availability/outdated/package-manager information and expose the preference separately. Skill notices, remote comparisons and the global opt-out are unchanged.
• The dedicated schema-v1 file is separate from the update cache. Strict bounded JSON, private ownership/modes, no-follow directory-relative I/O, a stable bounded lock and atomic fsynced replacement protect the preference. Invalid/future files suppress nothing and are not silently overwritten by either operation. A post-replacement durability failure reports that the preference may have changed.
• Initial support is macOS/Linux. Windows rejects the per-version write before POSIX storage access, while normal updates/checks/global opt-out remain unchanged.
• Updates the four CLI reference locales and regenerates the dependency-free standalone artifact. No skill source, global preference, runtime version, release or operations change.

Refs #294

Validation

• Python 3.9 isolated environment: 61 focused update/preference tests OK (one native owner-policy skip); 65 documentation tests OK.
• Final Python 3.13 extras suite with -X dev -W error::ResourceWarning: 1,355 tests OK (20 skips). The initial pre-review candidate also passed 1,352 tests; the final run includes the added review regressions and reviewed translation corrections.
• Generator check and git diff --check pass. Tests require the actual generated standalone artifact, not a source fallback.

The approved local temp volume ignores ownership. Controlled-metadata fixtures use actual directory-relative I/O, inode checks, locks, fsync and atomic replacement; they are not native permission certification. The unmocked positive test skips on unsafe real ancestors and is intended to execute on an owner-enforcing Linux CI filesystem. No internal-disk permission-test exception or volume/global permission change was used.

Independent review

Before PR publication, independent-cli-review ran Claude Code 2.1.308 (requested/observed model claude-opus-5-5) and Codex CLI 0.163.0 (isolated default; actual model not exposed) in parallel with identical read-only, frozen, self-contained source/test/document excerpts. Both completed with exit 0; Claude reported no error, Codex completed its turn with zero command-execution events, and the input remained unchanged (SHA-256 5cad25dcb2732491872345adaf0858914c960210a1a40e0a0baaa222da004718). No runtime execution by the reviewers is claimed.

Root-verified corrections: inode-bound cleanup on temporary validation failure, inherited setgid-compatible private mode checks, unambiguous check wording, removal of test source fallbacks, real-CLI remote-check advisory/result separation, and an inherited four-locale correction that newer remote runtimes are not downgraded. The original input overclaimed that an actual remote CLI check never reads the caller's local preference; this was a wording error, not a remote comparison defect. The intended local ambient advisory reads once and does not hide remote versions/results. Narrow corrections were separately source-audited and regression-tested; no claim that the two CLIs re-reviewed the final corrected revision. CLI agreement is not proof of correctness.

Limits

• This hides an advisory only, not a security finding or update availability. The selected version need not exist or be trusted.
• No native agent, SSH account, live update, installer, deployment or package publication was exercised.
• No automatic repair of malformed preferences and no Windows per-version persistence in this change.

Translation quality review

At the user's request, Antigravity CLI 1.4.5 explicitly selected the available gemini-3.8-flash-medium model with medium effort, plan mode, sandbox and a five-minute limit. The task-local runner omitted permission bypasses; one fresh read-only review completed with exit 0 and structured output (149.536 seconds). The input contained the complete changed Updating section in English/Korean/Japanese/Simplified Chinese, with commands and surrounding context (20,196 bytes, SHA-256 c3e614fbbc308019f200239b1abd412dd5b42ccf49d0509c9dc5028b0dcd8dd3). The model selector is recorded; an independent backend identity is not exposed by the result.

Codex checked suggestions against the English source and applied precise fallback/ACL wording, invalid/future-schema file subjects, Japanese cache-write wording, and the Chinese newer-version distinction. Flags, code blocks, paths, numbers and behavioral limits remain unchanged. This is static linguistic/semantic QA, not native execution, CI evidence, security certification or native-speaker field testing. Final local docs checks are recorded separately.

[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 2026-10-23T01:36:06.945775Z	13ece10	PR opened
🔒 Security Review	🔄 Running since 2026-10-23T01:31:37.860408Z	13ece10	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.

[abruption]

CI follow-up at 3deadee: canonicalize only the isolated test fixture root, so macOS /var aliases do not prevent positive fixtures from exercising the preference logic. The production no-follow guard is unchanged. A real ancestor-symlink regression still refuses reads/writes without effects, while the explicitly resolved fixture succeeds. A synthetic alias reproduces the original unsafe-versus-missing failure signature; the historical runner path itself was not printed, so that exact path is not asserted. Python 3.9 and 3.13: 62 focused preference/update tests OK each (one existing permission-enforcement skip on the approved external test volume). No runtime, documentation, version or permission relaxation. New-head CI is pending.

[abruption]

Exact-head CI verification: 3deadee now has 28/28 successful checks, including both previously failing macOS Python jobs and the release gate. The macOS Python 3.9 log explicitly reports test_native_private_path_when_filesystem_enforces_metadata and test_temporary_root_alias_refuses_while_resolved_fixture_succeeds as OK (not skipped). This supplies real runner owner/mode-path evidence separate from the locally controlled-metadata fixtures; it does not prove ACL audit, arbitrary filesystem support or all deployment environments.