Cap bounded robustness worker resources

[Anionix]

Summary

• apply non-raiseable worker limits before importing the heavyweight robustness graph
• enforce Linux address space plus Unix file-size, open-file, and finite real-user process ceilings while preserving tighter inherited limits
• record requested, child-confirmed effective, unsupported, and application-status evidence per case and per run
• reject missing or contradictory child limit evidence as HARNESS_FAILED / UNCONFIRMED
• record Darwin RLIMIT_AS as unsupported and track the remaining containment risk in [Accepted risk] Enforce bounded-worker memory containment on macOS #510

Research and lifecycle contract

• Primary contracts: https://docs.python.org/3.12/library/resource.html and https://man7.org/linux/man-pages/man2/getrlimit.2.html
• Limit setup failure remains incomplete, non-ranking evidence and never advances lifecycle state.
• RLIMIT_NPROC is a real-user aggregate, not a per-worker child count; the 512 ceiling leaves headroom for existing runner processes and library threads while remaining finite.
• Real UID 0 and Linux CAP_SYS_ADMIN / CAP_SYS_RESOURCE contexts record NPROC as unsupported because the kernel exemption makes the numeric limit unenforced.

Validation at exact SHA

Target: a64942fd17d6ac866f24fde77a3d60826e2b9cbe

• pytest tests/test_robustness_limits.py tests/test_robustness_runner.py::test_default_worker_does_not_claim_unconfirmed_resource_limits tests/test_robustness_profile.py::test_public_cli_runs_and_reports_bounded_fixture: 6 passed
• Ruff changed-file check: passed
• ty focused implementation check: passed
• Pyright 1.1.411: 0 errors
• fixed-SHA spec review: no findings
• fixed-SHA standards review: no findings
• Lean: NO_LEAN (Python/config-only slice)

CI and review feedback incorporated

• First exact-SHA run exposed Pyright narrowing errors; fixed with explicit typed validation.
• Second exact-SHA run reached 600 passed / 3 skipped but showed RLIMIT_NPROC=32 blocked every Linux bounded worker. The real-user ceiling is now 512; tighter inherited caps still win.
• Third exact-SHA run passed all checks. Pre-merge review then identified privileged Linux NPROC exemptions; those contexts now record the cap as unsupported instead of applied.

Diff note

The cohesive change is above the usual 250-line risk signal because the bootstrap boundary, evidence schema, report output, and success/failure/privilege regressions must land atomically; splitting them would create a dependent stacked PR, which this repository forbids.

Fixes #301
Refs #510

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 668e3ed3e1

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".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".