[thomas-bartlett]

Summary

• Move Codex state outside the task workspace and give all conditions the same minimal environment notice.
• Expose and record generation limits; default to 16 turns and 600 seconds, with a 1,200-second task backstop.
• Support both CodeGuard skill layouts without inferring versions, and strengthen compatibility and sandbox tests.

Dataset, Semgrep contract, scoring semantics, dependencies, and skill contents are unchanged. Comparisons should use matching environment and budget settings.

Validation

• 286 non-Docker tests passed.
• Five mock-backed Docker integration tests passed; test containers cleaned up.
• Lockfile, pinned artifacts, task discovery, compilation, and Compose configuration verified.

Follow-up to #119.

[aacarter1]

No blocking issues found, lgtm