fix: stabilize and speed up Windows credential ACL verification

[kmbroai]

Summary

• Retry bounded Windows credential descendant enumeration when a concurrent ambient-auth import removes a temporary file between directory traversal and lstat. A missing credential-home root, persistent instability, symbolic links, junctions, and other filesystem failures still fail closed.
• Run the Windows-only ACL integration check through asynchronous execFile, preserving the actual subprocess error instead of surfacing only an unexplained synchronous status: null.
• Batch credential-home ancestry, home ACL, and descendant ACL inspection into one streaming PowerShell process per validation. The same owner, trusted-SID, inheritance, attacker-writable-ancestor, descendant-privacy, alias-resolution, repair, and constrained-language checks still run on every access.
• Keep the standard bun test --timeout 30000 timeout on every platform; the earlier Windows-only 60-second workaround is not part of the final diff.
• Add deterministic regression coverage for disappearing descendants, repeated instability, missing roots, one-process ACL snapshots, empty homes, unsafe ancestors, unsafe descendants, incomplete subprocess output, and the default CI timeout.

Root cause

The flaky failure came from two Windows-only races: a descendant could disappear between filesystem enumeration and lstat, and the synchronous PowerShell test could occasionally return only status: null. Separately, managed credential workflows repeatedly validated the same credential home, and each validation started separate PowerShell processes for ancestry, the home ACL, and descendants. That subprocess startup cost pushed legitimate tests close to the 30-second limit.

The batching change removes the repeated PowerShell startup cost without caching or skipping ACL validation. A populated credential-home validation now uses one streaming PowerShell process for all ACL descriptors instead of separate ancestry, home, and descendant PowerShell processes; repair paths still re-inspect after mutation.

Failure evidence

• Original Windows/Node 24 failure on main: https://github.com/openai/codex-security/actions/runs/31132067211/job/92723068033
• Windows/Node 22 concurrent credential import failure: https://github.com/openai/codex-security/actions/runs/31127158645/job/92702598234
• Earlier PR run showing the original PowerShell failure fixed but a different credential test reaching 30 seconds: https://github.com/openai/codex-security/actions/runs/31135518651/job/92733857846

Verification

• The new disappearing-descendant and persistent-instability regressions failed against the original implementation, then passed with the bounded retry.
• The combined-inspection regression failed before the batching implementation, then passed with the single-process snapshot path.
• Randomized local SDK suite after batching (--randomize --seed 12345 --timeout 30000): 950 passed, 10 platform-specific skips, 0 failed.
• TypeScript typecheck, generated-model validation, full Prettier check, production build, pnpm pack, and installed-package smoke validation passed.
• Current GitHub Actions run: https://github.com/openai/codex-security/actions/runs/31205123609
• All 11 applicable checks passed, including Windows Node 22 and Node 24; 5 publication checks were intentionally skipped.
• Windows Node 24 passed 972 tests with the default 30-second timeout. The formerly slow parallel-scan test improved from 25.02s to 14.46s, delegated credential login from 30.87s to 17.70s, and the full Windows suite from 399.86s to 328.22s despite the combined PR running more tests.
• Windows Node 22 passed 972 tests with the default 30-second timeout; hostile ancestor rejection, nested credential ACL repair, and constrained PowerShell regressions passed on both Windows jobs.