fix(ci): stabilize Windows race regressions

[steipete]

Summary

• inject the retryable path-mismatch failure directly when testing JSON retry exhaustion, preserving the exact five-attempt budget without depending on Windows rename identity timing
• retry only transient Windows directory-rename denials while arranging the real parent-swap regression
• propagate post-lstat test-hook failures so a failed race setup cannot be silently treated as an ordinary path inspection error

Proof

• pnpm test test/json.test.ts test/path-stress-regression.test.ts (10 repeated runs)
• pnpm check (100 files passed, 1,035 tests passed, 61 skipped)
• git diff --check
• Codex autoreview: clean, no accepted/actionable findings

Fixes the Windows failures in coverage run 30976077742.

[steipete]

Landed as 8cc745a20b1d19340b27aaf1e73b1247a02297cd.

Verification:

• pnpm test test/json.test.ts test/path-stress-regression.test.ts passed across 10 repeated local runs.
• pnpm check passed: 100 test files, 1,035 tests passed, 61 skipped, plus build, docs, filesystem-boundary lint, file-size lint, and package checks.
• git diff --check passed.
• Codex autoreview was clean before commit and again immediately before landing.
• Exact-head CI at 1307bf321daee2d52013c602dff5c924edf317da passed the full CI, CodeQL, benchmarks, and coverage matrices, including Node 22 coverage on Windows, Windows Node 22/24 checks, native Windows checks, and Windows package smoke.

One unrelated Ubuntu coverage attempt timed out in native-publish-equivalence.test.ts; rerunning the failed job passed, and merged coverage then passed. No verification images applied to this test-only CI fix.