[danshapiro]

Summary

A cloud vitest shard lost a whole base-gate run (2026-10-29) to an unhandled
ENOTEMPTY from fixture.cleanup() in the Claude permission-contract harness:
the sidecar's native CLI outlived the sidecar-only SIGKILL and kept writing
into the fixture while fs.rm swept it, the rejection escaped
runScenario's finally, and the harness died before printing its one-line
receipt — so the suite failed with "expected one receipt line" and the shard
crashed out.

Three layers, test-file-only (no product code):

1. runScenario now spawns the sidecar detached and, on the force path,
   kills its whole process group (the native CLI is the fixture's concurrent
   writer), then awaits the exit before cleanup.
2. fixture.cleanup() retries ENOTEMPTY/EBUSY stragglers
   (maxRetries: 10, retryDelay: 100) — fs.rm's default maxRetries=0
   turns any file recreated between the unlink sweep and the final rmdir
   into a rejection.
3. A cleanup failure is caught and logged (fixture.cleanup_failed); the
   receipt reports it as fixtureRemoved=false instead of the process
   crashing and eating the receipt for every scenario at once.

Test evidence

• New sabotage test (red first — the harness crashed with zero stdout on
  the unwritable-fixture EACCES; then green): with every fixture root made
  unwritable after its setup, the harness still prints one parseable receipt
  with fixtureRemoved=false for all eight scenarios and exits 1.
• Impacted sweep: all 7 suites in test/integration/providers/ pass
  (110 tests), including the real-CLI contract and sidecar-shutdown pins.
• pnpm run typecheck clean. No Rust touched.

References

• Base-gate failure attribution: Cloud Run job freshell-vitest-31cb9ad69f5b-9qozeu
  (2026-10-29T23:05Z), shard stderr shows the unhandled
  ENOTEMPTY rmdir /tmp/freshell-claude-perm-shapes-* crash.