[avrabe]

What failed

PR #428 changes only a workflow file, yet its Verification Gate went red on a Rust unit test:

thread 'stats::tests::record_and_read' panicked at loom-core/src/stats.rs:197:9:
assertion `left == right` failed
  left: None
 right: Some(3)

Why

record_and_read writes its own pass name three times and then reads it back. Its comment says the unique name prevents races. It doesn't: take_resets in the same module calls take_revert_summary(), which clears the whole process-global map. When the two tests run on parallel threads, the clear can land between the writes and the read, and the read returns None.

Evidence

	runs of stats::tests --test-threads=8	failures
main	400	1
this PR	2000	0

The counts alone don't prove the fix: one failure in 400 runs is a single observation, not a known rate. The argument rests on the mechanism. Both tests now hold one test-only lock, so the clear and the read can't interleave. The lock recovers from poisoning, so one failing test can't take the other down with it.

Is that the only race on these globals? I checked every reader of the global stats in the workspace:

• take_resets is the only test that clears the revert map.
• record_and_read is the only test that reads it.
• Every other test adds keys through record_revert, and an added key can't disturb a read of a unique key.
• The coverage tests build local CoverageSummary values and never touch the global.

So the lock covers every test pair that can race. cargo clippy --all-targets -D warnings and cargo fmt --check are clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01RZof1M5HMBSYVZPeoT4MEn

[github-actions]

Rivet verification gate

✅ 30/38 passed

	count
Passed	30
Failed	0
Skipped (no steps)	8

Filter: (and (= type "feature") (matches id "^TEST-"))

Failed artifacts

(none)

_{Updated automatically by tools/post_verification_comment.py. Source of truth: artifacts/verification.yaml.}