Read docs/testing.md#integration-databases for isolation and contamination rules.
Run a focused file through the launcher, not bare Vitest:
pnpm test:integration path/to/file.integration.test.ts
pnpm test:integration --preflight path/to/file.integration.test.ts
Use Docker, or supply both caller-owned database URLs named in testing.md. Inspect the first failure before downstream contaminated cases. Check bundle manifest diagnostics and use their rebuild command; do not bypass stale evidence or fall back to tsx for recovery forks. Stop a broad run if unrelated cases fail shared setup. Observe lock sets rather than backend PIDs across transaction retries.
Keep enforcing mode for acceptance. For diagnosis only:
OPENSTEER_LOCK_ORDER_MODE=record pnpm test:integration path/to/file.integration.test.ts
For an eligible complete unfiltered attempt, diagnose once at its original SHA:
pnpm integration:rerun /absolute/run-record/attempt
Read the JSON decision; exit 1 is intentional even when failures disappear.
See scripts/run-record.ts for attempt storage and contract requirements. Do not
use a rerun as a qualification receipt or rerun until green.
After SIGKILL or a host crash, identify the printed run-owned container before
removing it: docker rm --force --volumes <container-id>. Never stop caller-owned
servers or another run's resources. Report the first failure, command, SHA and
cleanup outcome.