[bokelley]

Summary

• pause the provider-intent test wrapper only after PostgreSQL acknowledges COMMIT, when its target-plan row lock is definitively released
• start and verify standalone recovery inside that deterministic post-commit window
• release and settle both deferred operations during cleanup so assertion failures cannot strand pool clients

This changes only the integration test. Production timeout, retry, fail-closed, ledger, and migration behavior are unchanged.

Failure evidence

The unchanged baseline test failed on PRs #7446, #7463, and #7475 when the unsynchronized recovery race returned persistence_uncertain. Local baseline stress reproduced the line-414 failure in 15 of 16 concurrent runs.

Verification

Exact reviewed head: fd82d7835dbc9ed57ccfc7626f248ff2212c9235.

• focused requested PostgreSQL test: passed
• stress: 16/16 concurrent runs passed against isolated PostgreSQL databases
• pre-commit: 71 root unit files / 1,063 tests passed; dynamic-import, format-identity, and call-API guards passed
• npm run typecheck: passed
• git diff --check: passed
• independent code review: no findings
• GitHub CI terminal: 28 successful checks plus expected skips/status; zero pending checks and zero failures
• all four PostgreSQL integration shards and their aggregate: passed
• CodeQL review comments: none

The complete 36-test PostgreSQL file was also exercised locally. The modified test passed, but the VM consistently hit a separate unchanged timeout race in the preceding runs provider-intent recovery independently of a concurrent standalone recovery test at line 372 (35/36 passed). That test fails before the modified case and was not broadened into this narrowly scoped fix. GitHub's clean PostgreSQL integration matrix passed all four shards at the exact head above.

Remaining gate

Human maintainer review is requested from @adcontextprotocol/maintainers. Keep this PR Draft with auto-merge disabled; do not merge as part of this handoff.

---

Open workspace in Conductor

[bokelley]

Closing as superseded by #7514, which provides the broader deterministic PostgreSQL barrier fix for this recurring ledger-test race.