Four integration tests shut the API down and immediately count the
database's rows in pg_stat_activity, expecting zero. PostgreSQL ends a
backend asynchronously: pool.end() resolves once the client side has
closed, and the server can still list that backend for a moment. On a
loaded runner an immediate count saw one, failing migration-smoke on a
docs-only PR (Ekklesia CI run #238, module-database.integration.ts:764,
"expected 1 to be +0").
Reproduced against postgres:18.4-alpine3.24 limited to 0.05 CPU: after
pool.end() resolved, the backend was still listed in 7 of 300 trials,
and gone about a millisecond later.
The checks now wait up to 3 s for the count to reach zero. A connection
that is genuinely left open is still caught: an idle pooled client
stays listed for pg-pool's 10 s idle timeout, far past the deadline.
Co-Authored-By: Claude Opus 5.5 <[email protected]>