test(integration): background work a test starts is cancelled before the next (B-21)

[sshlg]

B-21 — the test_learnings_api StaleDataError (CI on 2026-09-23 and 2026-09-26; passes alone and on rerun).

Mechanism. On SQLite every session shares one physical connection (StaticPool), and work started with spawn_tracked or an in-process task_queue fallback is never awaited by the request that started it. A leftover task that rolls back on that connection can undo the NEXT test's flushed-but-uncommitted insert; that test then commits nothing, finds the object in its identity map without a query, and its UPDATE matches zero rows. That is exactly UPDATE … expected to update 1 row(s); 0 were matched. Not reproduced deterministically — this is the mechanism the evidence fits, and the fix removes the precondition.

Fix. db_session teardown cancels and awaits every such task before the rollback and the truncate (_drain_background_work).

Guard. test_background_work_does_not_outlive_its_test.py: one test leaves a task running, the next asserts it did not inherit it — red without the drain.

Checks: integration 702/702 on SQLite locally (the PostgreSQL job runs here); ruff clean. Only tests/integration changed, so the unit suite is unaffected.

🤖 Generated with Claude Code

[sshlg]

First CI run: the drain cancelled leftover tasks, and a task cancelled mid-statement makes SQLAlchemy invalidate the connection — under StaticPool that swaps in a new empty :memory: DB, so everything after failed with no such table: users (263 failed, 965 errors). Pushed a wait-only drain (30 s, never cancels, names stragglers) plus _assert_schema_survived, which fails at the cause if the shared DB ever loses its tables. Local CI-like run under load average ~200 gave 700/702: test_performance_smoke (fails alone on this machine — timing) and one test_investigate_tenant_isolation FK failure that passes alone; this CI run is the real measurement.