Fix attachment retention test lifecycle race

[baixiaohang]

Summary

• disable automatic attachment-retention scheduling in generic Server test apps while keeping production and explicit opt-in behavior enabled
• track active retention sweeps and wait for them during background-task shutdown before the database pool closes
• add deterministic regressions for the test-helper wiring and the in-flight shutdown boundary

Root cause

Main CI run 31784141632 failed when the per-test TRUNCATE ... CASCADE deadlocked with the startup attachment-retention query. The retention sweep was launched fire-and-forget from backgroundTasks.start(), so a test app could return from ready() while the sweep was still using the same worker database.

The failed me-multi-org test was only the cleanup victim. PR #2334 did not change the test harness, background task, retention service, or failed test.

Validation

Passed locally:

• related Server suites: 4 files / 49 tests with --maxWorkers=2
• background-task and bare-test-app final suites: 2 files / 7 tests
• full Server suite: 293 files / 3441 tests with --maxWorkers=2
• pnpm --filter @first-tree/server typecheck
• changed-file Biome checks
• git diff --check

Mutation checks also proved both regressions are effective:

• removing the in-flight stop() wait changes the settlement race from sweep-pending to stopped
• removing the createTestApp() default-disable wiring starts the mocked retention sweep and fails the helper-level test

[yuezengwu]

Reviewed the lifecycle fix. createTestApp() now disables the retention sweep by default while production remains enabled unless explicitly opted out; active retention sweeps are tracked and awaited before the database pool closes. The helper-level and in-flight shutdown regressions cover both failure boundaries. No schema or core data-structure changes. No blocking findings.