## Problem Tests that call `handlers.NewApp()` or `newTestEnv()` never shut down the `Broadcaster` created by `configureEvents()`. Each leaked `Broadcaster.run()` goroutine blocks on a `select{}` for the lifetime of the test binary. These accumulate across tests and eventually cause the 35-minute CI timeout. The flaky test name (`database-api-load-balancing-discovery-pg_curr_version-1_25-valkey::TestMain`) is misleading. The timeout fires while unrelated tests are running because the goroutine leak is cumulative across the entire test binary. ## Root cause 1. `NewApp` did not clean up the event sink if an error occurred after `configureEvents()`. 2. Integration tests that create an `App` (via `NewApp` or `newTestEnv`) did not call `GracefulShutdown` on cleanup. ## Fix 1. **Close Broadcaster on `NewApp` error** — add a deferred cleanup in `NewApp` so that if any step after `configureEvents()` fails, `GracefulShutdown` closes the event sink. 2. **Refactor `NewApp` error cleanup** — replace the `eventsSinkConfigured` bool flag with a named error return and a defer that calls `GracefulShutdown` on failure, avoiding duplicated sink-close logic. 3. **Add `GracefulShutdown` to test cleanup** — all `app_integration_test.go` tests that successfully create an `App` now register `GracefulShutdown` via `t.Cleanup`. 4. **Add missing `Cleanup` calls** — several tests in `api_integration_test.go` create a second `newTestEnv` (typically with auth options) without calling `Cleanup`, leaking a `Broadcaster`. These now call `env.Cleanup(t)`. 5. **Use `skipDatabaseNotEnabled` helper** — lockfile tests replaced manual `REGISTRY_DATABASE_ENABLED` env var checks with the existing helper for consistency.