[haksungjang]

test (backend-integration) took 22 to 37 minutes and was the critical path of every pull request. The integration suite now runs as three jobs, test (backend-integration, shard K), each with its own Postgres and Redis service containers. pytest --shard K/N (tests/_shard.py) splits by whole file, longest first onto the least loaded shard, using weights in tests/integration/shard_durations.json; files keep their collection order inside a shard. I chose a file-level split over pytest-split because it keeps module fixtures and in-file ordering intact, depends only on file paths (parametrised ids need not be stable across processes), and adds no dependency. N is strategy.job-total. Three is enough: the tests add up to about 31 minutes, so a shard is about 10 minutes of tests plus 2 of setup, which is below the unit leg (12 to 21 minutes), now the critical path.

The required check keeps its name. It belongs to an aggregate job (needs: the shards, if: always()) that is red unless the shards' result is success and python -m tests._shard verify confirms that the shards ran every collected file exactly once with the same collection. It also uploads merged durations for refreshing the weights. coverage-gate now needs test and the aggregate, and combines .coverage.backend-unit with .coverage.backend-integration-K-of-N after checking that every shard's file is present. The combined total matches main (91%, 3131 missed statements vs 3135), and 2724 tests pass with 16 skipped (main: 2722 and 16, plus the two tests added here).

For #591, CI passes --timeout=300 --timeout-method=thread (pytest-timeout 2.4.0, hashed in requirements-dev.lock). The slowest test takes 70 s. The thread method fires even when the main thread is blocked in C code, and a conftest hook puts the test id in the main thread's name, so the stack dump is headed MainThread running <test id>, including for hangs in fixtures. This PR does not look for the lock itself.

Verification on this branch, all in throwaway commits that have since been dropped: a test hung on a lock was killed after 20 s with its stack and id printed, and the aggregate went red (failure). A shard cancelled by a one-minute timeout-minutes also turned the aggregate red (cancelled) while the other shards passed. Breaking the audit export's 413 mapping made the new 413 test fail.

Tests that relied on state left by other files, found by running the partitions:

• test_seed_verify_baseline_webhook_secrets.py failed in a shard where test_scan_scheduler.py (whose TRUNCATE ... CASCADE reaches projects) ran before it with no fresh seed in between. It now drops and reseeds the demo itself.
• Two code paths were covered only by accumulated rows: the audit CSV export's 413 (the unfiltered export in test_demo_admin_scope_db.py passed 100,000 audit rows) and the catalog refresh skipping a clean row. Each now has a test that reaches it on purpose.

Closes #575
Refs #591