[hyoban]

Summary

Follow-up to #45381. In CI even the fastest 10% of API unit tests take ≥18ms, and about two thirds of unit test time is spent in tests under 0.1s. Most of that is autouse setup that every test pays before it runs:

Autouse fixture	Local cost per test	What it did
_sqlite_engine	~0.8ms	Copied the 3.5MB full-schema SQLite template for every test, connected or not
reset_redis_mock	~1.0ms	Recursively reset ~370 child mocks of the shared redis_mock

Changes:

• Lazy SQLite copy: _sqlite_engine copies the template on the engine's first connection (do_connect), guarded by a lock so the copy happens at most once. sqlite_engine still copies up front, because several tests hand sqlite_engine.url to Flask-SQLAlchemy, which opens the file with its own engine.
• Skip the empty Redis reset: redis_mock.reset_mock() runs only when redis_mock.mock_calls is non-empty. mock_calls records calls on every attached child and return value, so an empty list means the recursive reset has nothing to clear. The explicit per-method resets and return_value assignments still run on every test.
• No fixed 0.5s waits in teardown: seven local test HTTP servers ran serve_forever() with the default 0.5s poll interval, so each shutdown() waited up to half a second. They now poll every 10ms, as tests/helpers/unit/model_provider_fixtures.py already did. In CI, 82 tests sat at ~0.5s for ~44s in total.
• One source scan in test_enums: two tests each walked the whole api/ tree (including .venv) and parsed every production module. The walk now prunes tests and .venv, and both tests share one cached parse.
• One app per controller test module: five human input and data source controller test modules (test_contact, test_human_input_form_compatibility, test_human_input_v2_form_read, test_data_source, test_human_input_channel_wire) built a fresh app per test and registered the whole console blueprint, console namespace, or service API namespace on it. That took a few hundred milliseconds per test, and each fresh ExternalApi also made the error contract check rebuild its OpenAPI document. Each module now builds its app once; per-test state (login user, scenario, SQLite engine binding, application services) stays in function fixtures via monkeypatch.setitem, so it is undone after each test. No test in these modules mutates the app itself.
• Freeze collection-time objects out of GC: collection imports the whole application, so every automatic full collection during the run rescanned those long-lived objects and stalled whichever test triggered it (a cluster of slack tests sat at exactly 0.25s). tests/conftest.py now runs gc.collect() and gc.freeze() in pytest_collection_finish, which runs in every xdist worker. Objects created by tests are still collected normally.
• Shared swagger exports: 11 read-only tests in test_generate_swagger_specs.py share one module-scoped generate_specs output instead of each exporting every surface (~4s each in CI). Tests of generation itself (idempotency, live payload match, subprocess exports) still generate on their own.

Local measurements (-n 8 unless noted, no coverage):

• 2000 no-op tests (-p no:xdist): 11.2s → 5.5s, about 4.3ms → 1.4ms of overhead per test.
• tests/unit_tests/services (10.4k tests): 80–89s → 69–78s.
• The six files with local HTTP servers (-p no:xdist): 55.0s → 11.7s.
• tests/unit_tests/models/test_enums.py: 6.6s → 4.1s.
• The five controller modules (-p no:xdist, 288 tests): 50.1s → 10.3s.
• Controller suite (9071 tests, -n 8): 59–60s → 50s.
• gc.freeze: human_input_v2 unit tests (-p no:xdist) 59s → 54–55s; tests/unit_tests/services (-n 8) 75–78s → 69–74s.

CI results

Baseline: the five merge-queue runs on main right after #45381 (median job time). Each commit of this PR ran once. Job times vary by about ±30s between runners, so each change is also measured by the junit time of the tests it touches (summed across workers), which is far less noisy.

	API Controller job	API Unit jobs (shard 1 / 2)
main baseline (median)	405s	448s
+ per-test fixtures, shared swagger exports	367s	419s / 397s
+ no fixed waits, single enum scan	383s	427s / 429s
+ one app per controller module	356s	426s / 394s
+ gc.freeze (this PR)	318s (−21%)	398s / 404s (−10%)

Per change, junit time of the affected tests:

Change	Tests measured	Baseline	After
One app per controller module	the five controller modules	152–166s	33–38s
No fixed 0.5s waits	modules with local HTTP servers	61–63s	12–13s
Shared swagger exports	test_generate_swagger_specs	78–79s	62–66s
gc.freeze	whole controller suite	469s (previous commit)	420s
gc.freeze	whole unit suite	1049s (previous commit)	1027s
Lazy SQLite copy, skipped Redis reset	unit tests outside the groups above	1037–1044s	926–1035s (within runner variance)
Single enum source scan	test_enums	11–12s	8–11s (negligible)

Collection (~55–60s per unit shard, ~25s for controllers) is unchanged; it is dominated by importing the application in every xdist worker.

Checklist

• I've verified the change and added or updated tests where meaningful regression risk justifies coverage.
• I ran make lint && make type-check (backend) and vp staged (frontend) to appease the lint gods

From Claude Code

[github-actions]

Pyrefly Type Coverage

Metric	Base	PR	Delta
Type coverage	83.07%	83.08%	+0.00%
Strict coverage	82.73%	82.73%	+0.00%
Typed symbols	85,682	85,697	+15
Untyped symbols	17,529	17,528	-1
Modules	4373	4373	0

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 89.79%. Comparing base (24b4275) to head (b481a38).
⚠️ Report is 6 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main   #45388   +/-   ##
=======================================
  Coverage   89.79%   89.79%           
=======================================
  Files        6239     6239           
  Lines      390966   390966           
  Branches    79305    79305           
=======================================
+ Hits       351075   351086   +11     
+ Misses      33791    33782    -9     
+ Partials     6100     6098    -2     

Flag	Coverage Δ
api	89.63% <ø> (+<0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.