fix(tests): apply integration marker and timeout defaults per directory

[pofallon]

Closes #163.

The flakiness

Three integration tests intermittently timed out under -n auto and passed in isolation. Not a hang — insufficient headroom against the global 30s default in pyproject.toml. Serial runtimes on an idle 8-core box:

Test	Serial	Headroom
test_scenario_2_rollback_on_gate_failure	19.9s	10.1s
test_scenario_1_clean_retroactive_application	9.2s	20.8s
test_record_commit_stamp_flow	6.2s	23.8s

test_scenario_2 needed only a ~1.5x contention slowdown to breach the budget. These tests shell out to real jj/bd/git, so under seven competing xdist workers that's routine.

Where they bite: local full-suite runs, not CI. GitHub Actions installs jj but not bd, and both modules pytest.skip(allow_module_level=True) without their tooling — so they never run here. make ci on a dev box is the documented pre-push gate (CLAUDE.md), which is where they failed. (I had this wrong in the first draft of this PR and in the conftest docstring; corrected in 3ef9746.)

A second bug found while fixing it

tests/integration/conftest.py applied its marker via a module-level pytestmark — which pytest ignores in a conftest; it only honours pytestmark in test modules. So the integration marker only ever covered the four files that happened to declare it themselves, and -m integration silently under-collected the other thirteen. Anyone running -m integration or -m "not integration" has been getting wrong results.

Both are the same root cause — the directory had no working mechanism for applying defaults — so both are fixed by the same replacement: a pytest_collection_modifyitems hook that applies the marker and a default timeout, filtered to this directory and skipped wherever a test already said otherwise.

-m integration now collects 113 tests.

Sizing

120s, chosen from the measurements above rather than picked round: it absorbs a ~6x contention slowdown over the slowest observed test while still failing a genuinely hung test in bounded time. Explicit per-test overrides already in the tree (60s, 90s, 150s) still win — verified.

Not doing

The issue also suggested adding @pytest.mark.slow to integration tests over ~5s "so make test-fast excludes them consistently". Checking the Makefile, test-fast runs tests/unit/ by path — integration tests are already excluded regardless of markers. That suggestion was unfounded and is skipped.

Verification

make ci-coverage (the exact CI target): 4388 passed, 1 xfailed, coverage 85.45%.

New tests/unit/test_integration_suite_defaults.py covers the conftest itself, since both bugs here were the invisible kind — the suite passed the whole time, just with the wrong marks and too little headroom. It asserts marker coverage, that no integration test falls through to the 30s global, that explicit overrides win, and that nothing leaks into the unit suite.

Two things I checked rather than assumed:

• The tests genuinely catch the regression — restored the old conftest and watched 3 of them fail.
• They're environment-independent — first version named specific tests and broke in CI for the collection reason above; rewritten as invariants over whatever collects, and verified by running with bd removed from PATH (7 passed, 1 skipped with a stated reason).

🤖 Generated with Claude Code

https://claude.ai/code/session_012av72R5SPCc84chsUNjhLC