Nothing prevents a test module from mutating global state at import — add a collection-time guard

[lmoresi]

Found via #567 (PR #574). tests/test_0741_expression_arithmetic_units.py reached development as a converted script with no test functions at all, whose entire body ran at module level — including uw.Model().set_reference_quantities(length=2900 km, time=1 Myr), which switches the units system on process-wide, during pytest collection.

The consequence was subtle and expensive: a function-scoped reset cannot protect a module-scoped fixture (pytest builds higher-scoped fixtures first), so any test file whose module-scoped fixture ran first in its worker built its mesh under dimensional coordinates. test_0761_point_locator's field was sampled 2.9 million times too far apart — worst error 1.6747e+00 against a clean 2.9e-15. It only showed up under pytest-xdist because --dist loadfile decides which file is first in a process, which is why it looked like a parallelism bug and blocked enabling CI parallelism for days.

PR #574 fixes the instance (module-scoped autouse reset, and 0741 rewritten as real tests) and pins the consequence with test_0742. It does not pin the practice.

The ask

A collection-time guard that fails when a test module mutates global state at import. Candidates, cheapest first:

• a conftest.py hook that snapshots the relevant globals (the default uw.Model, the units/scaling registry, any process-wide PETSc options UW3 sets) before and after collection, and fails with the offending module named;
• a check that every tests/test_*.py contains at least one test function — a file with none is either dead or a script, and both are worth knowing about (this alone would have caught 0741);
• the same snapshot applied per-module during collection, which localises the culprit rather than reporting a global diff.

Related instances of the same family, so this is not a one-off: tests/test_0050_utils.py runs two Stokes solves at module scope (#505), which is why whole-directory collection hangs and why pytest -m "level_1 and tier_a" without a path dies silently. Both files predate any of today's work.

Related: #567/#574 (the instance), #505 (the sibling), #570 (the coverage audit — a file with no tests is invisible to it too).

Underworld development team with AI support from Claude Code

[lmoresi]

Implemented by PR #588, merged to development as 2ab2484.

tests/conftest.py fingerprints the uw_object count, the active model's reference quantities and the strict-units flag around each module's collection, and reports a difference as a collection error against that module. Eighteen modules already offend and are exempted by a ratchet list, tracked in #587 with test_0601_mesh_vector_calc.py named first as the only one that moves the units state.

One thing worth recording for anyone writing a similar guard: aborting the session from a collection hook does not work under xdist. Both pytest.UsageError from pytest_collection_finish and pytest.exit(returncode=4) leave a worker part-collected, and the controller reports INTERNALERROR> assert not crashitem at return code 3 — the mode CI runs in. A pytest_make_collect_report wrapper that marks the offending module's report failed is carried by both runners.

Auto-close does not fire because PRs target development, not the default branch.