Test suite

pyproject.toml sets addopts = -m 'not medium and not hpc', so the default invocation runs only the fast, hermetic unit tests. medium and hpc tests are opt-in — pass an explicit -m to include them (this overrides addopts, it does not combine with it).

# Fast unit tests only (no binary, no SLURM — ~20 s). This is the default.
uv run python -m pytest tests/

# Include numerical verification tests (runs the real binary at fidelity 3, ~2 min each)
uv run python -m pytest tests/ -m medium

# Full suite including SLURM integration tests (submits real jobs, needs cluster allocation)
uv run python -m pytest tests/ -m "medium or hpc"

Run medium/hpc via CI, not locally

The medium and hpc marked tests are wired to run on GitHub Actions, not routinely on OSCAR — that's the whole point of the CI setup below. Push your change and let CI validate it; only run these locally when you're specifically debugging something that requires comparing local OSCAR behavior against CI (e.g. isolating a toolchain-specific difference).

GitHub Actions (.github/workflows/ci.yml) runs the fast suite on every push/PR, and the medium numerical-verification suite (real fidelity-3 CFD runs) on pushes to main. It builds Basilisk from the official source tarball into a cached basilisk/ directory rather than relying on any OSCAR-specific path — hpc-marked tests (real SLURM submission) never run in CI, since they need an actual cluster allocation.

Known open issue: the medium job's numerical-verification step is currently non-blocking (continue-on-error). test_interface_oscillates_at_rocking_frequency passes reproducibly on OSCAR — including against the exact same Basilisk tarball CI uses, built fresh there — but fails on the CI runner's Ubuntu/gcc toolchain (6.5% spectral power in the expected band vs. a 20% threshold, typical correct value ~37%; not a borderline rounding difference). Basilisk source version and simple run-to-run nondeterminism have both been ruled out as the cause; not yet root-caused.