Testing

dough.testing ships shared pytest fixtures for downstream dough-based packages.

Scope

dough.testing provides cross-package test helpers only. It does not ship test data, assertions, or anything beyond regression-snapshotting fixtures.

Fixtures

Two fixtures are exposed via dough.testing.plugin. Both accept (data, max_number=None). Pass max_number=N when the regression snapshot would otherwise be huge — the fixture thins lists/arrays down to N equally spaced elements before the diff.

json_serializer

Factory that makes arbitrary data JSON-serializable for regression snapshots. Rounds floats to 5 digits, coerces numpy.ndarray via .tolist() (complex arrays split into [real, imag] pairs), and optionally subsamples lists/arrays to max_number equally spaced elements. Raises TypeError on unsupported types.

robust_data_regression_check

Wraps pytest-regressions' data_regression.check, piping the data through json_serializer first. Accepts the same max_number kwarg for subsampling.

Registration

The plugin is opt-in. Downstream packages activate it by adding one line to their top-level conftest.py:

pytest_plugins = ["dough.testing.plugin"]

This is intentional. Registering dough.testing.plugin as a pytest11 entry point would auto-load the fixtures in any environment that happens to have dough installed — even environments that use dough only as a runtime dependency and never call into the testing layer. Explicit registration keeps the "where do these fixtures come from?" trail one grep away, and avoids injecting pytest fixtures into unrelated test suites.

Dependency policy

dough itself declares pytest-regressions and numpy under the [tests] extra. numpy is imported lazily inside dough.testing._serialize so that dough's runtime surface stays numpy-free — the fixture still works on pure-Python data without numpy installed, mirroring the lazy-import rule already used for ase/pymatgen/aiida-core in the converters.