Mixle Core testing

Document ID: CORE-DOC-TESTING-001 Version scope: 0.8.x development Owner: PRJ-CORE

Strategy

Tests are organized by claim, not only by module. A behavior change should cover the successful path, invalid inputs, boundary values, preserved compatibility, and any scientific invariant it claims.

Pytest collects both suffix-named files ending in test.py and prefix-named files beginning with test. Test functions begin with test_, and legacy unittest classes remain supported.

Local tier

Run only directly affected nodes with serial execution and no marker filter. Each local command has a hard 30-second deadline. A timed-out command is evidence of an incomplete local check, not a pass.

Hosted tiers

• fast jobs exercise the commit gate on supported Python and platform entries;

• full jobs exercise non-optional correctness with coverage;

• minimum-version jobs check declared dependency floors;

• clean-wheel jobs check packaging and import behavior;

• optional jobs exercise selected accelerators, connectors, and distributed backends when their dependencies are installed.

The workflow file is .github/workflows/tests.yml. Marker ownership is centralized in mixle/tests/conftest.py.

Evidence mapping

Record exact node IDs or file selections, revision, environment, elapsed time, result, and known omissions. Scientific claims additionally record seed, tolerance, data or fixture identity, and the invariant being checked.

Coverage limits

Passing the base environment does not establish optional-backend parity. Coverage percentage does not establish numerical validity. Stochastic success at one seed does not establish calibration or convergence. Release evidence must combine functional, negative, compatibility, scientific, packaging, and supported-environment results applicable to the changed surface.