tests/test_audit_split_hygiene.py

from ylff.services.audit.audit_runner import run_audit
from ylff.services.audit.models import ExternalReferenceMeasurement, OperatingRegime


def _m(capture_id: str, *, err: float, sigma: float):
    return ExternalReferenceMeasurement(
        capture_id=capture_id,
        regime=OperatingRegime.INDOOR_CONSTRAINED,
        measurement_type="tag_to_tag",
        d_pred=10.0 + float(err),
        d_star=10.0,
        sigma_d=float(sigma),
    )


def test_run_audit_splits_by_capture_id_scene_disjoint():
    # Two scenes, multiple measurements each. Ensure split happens at scene granularity.
    ms = []
    for _ in range(10):
        ms.append(_m("scene_a", err=0.1, sigma=0.2))
        ms.append(_m("scene_b", err=0.1, sigma=0.2))

    res = run_audit(ms, calibrate=True, calibration_split_fraction=0.5)
    split = res.summary.get("split", {})
    assert split.get("mode") in ("scene_stratified_by_regime", "row_fallback")
    if split.get("mode") == "scene_stratified_by_regime":
        assert split["num_scenes_total"] == 2
        assert split["num_scenes_cal"] + split["num_scenes_audit"] == 2


def test_dataset_level_gate_2b_enforced_only_when_calibrated():
    # If calibration set ends up empty (e.g., only one scene and rounding), Gate 2b is skipped.
    ms = [_m("scene_only", err=0.2, sigma=0.05) for _ in range(20)]
    res = run_audit(ms, calibrate=True, calibration_split_fraction=0.5)
    gates = {g.name: g for g in res.gates}
    assert "gate_2b_dataset_level_coverage" in gates
    # Skipped OR enforced is fine; what we require is the gate exists and is reported.
    assert isinstance(gates["gate_2b_dataset_level_coverage"].details, dict)