Detect Null Drift

Task

Implement detect_null_drift(reference_missing: np.ndarray, current_missing: np.ndarray, threshold: float) -> np.ndarray. Repair the buggy null-drift detector so it flags feature columns whose null rate changed by more than an absolute threshold between a reference dataset and a current dataset.

Requirements

• reference_missing and current_missing are rectangular row-major boolean masks with the same number of feature columns whenever both are non-empty.
• True means the corresponding value is null and False means it is observed.
• For each feature, compute its null rate independently in each mask as true_count / row_count.
• Define every feature's null rate in an empty dataset as 0.0; infer the feature count from the non-empty mask. If both masks are empty, return an empty output.
• Flag feature j only when abs(current_rate[j] - reference_rate[j]) > threshold; equality with the threshold is not flagged.
• threshold is between 0.0 and 1.0 inclusive.
• Return flagged zero-based feature indices in increasing order, and do not mutate either input.

Example

reference_missing = np.array([
    [False, True],
    [False, False],
    [False, False],
    [False, False],
])
current_missing = np.array([
    [True, True],
    [True, False],
    [False, True],
    [True, False],
])
detect_null_drift(reference_missing, current_missing, 0.25)
# array([0])