• new random_state parameter on all estimators (FunctionEstimator, DensityEstimator, DimensionalityEstimator, TimeSensitiveDensityEstimator) — controls the seed used for k-means landmark selection and PyNNDescent nearest-neighbor index initialization. Defaults to 42, preserving prior behavior. Previously, only the module-level compute_landmarks / compute_nn_distances functions exposed the seed; estimators silently used the hardcoded default.
• bugfix: GaussianProcessType now inherits from (str, Enum) so string literals (e.g. "fixed") compare equal to enum members (GaussianProcessType.FIXED). Previously, public functions like compute_landmarks checked gp_type == GaussianProcessType.FIXED directly, which silently returned None when callers passed the documented string literal — bypassing the intended fall-through behavior (returning all datapoints as landmarks when n_landmarks >= n_samples). Affected any caller using compute_landmarks(..., gp_type="fixed") directly without going through an estimator.
• downgrade compute_landmarks log message from WARNING to INFO when gp_type="fixed" is requested with n_landmarks >= n_samples. This is a documented, intentional fallback ("use every cell as a landmark") and callers who pin a high n_landmarks for cross-condition consistency should not be pestered with a warning about expected behavior.