Robust feature geometry: synthetic OOD example
================================================
Reference features: 600
Feature dimension:   16
Reference leverage contamination: 12%

Method                       AUROC    mean in-score    mean OOD-score
Empirical covariance          0.528            3.856            3.922
Robust FastMCD geometry       1.000            4.029            8.499

Interpretation
--------------
The empirical covariance is fitted on a reference feature set that contains leverage-like contamination.
This inflates variance along the contaminated direction and weakens Mahalanobis-based OOD scores.
Robust feature geometry estimates the central feature-space shape more stably and restores separation in this synthetic example.
