Robust feature geometry: similarity-kernel diagnostic
=====================================================
reference fit samples:        800
reference anchor samples:     400
held-out reference samples:   400
shifted/new samples:          400
feature dimension:            20
reference contamination:      12%
contaminated reference count: 96
shift direction:              low-variance feature axis
leverage strength:            18.0
new-distribution shift:       4.0
RBF length scale:             6.325
top-k similarity summary:     10

Method                              same-ref top-k   new-ref top-k   contrast
Empirical clean references                0.784          0.680      0.104
Empirical contaminated                    0.798          0.793      0.005
Robust FastMCD clean refs                 0.776          0.670      0.106
Robust FastMCD contaminated               0.777          0.672      0.105

Diagnostic summary
------------------
empirical contamination contrast drop: +0.099
robust contamination contrast drop:    +0.001
robust contrast gain over contaminated empirical: +0.099
pattern: empirical RBF geometry loses contrast under leverage contamination, while robust feature geometry preserves kernel sensitivity to the shifted distribution.

Interpretation
--------------
The RBF kernel is computed from distances induced by a fitted feature geometry.
When empirical covariance is contaminated in a sensitive low-variance direction, shifted features can remain spuriously similar to reference features.
A robust scatter estimate gives a more stable feature metric, so the kernel similarity to shifted features drops as expected.
