test(eval): pin biased linear CKA behavior across sample regimes

[aurascoper]

Third of three PRs resolving #38. Stacked on #40 — merge #39, then #40, then this. Closes #38.

The estimator is deliberately unchanged

cka() implements biased linear CKA over centered Gram matrices, and it stays that way. Switching to unbiased CKA would alter every prior representation-comparison result and requires a registered analysis migration, not a test cleanup.

The expectation was wrong for that estimator

X = randn(50, 128); Y = randn(50, 256)
cka(X, Y)   # 0.785, asserted < 0.3

The biased estimator's null is inflated when feature dimension approaches or exceeds sample count. At n=50 with d=128 and 256, ~0.78 on independent data is the estimator behaving as designed. Asserting < 0.3 there assumed either an unbiased estimator or n ≫ d.

Two tests, both regimes pinned

regime	shapes	score
n ≫ d	500 × 16 vs 500 × 32	0.042
d ≫ n	50 × 128 vs 50 × 256	0.785

test_cka_independent_is_small_when_samples_exceed_dimensions
test_biased_cka_null_is_inflated_when_dimensions_exceed_samples

The second is explicitly labelled as characterizing the estimator, not as evidence the representations are similar — so a future reader can't mistake a high null score for a finding. Neither the estimator nor the inflation can now change unnoticed.

cka() also gains a docstring recording the inflation and warning against comparing scores across n/d regimes without a matched null or permutation baseline.

The eval gate reaches zero deselections

before #39   83 passed,  3 deselected
after  #39   86 passed,  2 deselected   (suite grew by 2)
after  #40   87 passed,  1 deselected
after  this  89 passed,  0 deselected   (suite grew by 1)

Final counts pending CI. The last --deselect line is removed from the python-contracts job, and its stale "8 of 62" comment is corrected.