Cross-fit the hit-calling covariance instead of splitting the controls in half

Problem

hit_calling splits the reference rows in half: one half fits the centroid and covariance, the other supplies the
null pool. That costs power twice over — the covariance comes from half the controls, and the null pool is half
as large as it could be — and it is why _hits.py warns that the covariance is singular at roughly twice as many
features as it otherwise would be.

Cross-fitting keeps every control in the estimate while keeping each scored row out of its own fit: for each fold,
fit on the controls outside it and score the ones inside.

Measured

Prototype against #62's half-split, both with #62's pooled permutation null. Real DMSO cells from jump_cells
relabelled as pseudo-treatments, 96 controls, groups of 24, 10 folds, 60 seeds:

Two things follow.

It is a no-op in the normal regime. At 20 features against 48 fit rows the covariance is already comfortable,
and cross-fitting matches the half-split on calibration while losing a little power. Not worth the complexity on
its own.

It pays off where the covariance is starved. As features approach the fit size the half-split loses power
(0.892 to 0.850 from 40 to 60 features) while cross-fitting gains it (0.917 to 0.936). That is the regularization
ceasing to dominate the distances, and it is the same regime _hits.py already warns about.

The blocker

Cross-fitting drifts conservative as features grow — 0.011 at 60 features, five times below nominal. So it trades
the half-split's anti-conservative drift for a conservative one rather than fixing calibration.

The likely cause is in the prototype rather than the method: control rows were scored by a fit excluding their own
fold, treated rows by a fit on all controls. Those two are not quite exchangeable, and the gap widens as the
covariance gets harder to estimate. Resolving that is the real work here — score both sides under the same fits,
or fold the treated rows into the scheme — and it has to be settled before this can ship, since the whole point of
#62 was a null that is exchangeable with what it is compared against.

Proposed fix

• Score treated and control rows under the same fits so the two sides stay exchangeable
• Re-measure calibration and power across features from 10 to 60 against the half-split
• Decide the fold count, or use leave-one-out with a rank-1 downdate of the covariance
• Only adopt it if calibration holds across the range; a conservative test is not an improvement on an
  anti-conservative one, only a different error
• If adopted, revisit the singular-covariance warning, whose threshold is written for the half-split

Until then the guidance in #62's docstring stands: reduce the features with pp.feature_select or score
use_rep="X_pca", which puts a screen in the regime where both schemes agree.

Follows from #60 and #62.

Start by reading the hit-calling implementation in _hits.py and the guidance in #62's docstring. Reproduce the reported cross-fitting measurements, then compare schemes that score treated and control rows under the same fits. Done means a settled fold or downdate design with calibration and power measurements across 10–60 features, and adoption only if calibration holds.