Cross-fit clinicalWinRatio; characterize & document small-sample behavior

[blind-contours]

What

Two related improvements to the experimental clinicalWinRatio():

1. Cross-fitting (n.folds, default 5). The transition and censoring hazards
   are now fit out-of-fold, so each subject's influence-function contribution uses
   nuisances trained without them. This removes the empirical-process (over-fitting)
   term and gives honest inference when SL.library contains flexible learners
   (random forests, HAL, penalized regression). n.folds = 1 keeps the faster
   in-sample fits for simple parametric learners. Verified the estimate is
   unchanged at large n (WR 1.590 vs brute-force truth 1.585).

2. Small-sample behavior — characterized and documented honestly. The win
   ratio is a ratio, so it is mildly biased and anti-conservative at small n.
   This is a well-known finite-sample property of the win ratio (it affects the
   unadjusted Pocock win ratio too), not a defect of this estimator or its
   nuisances. A null simulation (both arms identical, true WR = 1) gives:

   n / arm	mean WR	95% coverage	type-I
   400	0.99	0.93–0.94	0.06–0.07
   800	1.00	0.97	0.03
   1600	1.00	0.97	0.03

   Bias ~1% at 400/arm, gone by 800; coverage nominal by ~800/arm.

   Crucially, cross-fitting does not change this — confirmed empirically
   (cross-fitted vs in-sample at n=400 gave identical coverage). It is a property
   of the win-ratio functional, not the empirical-process term, so the methodological
   reflex of "under-coverage → cross-fit" does not apply here.

Docs

• Function help: new @section Small-sample behavior; corrected @param n.folds
  (it no longer claims to fix the small-n issue).
• "Win ratios for trialists" vignette: new A note on small trials subsection +
  coverage-vs-n figure.
• NEWS.md updated.
• scripts/make-clinical-wr-smalln.R reproduces the sweep.

🤖 Generated with Claude Code