fix: warn when recommend_config extrapolates beyond validated p/n range

[jc-macdonald]

Summary

Fixes #116. recommend_config(n, p) buckets purely on p (smallp <=30, trans <=70, large >70), with no upper bound and no use of n at all -- p>70 always maps to the identical "large" bucket config no matter how large p gets or how small n is.

Why this matters

Found while extending pp-eigentest's trade study to genomics-representative data shapes (small cohort, thousands of features -- p/n of 50-1000x, common for bulk RNA-seq/microarray data). The Option A trade study's own regime grid (analysis/trade_study) only covers p up to 200 and p/n up to 2.0 -- computed directly from every TRAINING_REGIMES/VALIDATION_REGIMES entry. Anything past that is silently extrapolated.

Empirically this breaks down: recommend_config(n=30, p=2000) returns the identical config as recommend_config(n=100, p=100), but recovers rank 9 instead of the true rank 5 (vs. correct recovery at the balanced shape).

Change

recommend_config now warns (UserWarning) when p > 200 or p/n > 2.0, mirroring the existing missingness != "auto" warning from #110. This doesn't fix the underlying gap -- that needs the Option A trade study extended with an aspect-ratio axis and re-run, a bigger effort tracked in #116 for later -- but callers are no longer silently hunted an untested extrapolation.

Test plan

• just ci (format, lint, mypy --strict, pytest --cov) passes, defaults.py at 100% coverage
• New tests: warns past p=200, warns past p/n=2.0 even when p itself is small, silent within the validated region (boundary values p=200/p/n=2.0 inclusive)

🤖 Generated with Claude Code