feat: wire #120's aspect-ratio buckets into recommend_config()

[jc-macdonald]

Summary

Closes the last piece of #116: recommend_config() now actually returns an aspect-ratio-aware recommendation for genomics/ecological-scale data, instead of just warning that the standard buckets don't cover it.

Change

_bucket() now takes n (not just p), routing extreme-aspect-ratio regimes to one of #120's four new buckets:

bucket	trigger	validated at
wide_extreme	p/n > 10	bulk_rnaseq: n=30, p=2000, p/n=66.7
wide_moderate	p > 200 or p/n > 2.0	microbiome (p/n=6) + single_cell (p=500)
tall_extreme	n/p > 50	ecological: n=3000, p=30, n/p=100
tall_moderate	n/p > 15	cultural: n=1000, p=50, n/p=20

Falls back to smallp/trans/large otherwise, unchanged. recommend_config() still warns when it returns one of the four new buckets (each validated at a single example regime, not a dense grid like the original three) -- but now it's a real, empirically better recommendation rather than a warning with nothing behind it.

Also closes a gap in #117's warning: it only checked p/n (the wide direction), never n/p -- the original Option A grid's own tested n/p ceiling (15.0) is much tighter than the p/n one (2.0), so n >> p data was silently extrapolating too.

Two real bugs found while wiring this in

Neither was exercised by #120's own validation script, which called VBPCASimulator.generate() directly rather than round-tripping through recommend_config():

1. criterion_order/active_criteria were stored as trade-study preset names ("cost_angle", "all"), not the list/dict VBPCA's constructor actually accepts -- VBPCA(**recommend_config(n=30, p=2000)) raised TypeError immediately. Resolved to real values; active_criteria renamed to convergence_criteria (VBPCA's actual kwarg).
2. rmsstop_window/rmsstop_atol/rmsstop_rtol were three separate keys; VBPCA takes one compound rmsstop=[window, atol, rtol] kwarg.

Found a third issue alongside (same root cause): the shallow dict(_BUCKET_CONFIGS[bucket]) copy left the new buckets' nested list/dict values as shared references across calls -- mutating a returned config would corrupt the module-level bucket for every future caller. Now a deep copy.

Test plan

• just ci passes, defaults.py at 100% coverage
• New tests: all four new buckets actually VBPCA(**cfg).fit() without error (regression test for both bugs above), mutation-safety (regression test for the third), bucket routing at each threshold (wide_extreme/wide_moderate/tall_extreme/tall_moderate), boundary values stay in the non-extreme bucket, extreme configs differ from their moderate siblings

🤖 Generated with Claude Code