feat: derive VBPCA hyperparameter recommendations for extreme aspect-ratio regimes

[jc-macdonald]

Summary

Adds analysis/trade_study/option_a_aspect_ratio.py, addressing #116's finding that recommend_config's shipped buckets were never validated against genomics-scale (p >> n) or ecological/survey-scale (n >> p) data -- the Option A regime grid only covers p up to 200 and p/n up to 2.0.

Approach

Uses trade_study.recommend_per_regime()/aggregate_bucketed_config() (jcm-sci/trade-study#123, split into two functions in #126) to search VBPCA hyperparameters independently per regime via NSGA-II, then aggregate into named buckets. search/aggregate are separate CLI subcommands rather than one combined call -- re-running the ~2hr NSGA-II search to try a different bucket grouping would have been wasteful, so search runs once and caches every regime's best config; aggregate re-groups from that cache for free.

Result: 2 buckets isn't enough, 4 is

A first 2-bucket grouping (wide vs tall) performed badly for the two most extreme regimes even with n_reps=3 replicate averaging to guard against single-draw luck. A second, finer 4-bucket grouping by aspect-ratio extremity -- giving the extreme regimes their own dedicated bucket instead of averaging them in with milder ones -- fixed it:

regime	wide_tall (2 buckets)	extremity (4 buckets)
bulk_rnaseq (p/n=66.7)	rank_mae=4.0, rmse=2.16	rank_mae=0.0, rmse=0.78
microbiome (p/n=6)	rank_mae=0.0	rank_mae=0.0
single_cell (p/n=1)	rank_mae=0.0	rank_mae=0.0
ecological (n/p=100)	rank_mae=5.0	rank_mae=4.0
cultural (n/p=20)	rank_mae=1.0	rank_mae=1.0

bulk_rnaseq goes from a real failure to perfect recovery; ecological improves substantially (still not perfect, likely because its "best" config was chosen by 3-rep-averaged score during search but validated here on a single draw -- some variance is expected). Both moderate-aspect-ratio regimes and their configs are completely unaffected, and getting worse than 2-bucket wasn't observed anywhere.

Not included

Not wired into defaults.py's shipped _BUCKET_CONFIGS -- whether/how to extend recommend_config() itself with these buckets is a separate decision (bigger blast radius, changes production behavior for all callers) I'd like to check with the maintainer before doing.

Test plan

• just ci passes
• Validated both groupings' configs against all 5 regimes directly via VBPCASimulator/VBPCAScorer (table above)

🤖 Generated with Claude Code