feat: migrate sensitivity_report.py to trade_study.sensitivity_from_table()

[jc-macdonald]

Summary

Replaces the hand-rolled marginal Spearman correlation in sensitivity_report.py with trade_study.sensitivity_from_table() (jcm-sci/trade-study#113, built for exactly this) for the 9 continuous VBPCA hyperparameters, per p-bucket.

Why this matters

Unlike Spearman, Sobol-via-surrogate correctly detects non-monotonic (e.g. U-shaped) effects -- exactly the kind defaults.py's own docstring already documents for hp_va: "its tercile means do show a real, nonlinear U-shape a monotonic Spearman correlation understates." That gap was known but this script kept using the method that produces it.

sensitivity_from_table only screens continuous factors (matching screen()'s own contract), so the 4 discrete VBPCA hyperparameters (niter_broadprior/maxiters/patience/rmsstop_window) keep the original marginal-Spearman treatment -- now explicitly labeled "spearman_marginal" in the output rather than mixed in unlabeled alongside the Sobol results.

Verification

Ran against the real, current surrogate_train.json (not synthetic data):

- smallp: top rank_mae factors = xprobe_fraction (sobol_via_surrogate=0.46), niter_broadprior (spearman_marginal=0.17), hp_va (sobol_via_surrogate=0.16)
- trans: top rank_mae factors = xprobe_fraction (sobol_via_surrogate=0.77), hp_v (sobol_via_surrogate=0.06), niter_broadprior (spearman_marginal=0.05)
- large: top rank_mae factors = xprobe_fraction (sobol_via_surrogate=0.75), niter_broadprior (spearman_marginal=0.22), maxiters (spearman_marginal=0.08)

xprobe_fraction dominates in all three buckets via this completely different method -- corroborating the same finding the old marginal-Spearman analysis already reported in defaults.py's docstring ("xprobe_fraction is the strongest, most significant, and most consistent per-bucket predictor of rank_mae"), not hp_va despite it being called the "dominant lever" there. Figure regenerated cleanly (figure_sensitivity_rank_mae.png etc.) with sane-looking bars.

Test plan

• just ci passes
• Ran end-to-end against real local surrogate_train.json, inspected the regenerated figures and JSON output for sanity

🤖 Generated with Claude Code