feat: sensitivity_from_table() -- post-hoc Sobol/Morris from a table (#113) - #118

Stacked on #117

This PR is stacked on feat/114-surrogate-accuracy (#117) -- it uses fit_surrogate()'s new cv_r2 (from #114) as a trust signal for the sensitivity result. Base is set to that branch rather than main; retarget to main once #117 merges (or I'll do it).

Summary

screen() needs a live, re-runnable simulator and a proper Saltelli/Morris sample design -- neither of which an arbitrary already-collected ResultsTable has, so its indices can't be computed retroactively from whatever points happen to already be in a table. That rules out the literal ask in #113 ("reuse whatever Sobol machinery already backs screen()" applied directly to stored data) -- but there's a standard, sound alternative: fit a cheap surrogate over the table, then run screen()'s existing machinery against the surrogate's predict() instead of a fresh, expensive simulator call. Since the surrogate is cheap to query, you get proper variance-based indices without any new evaluations.

Why this actually fixes the motivating case

#113 was filed because VBPCApy's ARD prior strength (hp_va) has a mechanistic U-shaped effect (too weak = no pruning, too strong = over-pruning) that a marginal Spearman correlation shows as near-zero, while a co-varying, purely-monotonic factor (xprobe_fraction) that had nothing to do with the believed-dominant effect showed up as the strongest correlate. The new test test_sobol_detects_nonmonotonic_effect reproduces exactly this shape (y = (a - 0.5)**2 + 0.1*b, a U-shaped and symmetric, b linear) and confirms Sobol S1 through the surrogate correctly ranks a as dominant (importance > 0.5 vs b's remainder) -- the case a marginal correlation would get backwards.

Design

• sensitivity_from_table(results, factors, *, method="sobol", surrogate_method="rf", ...): fits fit_surrogate() on the continuous subset of factors, then calls screen() with the surrogate's .predict as run_fn.
• Only continuous factors are screened (matches screen()'s own contract exactly); non-continuous keys elsewhere in the table's configs are simply ignored by the encoder rather than causing a mismatch.
• Returns TableSensitivity(importance, surrogate_cv_r2) -- surrogate_cv_r2 (fit_regime_surrogate / RegimeSurrogate should report held-out or OOB accuracy #114) is included directly in the result so callers don't have to separately fit and check accuracy before trusting the indices for a given observable.
• Fails fast with screen()'s own "at least one continuous factor" message if none remain, rather than letting a degenerate empty-feature surrogate fit produce a confusing sklearn error.

Test plan

• just ci green locally (ruff format/lint, mypy --strict, coverage 99.51%, 340 passed -- 8 new)
• test_sobol_detects_nonmonotonic_effect: reproduces the exact U-shaped-effect failure mode from Post-hoc Sobol/Morris total-effect sensitivity from an existing ResultsTable #113 and confirms it's fixed
• Morris method also runs end-to-end
• Non-continuous factors in the input list are dropped, not an error
• Empty continuous-factor list raises with screen()'s own message
• A surrogate fit to pure noise triggers the forwarded fit_surrogate low-accuracy warning