Fix multivariate default n_best scaling and mirror R's dependence tryCatch chain

[OVVO-Financial]

Changes

Two fixes to keep order=None behavior exact against R NNS, verified with a live R 4.3.3 + NNS 13.1 build:

1. _resolve_n_best: scale default n_best by sqrt(n_obs), not sqrt(n_features) (multivariate_regression.py)

The default n_best = max(1, floor((1-dependence)*sqrt(x.shape[1]))) mirrored R's Multivariate_Regression.R:129, where n is shadowed by ncol(original.IVs). With sqrt(#features) ≤ 2.24, the default collapsed to n_best = 1 for any dependence above ~0.1, making order=None predictions identical to order="max" (pure 1-NN) on continuous multivariate data. Now uses sqrt(x.shape[0]), consistent with the sqrt(n_obs) grid nns_stack already cross-validates over. Companion R PR: OVVO-Financial/NNS#39.

2. _regression_dependence: mirror R's tryCatch fallback around NNS.dep (regression.py)

R wraps the NNS.dep call in tryCatch(..., error = function(e) .1) before averaging with the copula measure; the port only guarded the copula half. The order=NULL dependence chain is now a literal transcription of the R block.

Verification

• order=None ≠ order="max": max abs diff 2.539 on 200×5 shared test data, identical value in both languages; previously bit-identical predictions.
• Cross-language sweep (9 datasets, n=50–500, spanning n<100 halving, weak/strong dependence, dependence==1): NNS.dep and NNS.copula components bit-identical (≤6e-16), composed dependence within 1 ulp, all 9/9 resolved dep.reduced.order values equal for order=None and integer overrides.
• Integer order parity: univariate regression point counts (3/5/11/200/19 for order 1/2/4/max/None) and multivariate estimate diffs (4.74844/2.92378/0/0 vs "max") match live R digit-for-digit.
• Point-estimate parity post-fix: 59/60 rows bit-exact vs patched R; the remaining row is the out-of-support gradient-extension path (diff 3.4e-3).
• Full suite: 2226 passed, 12 skipped — no recorded R-parity case affected (all cached multivariate fixtures resolve to n_best=1 under both formulas).
• Benchmark effect (5 seeds × 3 regimes, n_train=500): default nns_reg RMSE improves iid 1.39→1.12, support_shift 2.22→2.07, prior_trap 1.58→1.25.

🤖 Generated with Claude Code

https://claude.ai/code/session_01BPbZvtDw4h2XJo9w5hw57h

---

Generated by Claude Code