Fix multivariate default n.best to scale with observations, not features

The default n.best in nns_m_reg used floor((1-dependence)*sqrt(ncol)),
mirroring R's Multivariate_Regression.R:129 where n is shadowed by
ncol(original.IVs). With sqrt(#features) <= 2.24 for typical widths,
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.

Use sqrt(nrow) instead, consistent with the sqrt(n_obs) grid nns_stack
already cross-validates over. Verified against a live patched R NNS
13.1 build on shared data: order=None now differs from order="max"
identically in both languages (59/60 test rows bit-exact, remaining row
is the out-of-support gradient-extension path). Full test suite passes;
no recorded R-parity case is affected because all cached multivariate
fixtures resolve to n.best = 1 under both formulas.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BPbZvtDw4h2XJo9w5hw57h

Author: claude

src/nns/multivariate_regression.py

@@ -385,7 +385,7 @@ def _resolve_n_best(
     if n_best is not None:
         return max(1, int(n_best))
     dependence = _copula_matrix(np.column_stack((x, y)))
-    return max(1, math.floor((1.0 - dependence) * math.sqrt(x.shape[1])))
+    return max(1, math.floor((1.0 - dependence) * math.sqrt(x.shape[0])))
 
 
 def _k_as_count(k: KValue, row_count: int) -> int: