Scope 4.2: Implement _compute_fold_metrics() helper for per-fold RMSE and MAE

[DEVunderdog]

Parent

#93

What to build

Add _compute_fold_metrics(y_true, y_pred) -> tuple[float, float, float] to _numeric_imputer.py. The function takes two numpy arrays — the true values and predicted values from a single k-fold validation fold — and returns (r2, rmse, mae) as a 3-tuple of floats. No model dependency, no Polars dependency, no I/O. Pure numpy/sklearn.

This function replaces the direct _r2_score(y_true, y_pred) call currently inlined inside the k-fold loops in _compute_regression_diagnostic, _compute_knn_diagnostics, and _compute_mice_diagnostics. Issue #227 wires it into those three loops and accumulates rmse_fold and mae_fold lists alongside the existing fold_r2s list.

The minimum-rows guard (n_complete < refit_r2_min_complete_rows) stays on the outer loop in each diagnostic function — _compute_fold_metrics is only called when the guard passes and the fold has variance. It does not handle the None case internally.

Acceptance criteria

• _compute_fold_metrics(y_true, y_pred) exists in _numeric_imputer.py as a private function
• Returns tuple[float, float, float] of (r2, rmse, mae)
• rmse and mae are non-negative for any finite input
• rmse >= mae always holds (Cauchy-Schwarz inequality)
• rmse == 0.0 and mae == 0.0 when y_pred == y_true exactly (zero residuals)
• Unit test: returns a 3-tuple of floats for a simple predictable case (e.g. constant target perfectly predicted)
• Unit test: rmse and mae are non-negative
• Unit test: rmse >= mae holds
• Unit test: rmse == 0.0 and mae == 0.0 when predictions equal targets

Blocked by

• Blocked by Scope 4.1: Add rmse and mae fields to ImputationFitDiagnostic schema #225