Scope 4.3: Wire _compute_fold_metrics() into Regression, KNN, and MICE diagnostic paths

[DEVunderdog]

Parent

#93

What to build

Extend _compute_regression_diagnostic, _compute_knn_diagnostics, and _compute_mice_diagnostics in _numeric_imputer.py to accumulate rmse and mae alongside r2 inside their existing k-fold CV loops. Replace the current direct _r2_score(y_true, y_pred) call in each fold with _compute_fold_metrics(y_true, y_pred) (from #226), which returns (r2, rmse, mae) for that fold. Average rmse_fold and mae_fold across folds exactly as fold_r2s is averaged today. Populate diagnostic.rmse and diagnostic.mae on the returned ImputationFitDiagnostic.

Regression path: the existing k-fold CV on complete rows of the joint array [col] + feat_cols gains two accumulated lists — fold_rmses and fold_maes. When the minimum-rows guard fires (n_complete < refit_r2_min_complete_rows), all three fields stay None.

KNN path: the existing per-column k-fold CV on the scaled KNN matrix gains the same two lists per column. Inverse-scaling is already applied to y_pred before _r2_score — the same inverse-scaled values feed _compute_fold_metrics.

MICE path: the existing k-fold CV on the intersection of complete rows across all MICE columns gains fold_rmses and fold_maes per column. All MICE columns share the same fold splits; if the intersection is below refit_r2_min_complete_rows, all three fields go None for every column in the block simultaneously.

No changes to scalar strategy paths (Mean, Median, Mode, Constant, Dropped, Passthrough, MNAR) — diagnostic remains None for those.

Acceptance criteria

• _compute_regression_diagnostic accumulates fold_rmses and fold_maes via _compute_fold_metrics() and sets diagnostic.rmse and diagnostic.mae as the mean across folds
• _compute_knn_diagnostics does the same per KNN column, using inverse-scaled predictions
• _compute_mice_diagnostics does the same per MICE column from the shared fold splits; all MICE columns have the same None / non-None status on r2_train, rmse, and mae simultaneously
• diagnostic.rmse and diagnostic.mae are non-negative floats for model-based columns with sufficient complete rows
• diagnostic.rmse and diagnostic.mae are None when n_complete < refit_r2_min_complete_rows
• diagnostic.rmse and diagnostic.mae are None for Mean, Median, Mode, Constant, Dropped, and Passthrough columns (no change to scalar paths)
• Integration test: after NumericImputer.fit(), diagnostic.rmse and diagnostic.mae are non-negative floats for KNN, Regression, and MICE columns
• Integration test: diagnostic.rmse and diagnostic.mae are None for all scalar strategy columns
• Integration test: diagnostic.rmse and diagnostic.mae are None when complete rows < refit_r2_min_complete_rows
• Integration test: for a MICE block, all columns share the same None / non-None status across all three metric fields
• FittedImputer.to_dict() / from_dict() round-trip preserves rmse and mae on a fitted imputer with model-based columns

Blocked by

• Blocked by Scope 4.2: Implement _compute_fold_metrics() helper for per-fold RMSE and MAE #226