# Changes in Dr_SFC GĐ1 simple v0.2

This patch corrects hyperparameter tuning for precipitation.

## What changed

- `src/models.py`
  - Added `PhysicalTargetScoringWrapper` and `physical_neg_mae_scorer`.
  - Ridge/ElasticNet still use `GridSearchCV`.
  - RF/GBM/XGBoost still use `RandomizedSearchCV`.
  - For precipitation, tuning now selects hyperparameters by MAE in physical units after inverse transform:
    `pr_bc = max(pr_factor_hat * pr_o_cli, 0)` versus `pr_o`.
  - For temperature, tuning by anomaly MAE is retained because it is exactly equivalent to physical MAE after adding `t2_o_cli`.

- `src/bc_runner.py`
  - Passes `pr_o_cli` as a scoring-only physical scale for precipitation CV.
  - The scale is not used as an input predictor.

- `src/features.py`
  - Added `select_samples_with_aux()` for aligned sampling of the scoring-only precipitation climatology scale.

## Why

The previous v0.1 tuning optimized transformed-space MAE for precipitation (`pr_factor`). The requested criterion is MAE of corrected precipitation (`pr_bc`) against observed precipitation (`pr_o`).