Session 34: Fix sklearn 1.6+ squared= deprecation in regression metric registry

Baseline: session 33 closed the public-surface drain. Test runs were emitting ~70 DeprecationWarnings about mean_squared_error(squared=False) per regression CV — a real bug introduced when sklearn 1.6 deprecated that path in favor of a dedicated root_mean_squared_error function. Session 34 fixes the registry.

FIXED — engine#

• FIXED — packages/engine/pycaret/containers/metrics/regression.py — RMSEMetricContainer: now uses sklearn.metrics.root_mean_squared_error directly when available, with a sqrt(mse) shim fallback for sklearn < 1.6. args dict no longer carries squared=False. Eliminates ~7 DeprecationWarning emissions per regression CV (n_folds + Mean + Std rows).
• FIXED — RMSLEMetricContainer — same pattern. Prefers sklearn.metrics.root_mean_squared_log_error when available; otherwise the existing sqrt(msle) shim with np.abs() for negative-input safety.

ADDED — tests#

• ADDED — packages/engine/tests/test_session34_metric_warnings.py — 4 tests:
  • test_rmse_metric_does_not_use_mean_squared_error_directly — locks the score_func to be the dedicated function (or our shim), never the deprecated reference.
  • test_rmse_metric_args_does_not_carry_squared — ensures args["squared"] is gone.
  • test_rmse_score_func_handles_kwargs_without_deprecation — calls the score_func directly with warnings.simplefilter("error", DeprecationWarning).
  • test_regression_cv_does_not_emit_squared_deprecation — full regression create_model under strict warning filtering.

INTERNAL#

• INTERNAL — Why a shim instead of just hard-requiring sklearn 1.6+. The 4.0 dependency floor is scikit-learn>=1.7 (from pyproject.toml), so in practice the shim never runs. But it's cheap insurance: if a downstream user is on a slightly older sklearn (e.g. some numerical / scipy package pinning), they still get correct RMSE without the deprecation. Same logic for RMSLE.
• INTERNAL — What the deprecation was actually doing. Each DeprecationWarning was harmless — sklearn would silently sqrt the MSE anyway and return the right value. The pollution was just to test output. But hiding deprecations would mask real ones; fixing them properly keeps the warning channel meaningful.

Session 34 delta summary#