scikit-learn
diff --git a/‎sklearn/ensemble/_bagging.py‎
Lines changed: 3 additions & 3 deletions b/‎sklearn/ensemble/_bagging.py‎
Lines changed: 3 additions & 3 deletions
diff --git a/‎sklearn/ensemble/_forest.py‎
Lines changed: 7 additions & 6 deletions b/‎sklearn/ensemble/_forest.py‎
Lines changed: 7 additions & 6 deletions
diff --git a/‎sklearn/tree/_classes.py‎
Lines changed: 4 additions & 3 deletions b/‎sklearn/tree/_classes.py‎
Lines changed: 4 additions & 3 deletions
@@ -159,9 +159,9 @@ def _parallel_build_estimators(
             sample_weight,
         )
 
-        # Pass the resolved categorical mask from the ensemble. Trees only see the
-        # already-encoded float ndarray, so user values like "from_dtype" or
-        # feature-name lists would otherwise silently resolve to None.
+        # Trees must get the bool mask, not categorical_features="from_dtype" (or
+        # column names). The ensemble already turned X into a NumPy array, so trees
+        # can no longer read dtypes/names and would treat all features as numeric.
         if getattr(ensemble, "is_categorical_", None) is not None:
             if requires_feature_indexing:
                 cat_subset = ensemble.is_categorical_[features]
 
@@ -354,9 +354,10 @@ def fit(self, X, y, sample_weight=None):
                 skip_check_array=True,
             )
             X = self._preprocess_X(X, reset=True)
-            # X has already been encoded to a numeric array. Do not call
-            # validate_data(reset=True) again here because ndarray input would
-            # remove feature_names_in_ captured from the original container.
+            # Feature names were already stored from the original dataframe above.
+            # Encoding turns X into a plain ndarray with no names.
+            # validate_data(reset=True) would treat that as "no feature names"
+            # and delete feature_names_in_.
             X, y = validate_data(
                 self,
                 X,
@@ -497,9 +498,9 @@ def fit(self, X, y, sample_weight=None):
                 self._make_estimator(append=False, random_state=random_state)
                 for i in range(n_more_estimators)
             ]
-            # Pass the resolved categorical mask, not the user parameter. After the
-            # forest encodes X, trees only see a float ndarray so values like
-            # "from_dtype" or feature-name lists would silently resolve to None.
+            # Trees must get the bool mask, not categorical_features="from_dtype" (or
+            # column names). The forest already turned X into a NumPy array, so trees
+            # can no longer read dtypes/names and would treat all features as numeric.
             if self.is_categorical_ is not None:
                 for tree in trees:
                     tree.set_params(categorical_features=self.is_categorical_)
 
@@ -303,9 +303,10 @@ def _fit(
             )
             check_y_params = dict(ensure_2d=False, dtype=None)
             if has_categorical:
-                # X has already been encoded to a numeric array. Do not call
-                # validate_data(reset=True) again here because ndarray input would
-                # remove feature_names_in_ captured from the original container.
+                # Feature names were already stored from the original dataframe above.
+                # Encoding turns X into a plain ndarray with no names.
+                # validate_data(reset=True) would treat that as "no feature names" and
+                # delete feature_names_in_.
                 X = check_array(X, input_name="X", estimator=self, **check_X_params)
                 y = check_array(y, input_name="y", estimator=self, **check_y_params)
                 _check_n_features(self, X, reset=False)