Summary

• relax _wrap_in_pandas_container so pandas outputs keep their own index when row counts change and only align when lengths match
• adjust ndarray wrapping to ignore mismatched indices while preserving column naming
• add targeted FeatureUnion and utility tests covering mismatch and aligned cases

Testing

• pytest -q sklearn/utils/tests/test_set_output.py::test__wrap_in_pandas_container_preserve_index_on_length_mismatch_dataframe
• pytest -q sklearn/utils/tests/test_set_output.py::test__wrap_in_pandas_container_ndarray_ignore_index_on_length_mismatch
• pytest -q sklearn/utils/tests/test_set_output.py::test__wrap_in_pandas_container_align_when_lengths_match
• pytest -q sklearn/tests/test_pipeline.py::test_feature_union_pandas_preserves_aggregated_index
• pytest -q sklearn/tests/test_pipeline.py::test_feature_union_pandas_aligns_index_when_lengths_match
• flake8 sklearn/utils/_set_output.py sklearn/utils/tests/test_set_output.py sklearn/tests/test_pipeline.py

Reproduction

Steps from #72:

import pandas as pd
from sklearn.base import BaseEstimator, TransformerMixin
from sklearn import set_config
from sklearn.pipeline import make_union

index = pd.date_range(start="2020-01-01", end="2020-01-05", inclusive="left", freq="H")
data = pd.DataFrame(index=index, data=[10] * len(index), columns=["value"])
data["date"] = index.date

class MyTransformer(BaseEstimator, TransformerMixin):
    def fit(self, X, y=None):
        return self

    def transform(self, X, y=None):
        return X["value"].groupby(X["date"]).sum()

set_config(transform_output="default")
print(make_union(MyTransformer()).fit_transform(data))

set_config(transform_output="pandas")
print(make_union(MyTransformer()).fit_transform(data))

Observed failure:

ValueError: Length mismatch: Expected axis has 4 elements, new values have 96 elements

Stack trace summary: the error is raised in sklearn/utils/_set_output.py when _wrap_in_pandas_container assigns the original input index to a pandas output with fewer rows.

Fixes #72.