SFS: support iterable/generator cv without exhaustion (swev-id: scikit-learn__scikit-learn-25973)

[rowan-stein]

Unable to pass splits to SequentialFeatureSelector

Describe the bug
Passing an iterable of splits (e.g., a generator from a CV splitter) to SequentialFeatureSelector fails. Using cv=5 works, but cv=splits generated via LeaveOneGroupOut.split(X, y, groups) triggers an IndexError.

Steps/Code to Reproduce

from sklearn.datasets import make_classification
from sklearn.feature_selection import SequentialFeatureSelector
from sklearn.neighbors import KNeighborsClassifier
from sklearn.model_selection import LeaveOneGroupOut
import numpy as np

X, y = make_classification()

groups = np.zeros_like(y, dtype=int)
groups[y.size//2:] = 1

cv = LeaveOneGroupOut()
splits = cv.split(X, y, groups=groups)

clf = KNeighborsClassifier(n_neighbors=5)

seq = SequentialFeatureSelector(clf, n_features_to_select=5, scoring='accuracy', cv=splits)
seq.fit(X, y)

Expected Results
Runs without errors.

Actual Results

IndexError: list index out of range
  File "sklearn/model_selection/_validation.py", line 1930, in _aggregate_score_dicts
    for key in scores[0]

Observed Failure and Explanation

• When cv is a generator, it is consumed during the first scoring pass. Subsequent passes see an exhausted iterable, leading to empty results and IndexError in _aggregate_score_dicts.
• Converting to a list (cv=list(splits)) works around the issue but is inconsistent with typical cv handling.

Specification (research by Emerson Gray)

• Align SequentialFeatureSelector cv handling with *SearchCV classes.
• In fit, call check_cv(self.cv, y, classifier=is_classifier(self.estimator)) once to normalize cv.
• If cv is an iterable/generator, check_cv will materialize it once to a reusable list via its wrapper.
• Thread the normalized cv into all cross_val_score calls within SFS to prevent re-listing/re-consuming.
• Update SFS cv parameter docs to mirror *SearchCV: accept int, splitter objects, or iterables of (train, test) indices. Note that iterables are materialized once (memory cost).

Implementation Outline

• File: sklearn/feature_selection/_sequential.py
  • Imports: add from ..base import is_classifier and from ..model_selection import check_cv, cross_val_score.
  • In SequentialFeatureSelector.fit: cv = check_cv(self.cv, y, classifier=is_classifier(self.estimator)), then pass cv into scoring helper.
  • Update _get_best_new_feature_score to accept cv and use it in cross_val_score.
  • Update docstring for cv parameter.

Non-regression Tests

• File: sklearn/feature_selection/tests/test_sequential.py
  • test_sfs_supports_iterable_cv_generator: build splits via LeaveOneGroupOut.split (generator), use as cv in SFS, ensure fit completes and selects expected number of features.
  • test_sfs_baseline_cv_int_runs: verify cv=5 baseline still works.

Versions
Originally observed in 1.2.2; branch scikit-learn__scikit-learn-25973 targets this fix.