Reference Issues/PRs

Fixes #33457

What does this implement/fix? Explain your changes.

Proposed Changes
This PR introduces sample_weight support to the NearestCentroid classifier, allowing for weighted means (Euclidean) and weighted medians (Manhattan), with strict parity between dense and sparse inputs.

What was built:

• NearestCentroid.fit(): Now accepts sample_weight and applies it correctly through all four code paths: weighted Euclidean mean (dense and sparse) and weighted median (dense Manhattan via _weighted_percentile and sparse Manhattan via a new utility).
• Shrinkage & Priors: Successfully applies weights to empirical priors, weighted pooled variance, and the shrinkage m-parameter.
• Sparse Utilities: Added csc_weighted_median_axis_0 and _get_weighted_median to sklearn/utils/sparsefuncs.py to handle implicit zeros correctly without spurious zero insertion.

Key guarantees and edge cases handled:

• Uniform Parity: Uniform weights produce results mathematically identical to the unweighted path.
• Fractional Weights: Supported seamlessly, provided sum(weights) > n_classes.
• Zero-Weight Classes: Immediately raises a ValueError with a clear message to prevent downstream NaNs.
• Float Arithmetic: Clamped sum_w_zeros to prevent tiny negative values caused by floating-point inaccuracies.
• NumPy 2.0 Compatibility: Replaced deprecated np.ptp with X.max(axis=0) - X.min(axis=0).

Files Changed:

• sklearn/neighbors/_nearest_centroid.py
• sklearn/utils/sparsefuncs.py
• sklearn/neighbors/tests/test_nearest_centroid.py
• sklearn/utils/tests/test_sparsefuncs.py

Test Coverage:
Added test_csc_weighted_median_axis_0 and test_nearest_centroid_sample_weight to verify identical behavior between dense/sparse inputs, correct centroid shifting, and proper error handling. All 24 tests pass cleanly.

AI usage disclosure

I used AI assistance for:

• Code generation (e.g., when writing an implementation or fixing a bug)
• Test/benchmark generation
• Documentation (including examples)
• Research and understanding

Any other comments?

All 24 tests pass locally, including parity checks between sparse and dense formats and against zero-weight edge cases. Looking forward to any feedback!