FIX prevent catastrophic cancellation in r_regression (#11395) - #34834

Reference Issues/PRs

Fixes #11395.

What does this implement/fix? Explain your changes.

r_regression computes centered feature norms using a moment-based expression. Floating-point cancellation can make the mathematically non-negative centered sum of squares slightly negative for constant or near-constant columns, causing an invalid square root and propagating invalid correlations. This change clamps the intermediate squared norms to zero before sqrt, preserving existing results for active non-constant features and keeping dense, CSR, and CSC inputs consistent. Regression tests cover constant features, force_finite=False, sparse formats, high-magnitude cancellation, and active-feature preservation.

Introduce yourself

I am FinalSunFlower, contributing a numerically focused bug fix to scikit-learn's feature-selection utilities. I use scikit-learn for classical machine-learning workflows and was prompted by issue #11395 to improve the stability of r_regression on degenerate inputs.

AI usage disclosure

I used AI assistance for:

• Research and understanding
• Code generation (implementation and tests)
• Test/benchmark generation
• Documentation (including examples)

Any other comments?

Local CPU validation passed for the targeted regression tests, the complete feature-selection test module, Ruff checks, and formatting checks. The changelog entry is included in doc/whats_new/upcoming_changes/sklearn.feature_selection/34834.fix.rst.