FIX Use uniform probabilities for zero-affinity rows - #34835

Reference Issues/PRs

Fixes #34351. Related to #34350.

What does this implement/fix? Explain your changes.

Zero-affinity rows currently normalize to all-zero values during fit and can produce NaN values in predict_proba. Neither result is a valid categorical probability distribution.

This change introduces one private normalization helper shared by fit and predict_proba. Rows with a nonzero sum retain the existing normalization, while zero-sum rows receive uniform probabilities. The uniform fallback represents the absence of class information without preferring a class and follows the direction agreed in #34351.

Regression tests cover disconnected training samples and far-away prediction samples for both LabelPropagation and LabelSpreading.

Tests:

• python -m pytest sklearn/semi_supervised/tests/test_label_propagation.py -q (175 passed, 60 skipped)
• python -m ruff check sklearn/semi_supervised/_label_propagation.py sklearn/semi_supervised/tests/test_label_propagation.py
• python -m ruff format --check sklearn/semi_supervised/_label_propagation.py sklearn/semi_supervised/tests/test_label_propagation.py

Introduce yourself

I am contributing this focused numerical correctness fix after following the maintainer-approved direction in #34351.

AI usage disclosure

I used AI assistance for:

• Code generation
• Test generation
• Research and understanding

Any other comments?

The main review point is applying the uniform fallback during LabelPropagation iterations as well as final normalization. This keeps disconnected rows valid throughout fitting and is covered by the existing closed-form and convergence tests.

This pull request includes code written with the assistance of AI.
The code has not yet been reviewed by a human.