FIX: Avoid deprecated pandas Series growth in reductions - #1801

Description

Fix the pandas 3.1 compatibility failures in the reductions suite without weakening warnings-as-errors.

The all-deps CI logs on #1741 show Pandas4Warning when constraint bases and ExponentiatedGradient enlarge a Series via .at. This is a separate compatibility fix, not a duplicate of #1741's deterministic-collection work.

• Use .loc for the seven insertions of new Series keys.
• Initialize neg_basis_present as boolean: after replacing .at, pandas also correctly exposes its previously incompatible float initialization.
• Update the same insertion pattern in the LP-failure test helper and replace deprecated reset_index(inplace=True) in the group-loss test.
• Add four regression cases covering demographic parity, equalized odds, bounded group loss, and end-to-end ExponentiatedGradient fit/predict. Add a release note.

No warning filters, algorithm changes, or dependency pins are introduced.

Reproduction on unchanged upstream

Base: 21c6b4c384126d388e2da7ba13ff0c434ebd42cd.

With only the new regression file added to unchanged production source:

pytest -q test/unit/reductions/test_series_growth.py
4 failed (Pandas4Warning)

The two LP-failure cases and group-loss GPA test also fail on a clean upstream worktree: 3 failed. These are the three existing failures remaining after fixing production Series growth.

Tests

• New tests added.
• Existing tests adjusted for the same pandas compatibility changes.

Executed on Python 3.12.15 / pandas 3.1.0:

pytest -q test/unit/reductions
3909 passed, 320 skipped, 888 deselected
ruff check <changed Python files>
passed
ruff format --check <changed Python files>
passed
uv build
sdist and wheel built

Minimum-dependency compatibility check on Python 3.11.17 / pandas 2.0.3 / numpy 1.24.4 / scikit-learn 1.6.0 / scipy 1.9.3:

pytest -q test/unit/reductions/test_series_growth.py
4 passed

Documentation

• Release note in the unreleased v0.16.0 notes.
• No API changes.

This addresses the reproduced reductions compatibility failure. It does not claim to fix separate OpenML network errors or other optional-ML CI failures.

AI disclosure: investigated, implemented and tested using Perplexity Computer. No independent human review is claimed.