FIX: match XL-SafetyBench cultural category labels to upstream

[Edson (Zhuoxi2000)]

Description

Fixes #2987

Three XLSafetyBenchCulturalCategory values were missing the commas that the upstream dataset uses in the category column of every data/cultural/<country>/scenario_prompts.csv (AIM-Intelligence/XL-SafetyBench on Hugging Face). The cultural loader filters rows by exact string match, so filtering by any of these members matched no rows and raised ValueError: No XL-SafetyBench cultural scenarios matched the configured filters. This PR changes the three values to the upstream labels:

• FOOD_DIETARY_LAW_AND_HOSPITALITY: Food Dietary Law & Hospitality -> Food, Dietary Law & Hospitality
• DEATH_GRIEF_AND_FUNERAL_PRACTICES: Death Grief & Funeral Practices -> Death, Grief & Funeral Practices
• HIERARCHY_ADDRESS_AND_SOCIAL_DEFERENCE: Hierarchy Address & Social Deference -> Hierarchy, Address & Social Deference

This also corrects the class-level harm_categories list on _XLSafetyBenchCulturalDataset, which is built from these values. Member names are unchanged. The category filter only accepts enum members, so callers that pass members need no change. Only code that compares against the old .value strings would notice, and those strings never matched any upstream row.

Tests and Documentation

New parametrized test test_cultural_category_filter_matches_upstream_labels_with_commas in tests/unit/datasets/test_xl_safety_bench_dataset.py. It feeds a row with each upstream comma label through the category filter.

Before the fix (the new test run against main at 4659ebe):

$ python -m pytest tests/unit/datasets/test_xl_safety_bench_dataset.py -q
FAILED ...::test_cultural_category_filter_matches_upstream_labels_with_commas[Food Dietary Law & Hospitality-Food, Dietary Law & Hospitality]
FAILED ...::test_cultural_category_filter_matches_upstream_labels_with_commas[Death Grief & Funeral Practices-Death, Grief & Funeral Practices]
FAILED ...::test_cultural_category_filter_matches_upstream_labels_with_commas[Hierarchy Address & Social Deference-Hierarchy, Address & Social Deference]
3 failed, 34 passed

After the fix:

$ python -m pytest tests/unit/datasets/test_xl_safety_bench_dataset.py -q
37 passed
$ python -m pytest tests/unit/datasets/ -q
4960 passed

Against the real dataset (France), every XLSafetyBenchCulturalCategory member now returns scenarios: 15 each, and 25 for Legal Landmines. Before the fix, the three affected members raised ValueError.

pre-commit run --files on both changed files passes, including ruff format, ruff check (v0.16.9, as pinned in .pre-commit-config.yaml) and ty. No documentation change is needed, and no notebooks or docs are touched, so JupyText was not run.

[hannahwestra25]

thanks for contributing