fix(metrics): score ChrF against each reference for multiple references

[chuenchen309]

Details

ChrF.score accepts reference: Union[str, Sequence[str]] and builds a list of references, so multiple references are a documented, type-hinted feature. But the default NLTK backend passed that whole list straight to nltk.translate.chrf_score.sentence_chrf(references, candidate), whose reference argument is a single reference string — not a list. NLTK misinterprets the list and returns a wrong score: a candidate identical to one of the references scored ~0.26 instead of 1.0 (even lower than scoring against that reference alone).

This scores the candidate against each reference separately and takes the best match — standard multi-reference chrF semantics (as in sacrebleu / HuggingFace evaluate). Single-reference scoring is unchanged, and the custom chrf_fn injection contract is preserved (it still receives the full references list).

• Before: ChrF().score(output="hello world", reference=["completely different text here", "hello world"]).value → 0.2552
• After: → 1.0

Verified against the real installed code before/after. Added test_chrf_metric__multiple_references__scores_against_best_match (the existing ChrF tests only covered single-reference and the custom-chrf_fn path, so the real multi-reference NLTK path was untested) — red before the fix, green after. Full ChrF test set: 3 passed. ruff check + ruff format: clean.

Change checklist

• User facing
• Documentation update

Issues

• Resolves #

AI-WATERMARK

AI-WATERMARK: yes

• Tools: Claude Code
• Model(s): Claude (Anthropic)
• Scope: Bug found via an AI-assisted review pass over sdks/python/src/opik/evaluation/metrics/; the fix and regression test were AI-drafted and then independently reproduced and verified (red-before/green-after against the real installed code, ruff clean) before submitting.