fix: score BleuScore as a single segment so every sentence counts

[Ziad-Abdalla]

Issue Link / Problem Description

• Fixes BleuScore evaluates only the first sentence (corpus_bleu reference structure bug) #2888

BleuScore only grades the first sentence of a multi-sentence response. Ragas splits both texts on ". " and passes the reference as [[s1], [s2], ...], which is NLTK's layout. sacrebleu reads references as a list of reference streams (one entry per hypothesis), so it sees one segment with several alternative references, and everything after the first response sentence is ignored.

Example, with reference "The capital of France is Paris. It is known for the Eiffel Tower.":

response	before	after
"The capital of France is Paris. Bananas are a popular fruit."	1.0	0.417
"Bananas are a popular fruit. The capital of France is Paris."	0.0	0.417

It also gets more urgent soon: sacrebleu has merged a stricter reference-length check (mjpost/sacrebleu#294, not released yet). Since Ragas doesn't pin sacrebleu, once that ships the current layout will raise on any multi-sentence response.

Changes Made

• Score the whole response against the whole reference as a single segment: corpus_bleu([response], [[reference]]). This is the same shape ChrfScore already uses, and it no longer depends on the ". " sentence splitter.
• Applied to both implementations: ragas.metrics._bleu_score and ragas.metrics.collections._bleu_score.
• User kwargs still pass straight through to sacrebleu.

Testing

How to Test

• Automated tests added/updated
• New tests/unit/test_bleu_score.py (18 tests, both implementations): later sentences affect the score, sentence order doesn't wipe out credit, unequal sentence counts don't truncate or raise, scores match whole-text BLEU, kwargs pass through, and sacrebleu is called with one segment.
• 16 of the 18 fail on current main and pass with the fix. The existing tests/e2e/metrics_migration/test_bleu_migration.py and the chrF tests still pass. ruff format and ruff check are clean.

References

• Related issues: BleuScore evaluates only the first sentence (corpus_bleu reference structure bug) #2888
• There are already open PRs for this issue (fix: BleuScore silently scores only the first sentence of the response #2908, fix(metrics): evaluate full text in BleuScore instead of first sentence only #2924, fix(metrics): BleuScore evaluates full text, not only first sentence #2941 and fix: evaluate full reference text in BleuScore instead of first sentence only #3012 take the same single-segment approach, credit to their authors). Opening this one after talking with @jjmachan so it can go into the next release. Happy to close it in favour of one of those if you'd rather merge an existing one.
• External references: Make corpus_score_args checks stricter mjpost/sacrebleu#294