BLEU Score

Implement single-reference BLEU with a brevity penalty.

Signature: def bleu(reference: list, candidate: list, max_n: int = 4) -> float

Tokens are pre-split lists. For each n in 1..max_n:

• Compute clipped n-gram precision: sum(min(cand_count, ref_count)) / total_cand_ngrams

Then:

• Geometric mean of precisions
• Brevity penalty: min(1, exp(1 - len(ref) / len(cand))) if len(cand) < len(ref) else 1
• Return BP * geo_mean

Return 0.0 for empty inputs or any zero-precision n-gram order.