You shipped expansion. Recall@1000 went up four points...

In BM25, the score is a sum over terms. If you add twelve terms of comparable IDF to a query that originally had three, the original terms lose their relative weight, causing documents that only match the synonyms to climb into the shallow cutoff. This displaces documents that matched the original query more precisely. The metrics are both correct; the superset grew, but the truncation performance worsened. The fix is to down-weight the expansion terms, re-measure, and ensure the metrics used for gating match the actual depth the generator reads.