v5.7.1: Calibrated Recall

Note recall returns fewer weak matches by default.

Changed

• Note recall defaults to a similarity floor calibrated per embedding model. For BAAI/bge-base-en-v1.5 the floor is 0.80 on the tools' (1 + cosine) / 2 scale, which is cosine 0.60. On 812 blind-labeled pairs from real recall queries, it raised the share of returned notes that are relevant from 0.31 to 0.39 and cut notes per call from 5.0 to 3.9, while recall of relevant notes fell only from 0.75 to 0.72. Other models keep 0.30. session_recall and the notes section of query_signals use it when minSimilarity is omitted; a caller's value still wins, query_signals' code section keeps 0.5, and bookmark matching, which was not measured, keeps 0.30. An empty recall now says no note was close enough. Method and numbers: docs/code-graph-context/research/recall-relevance-calibration.md in the owner's docs tree. (src/core/embeddings/embedding-config.ts)