fix: normalize vector store scores to similarity (higher = better)

The threshold parameter in Memory.search() was silently broken for 11
vector stores because they returned raw distance scores (lower = better)
while the threshold check assumed similarity scores (higher = better).
This caused threshold filtering to be inverted — good matches got
dropped and bad matches passed through.

Convert all affected vector stores to return similarity scores:
- Cosine distance stores: score = max(0.0, 1.0 - distance)
- L2 distance stores: score = 1.0 / (1.0 + distance)
- Stores computing similarity then discarding it: return similarity directly

Fixes #4453
Related: #3283

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

tests/vector_stores/test_milvus.py

@@ -129,7 +129,7 @@ class TestMilvusDB:
         # Verify results are parsed correctly
         assert len(results) == 1
         assert results[0].id == "mem1"
-        assert results[0].score == 0.8
+        assert results[0].score == 0.8  # COSINE: Milvus returns similarity directly
 
     def test_search_different_user_ids(self, milvus_db, mock_milvus_client):
         """Test that search works with different user_ids (reproduces reported bug)."""
@@ -228,10 +228,10 @@ class TestMilvusDB:
         
         assert len(parsed) == 2
         assert parsed[0].id == "mem1"
-        assert parsed[0].score == 0.9
+        assert parsed[0].score == 0.9  # COSINE: Milvus returns similarity directly
         assert parsed[0].payload == {"user_id": "alice"}
         assert parsed[1].id == "mem2"
-        assert parsed[1].score == 0.85
+        assert parsed[1].score == 0.85  # COSINE: Milvus returns similarity directly
 
     def test_update_with_none_vector_fetches_existing(self, milvus_db, mock_milvus_client):
         """Test that update with vector=None fetches the existing vector (fixes #3708)."""