fix(runtime): improve local service reliability

Close SQLite handles safely on Windows, clean orphaned vectors, normalize Ollama endpoints, and surface model errors without leaking reasoning tags.

Co-authored-by: Cursor <cursoragent@cursor.com>

tests/test_documents.py

@@ -97,6 +97,61 @@ def test_conversation_recall_uses_vec_and_matches_brute_force():
     asyncio.run(run())
 
 
+def test_delete_conversation_removes_message_vectors():
+    """Deleting a conversation must take its embeddings with it — orphaned
+    message_vectors rows are invisible to recall but grow the DB forever."""
+    s = _store()
+
+    async def run():
+        s.create_conversation("c1")
+        s.add_message("c1", "user", "tell me about lego star wars")
+        s.add_message("c1", "assistant", "lego star wars is a fun game")
+        s.create_conversation("c2")
+        s.add_message("c2", "user", "gpu vega vram notes")
+
+        # the first search lazily backfills a vector for every message
+        await s.semantic_search_conversations("lego star wars", _fake_embed, limit=2, min_score=0.1)
+        conn = s._connect()
+        assert conn.execute("SELECT COUNT(*) FROM message_vectors").fetchone()[0] == 3
+        conn.close()
+
+        s.delete_conversation("c1")
+
+        conn = s._connect()
+        orphans = conn.execute(
+            "SELECT COUNT(*) FROM message_vectors v "
+            "LEFT JOIN messages m ON m.id = v.message_id WHERE m.id IS NULL"
+        ).fetchone()[0]
+        assert orphans == 0
+        assert conn.execute("SELECT COUNT(*) FROM message_vectors").fetchone()[0] == 1  # c2 untouched
+        if s.vec_enabled:  # the ANN mirror is pruned too, not just the JSON table
+            assert conn.execute("SELECT COUNT(*) FROM vec_messages").fetchone()[0] == 1
+        conn.close()
+
+    asyncio.run(run())
+
+
+def test_startup_sweeps_pre_existing_orphan_vectors():
+    """Databases written before delete_conversation cleaned up after itself are
+    repaired the next time the store opens them."""
+    import json
+    path = Path(tempfile.mkdtemp()) / "t.db"
+    s = PersistentMemoryStore(path)
+    s.create_conversation("c1")
+    mid = s.add_message("c1", "user", "lego star wars")
+    conn = s._connect()
+    conn.execute("INSERT INTO message_vectors (message_id, embedding) VALUES (?, ?)",
+                 (mid, json.dumps([1.0, 0.0])))
+    conn.execute("DELETE FROM messages WHERE id = ?", (mid,))  # the old leaky delete
+    conn.commit()
+    conn.close()
+
+    reopened = PersistentMemoryStore(path)
+    conn = reopened._connect()
+    assert conn.execute("SELECT COUNT(*) FROM message_vectors").fetchone()[0] == 0
+    conn.close()
+
+
 def test_projects_scope_documents_and_survive_delete():
     s = _store()