fix: RAG vector search SQL syntax error breaking assistant chat

- Use CAST(:embedding AS vector) instead of :embedding::vector to avoid
  SQLAlchemy named param conflict with PostgreSQL :: cast syntax
- Add db.rollback() before recording AI usage on failure to prevent
  InFailedSQLTransactionError cascade

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

backend/app/api/endpoints/assistant_chat.py

@@ -155,6 +155,7 @@ async def post_message(
         raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail=str(e))
     except Exception as e:
         logger.exception("Assistant chat message failed: %s", e)
+        await db.rollback()
         await record_ai_usage(
             user_id=current_user.id,
             account_id=current_user.account_id,

backend/app/services/rag_service.py

@@ -143,13 +143,13 @@ async def search(
                 te.chunk_text,
                 te.chunk_type,
                 te.node_id,
-                1 - (te.embedding <=> :embedding::vector) as similarity
+                1 - (te.embedding <=> CAST(:embedding AS vector)) as similarity
             FROM tree_embeddings te
             JOIN trees t ON t.id = te.tree_id
             WHERE te.account_id = :account_id
                 AND t.deleted_at IS NULL
                 {exclude_clause}
-            ORDER BY te.embedding <=> :embedding::vector
+            ORDER BY te.embedding <=> CAST(:embedding AS vector)
             LIMIT :limit
         """),
         params,