fix: use isolated session in _on_query to prevent premature commit

The _on_query method was calling db.session.commit() on the Flask-scoped
SQLAlchemy session, which committed all pending dirty state from the
current request — not just the DatasetQuery audit rows.

This broke transaction isolation: if the downstream workflow failed, the
subsequent db.session.rollback() could not revert the already-committed
modifications (e.g. token deductions, partial node executions), leaving
dirty data in the database.

The same file already demonstrates the correct pattern in
_on_retrieval_end, which uses sessionmaker(bind=db.engine).begin() with an
independent session. This change applies the same approach to _on_query.

Additionally fixed a latent bug where dataset_queries.add_all() was called
inside the loop on every iteration, re-adding previously accumulated rows.

Fixes #37886

api/core/rag/retrieval/dataset_retrieval.py

@@ -1030,6 +1030,9 @@ class DatasetRetrieval:
     ):
         """
         Persist dataset query audit rows for retrieval requests.
+
+        Uses an independent session to avoid committing the request-scoped
+        db.session, which would break transaction isolation for the caller.
         """
         if not query and not attachment_ids:
             return
@@ -1059,9 +1062,9 @@ class DatasetRetrieval:
                     created_by=created_by,
                 )
                 dataset_queries.append(dataset_query)
-            if dataset_queries:
-                db.session.add_all(dataset_queries)
-        db.session.commit()
+        if dataset_queries:
+            with sessionmaker(bind=db.engine).begin() as session:
+                session.add_all(dataset_queries)
 
     def _retriever(
         self,