The stream-end reaction cleanup now reads from _reaction_ids instead of metadata, so pre-populate the dict in the test instead of passing reaction_id via metadata.