I eventually gave up without finding the issue, but somewhere deep inside one version of the Sphinx full text search software was a bug that would sometimes switch query got what result set. It only happened sometimes when queries were within a few seconds of each other, but it wouldn't happen with only one front end process even in multithreaded mode and would disappear if requests were _too_ close together. If I'd found a way to reproduce it I'd have submitted it to the Sphinx team, but after a few days of potentially private info leaking I gave up and moved to PostgreSQL's FTS.