So, finally I have set up a system for doing bulk tests, which is surprisingly illuminating. Results of this will follow, but if you have a data set where xia2 fails but manual processing works please get in touch - I will be happy to add it to the test system and work on making it work.