It turns out that between writing the previous fixes and retesting, even more problems arose for java. I think java is more affected because Reftype is bigger than Class, but most uses only need Class so trivial type tests affect java more than other langauges.

List of changes:

• The params string change was more important than I thought. Although the fix only removed 10s, that predicate was evaluated multiple times so the saving multiplied. This caches it instead so we only evalaute it in full once.
• The changes to AndroidComponent were no longer good enough. Instead this forces the TC to be materialised rather than use the bad magic version (which was done repeatedly). I decided to do this via getAnAncestor so I didn't have to define new predicates and as we evaluate the full TC we always want to use the non-magicked version after that. Note that only one use in each stage needs to be converted but as AndroidComponent was top in the magic as it has the first name alphabetically. I have opened an internal issue about making doing this change easier.
• There was fairly bad magic for ConfusingOverloading.ql, the existing magic didn't help much so removing it seems better.