Okay, so I have the bug isolated... sort-of. I have two test cases which differ solely by 1 call. In the one case, there's a memory leak (beyond the garbage created by the bisect oid store). In the other, there's none.

The problem is, I can't see any way for the leak to fall out of the difference. The one function that's different just checks a few parameters and then passes them on to the part that doesn't leak when called directly. Argh.

This is probably the time when you say "okay, time for pair programming". Something I'm missing... it must be something trivial. Argh.

Argh.