Tracking down the memory leak is not going well. So far each time I think I've isolated a test case that produces a leak it eventually stabilises at some memory level after an hour or two of operation.

There's something funky going on that I'm thinking may be a reference loop that's requiring the cyclic garbage collector to do the actual collection. Given that one of the tests is stabilising at 400+MB, it's quite reasonable to imagine memory getting exhausted even if there's no memory leak per-se (simply collection running way behind generation of the garbage).

Will be looking into that possibility tomorrow. A few gc.collect() calls should be sufficient to prove/disprove the theory for the test cases.