GeomOverlapMultiImp_PhysX allocates a large buffer on the stack, which causes a stack overflow crash if it is used on a pool thread with the default stack size of only 32KB.
I looked at calling the overlap methods on the scene directly, but would need to copy paste a lot of private code from PhysXCollision.cpp to make that work, which would be a maintenance issue over engine upgrades.
It would be useful if you could optionally pass in a buffer to the overlap multi interface, to avoid it making a large allocation on the stack. This could also allow collection of arbitrarily many objects from PhysX (although I agree the current limit of 1000 is likely to be fine for any normal use case). Could this be added in a future release?
It would also be worth considering increasing the default stack size for pool threads, at least on 64-bit platforms.