Maya process memory stays high after File > New Scene: leak or allocator retention?

During long Maya sessions, resident memory climbs steadily and is not returned to the OS after closing a heavy scene with File > New Scene. The only eventual failure condition is a platform-dependent out-of-memory crash, so there is no error string to anchor diagnosis. The unresolved question is how to classify this growth before it reaches that point.

Two documented contributors complicate the picture: the undo queue, whose length is configurable via the undoInfo command and which grows without bound under default settings, and the evaluation/dependency-graph caches retained in Parallel and Evaluation Manager modes. On top of that, Maya's memory allocators and loaded plug-ins are known to retain arenas even after scene data is destroyed, so Task Manager or top numbers alone cannot distinguish a genuine leak from fragmentation and retention. Third-party plug-ins and Python API reference cycles (including legacy MScriptUtil misuse) add further suspects.

Assume a recent Maya version on Windows or Linux, with behavior potentially differing by evaluation mode and OS allocator.

1. Is there a supported way to attribute Maya's resident memory to the undo queue, evaluation caches, or plug-in allocations, short of external heap profiling?
2. Does bounding the undo queue via undoInfo and flushing caches reliably cap growth, or is a full restart the only guaranteed reclamation?
3. What minimal reproduction best separates a true plug-in leak from allocator retention?