Tracing NumPy RNG draws across processes: three hooking gotchas

To catch duplicated random streams without editing the code under test, a sitecustomize.py on PYTHONPATH can replace numpy.random.mtrand._rand, the np.random.* module functions, np.random.RandomState / Generator / default_rng and random.Random with traced subclasses. Each draw logs a hash of the state before it, so the same hash in two processes means the same numbers. Three gotchas (NumPy 2.5.4, CPython 3.12/3.14). (1) sys.setprofile never sees RandomState draws: the Cython methods are bound 'method' objects, so no c_call event fires. Subclass and forward instead. (2) RandomState/Generator __reduce__ rebuild a PLAIN instance through numpy's private ctor, so pickling or deepcopy silently drops your subclass; override __reduce__ and also skip logging the set_state() that unpickling calls, or copies look like reseeds. (3) If you replace the np.random.Generator name, plain Generators made in compiled code fail isinstance checks; give the subclass a metaclass whose __instancecheck__ delegates to the real class. Patch from a meta-path finder wrapping numpy.random's exec_module, because NumPy 2 imports numpy.random lazily.