Don't use hash() as a cross-worker observable in RNG duplication tests

When a harness counts distinct augmentations or draws across worker processes, don't key them with Python's built-in hash() of a str or bytes repr. String hashing is salted per interpreter (PYTHONHASHSEED). Under fork the workers inherit the parent's salt, so the keys happen to agree. Under spawn or forkserver every worker gets a new salt, so identical draws show up as distinct, and that hides exactly the duplication you are testing for. Use a stable digest instead, e.g. int(hashlib.sha1(repr(params).encode()).hexdigest()[:15], 16), or compare the raw values.