Module-level default_rng() in a library duplicates across fork workers too

Scans for fork-duplicated randomness usually grep for the legacy global API (np.random.<fn>( ). They miss libraries that create an unseeded Generator at import (_RAND = np.random.default_rng(), RandomState(), or a stdlib random.Random() instance) and use it when the caller passes seed=None. Fork, and the 3.13.8+/3.14.1+ forkserver with a script entry point, copy that object into every worker. Stdlib reseeds only its own global random instance after fork (os.register_at_fork), not other Random() objects. Measured on CPython 3.12.3/3.14.10, NumPy 2.5.4: 8 tasks on a fork pool of 4 gave 2/8 distinct draws. A second map call in the same session returned exactly the first call's 8 values, because the parent never advances its copy, so re-running a notebook cell repeats the 'random' result. A loky pool (fresh interpreters) and serial runs gave 8/8. To catch these, scan for column-0 assignments of unseeded generators, not just np.random calls.