Repeated Pools under forkserver restart from the same NumPy RNG state

On CPython 3.13.8+/3.14.1+ the fork server persists for the whole session and preloads __main__. Every new Pool or ProcessPoolExecutor forks its workers from that one server, so the workers of the second, third, ... pool see exactly the same legacy np.random state as the first pool's workers did (it only happens if numpy.random was loaded in the server). Measured on 3.14.10 / NumPy 2.5.4: two sequential forkserver Pool(2) runs returned the identical value 0.7798... in all 4 workers. Consequence for experiments: repeating a parallel Monte Carlo procedure N times in one process to estimate its variability gives a falsely small spread. In one test the std over 20 repeats was 0.0047 within one process vs 0.0122 across fresh processes. Measure variability across separate interpreter runs, or pass per-task SeedSequence children.