Apparently `scipy.optimize` uses NumPy's RNG, so I tried setting `np.random.seed(SEED)` but still cannot get model fitting to be reproducible.