tf.keras Sequence + use_multiprocessing: workers share RNG, every epoch replays

Read from the TF 2.4.4 source and the standalone keras 2.6/2.10/2.15 wheels, then run with the real data_utils.py (TF imports stubbed), CPython 3.8 / NumPy 1.19.5, fork. Three facts combine. (1) OrderedEnqueuer's pool initializer gets random_seed=None, so it never reseeds NumPy. (2) _run creates a new Pool every epoch, forked from a parent whose RNG state does not advance, because all draws happen in workers. (3) fit() wraps a Sequence in KerasSequenceAdapter, which calls x[0] in the parent to peek at shapes, so a generator or RandomState created lazily on first use is created in the parent and copied into every worker. Result with 4 workers, 24 batches, 3 epochs, draws via global np.random: 7-9 distinct of 72; with a lazily created RandomState: 8 of 72 (72 of 72 without the parent peek). Two fixes tested at 72/72 in the same setup: count epochs in on_epoch_end (it runs in the parent before the next pool is forked) and draw from np.random.default_rng([base_seed, epoch, index]) in __getitem__, which is also reproducible; or keep a generator plus the pid that created it and recreate it with default_rng() whenever os.getpid() differs.