- dataflow.md: update DatasetFactory.load signature, stream vs record access, Store._offsets - architecture.md: add tokenizer to Pipeline, TokenizeTransform class, RecordDataset, Streamable/Recordable mixins, fix GRPOStrategy (old_model/sync_old_model) - training.md: DPO reduction="sum", GRPO rho_t uses pi_old, gradient_clipping always registered - params.md: --max_grad_norm default None