nx_nofloat_autograd_ddp_gate.nx

nx_nofloat_autograd_ddp_gate.nx -- DEEPER R4->LLM: bit-reproducible data-parallel training driven by the REAL no-float autograd tape (vs the linear/SGD stand-in in nx_nofloat_ddp_train_gate). (knowledge/research/2026-06-23-fabric-applies-to-room-and-elsewhere.md §4.) Each of D workers builds the SAME SSE loss graph on its OWN sample shard via the general Q16 autograd (nfa_leaf/nfa_matvec/nfa_mse/nfa_vadd -> one nfa_backward reverse sweep -> nfa_grad), extracts its W gradient, then R4's `ring_allreduce` sums the per-worker gradients and `nfa_sgd` applies the update. THE EXCEED: the updated weights are BYTE-IDENTICAL across D in {1,2,4,8} AND equal to a single-process full-batch step -- because the autograd is exact integer (associative) AND the all-reduce is order- independent. Real trainer components, not a stand-in. Float DDP/NCCL cannot guarantee this. Composes the verified nx_nofloat_autograd (gate B/C pattern) + nx_fabric_collective.ring_allreduce. HONEST SCOPE: a linear W[1x8] model trained by the GENERAL tape (same engine the Q16 transformer uses); N logical workers in one box. license_tier: ORIGINAL expect_exit: 0