nx_distributed_train_gate.nx

nx_distributed_train_gate.nx -- HEADLINE EXCEED PROOF: deterministic DISTRIBUTED training. Composes nx_fabric_collective::ring_allreduce (the byte-exact INTEGER gradient all-reduce) into a minimal no-float data-parallel trainer (forward -> backward -> all-reduce gradient -> integer optimizer step) and PROVES the trained weights are BIT-IDENTICAL regardless of the process count N (the data-parallel sharding). This is the property NO float stack has: NCCL's float all-reduce is non-deterministic across nodes, and the batch-invariant frontier (Thinking Machines) is single-node only. Ours holds because integer reduction is associative + commutative -> the summed gradient is independent of how the global batch is sharded. criteria: 1 N=1 full-batch baseline trains AND W actually MOVES (real work, not all-zeros) 2 N=2 data-parallel -> final W byte-identical to the N=1 baseline 3 N=4 data-parallel -> final W byte-identical to the N=1 baseline 4 N=8 data-parallel -> final W byte-identical to the N=1 baseline (deterministic across ALL shardings) negative control: NC1 N=8 WITHOUT the all-reduce (proc-0 local gradient only) -> W DIFFERS (the all-reduce is load-bearing) expect_exit: 0 license_tier: ORIGINAL