nx_blockfloat_ddp_gate.nx

nx_blockfloat_ddp_gate.nx -- THE CONVERGENCE: block-float weights inside DETERMINISTIC DISTRIBUTED training. Unites this session's two threads: block-float quant (FP8-class dynamic range, deterministic) + the byte-exact integer ring all-reduce (nx_fabric_collective). Each step: quantize W to block-float (per-block power-of-2 scale), FORWARD through the block-float dot, straight-through gradient, ring all-reduce across N data-parallel processes, integer optimizer step. PROVES the trained weights are BIT-IDENTICAL regardless of N -- so the combination quantized + sovereign + distributed + bit-exact-reproducible holds, which nothing in the float world can claim. criteria: 1 N=1 block-float DDP trains (W moves off zero) 2 N=2 -> final W byte-identical to N=1 3 N=4 -> final W byte-identical to N=1 4 N=8 -> final W byte-identical to N=1 (deterministic across ALL shardings WITH block-float in the loop) 5 block-float is REAL: block-float W != full-precision W (the quant changed the training, not a no-op) expect_exit: 0 license_tier: ORIGINAL