nx_fabric_collective_gate.nx

nx_fabric_collective_gate.nx -- LIVE GATE for R4 (collective communication: ring all-reduce). Proves ring all-reduce is byte-exact correct AND its per-process bottleneck stays ~constant as N grows, where naive root-based all-reduce bottlenecks at O(N) -- with negative controls. criteria: 1 ring all-reduce byte-exact correct (N=8) (the algorithm is REAL, not a cost model) 2 naive all-reduce byte-exact correct (N=8) (baseline sanity) 3 ring bottleneck * 3 <= naive bottleneck (N=8) (>=3x less per-process data; actual ~Nx) 4 ring max_bytes(N=16) <= 2 * ring max_bytes(N=4) (ring per-process BOUNDED as N grows) 5 naive max_bytes(N=16) >= 3 * naive max_bytes(N=4)(naive root grows ~linearly -- the contrast) negative controls: NC1 ring with the REDUCTION disabled -> WRONG result (the sum is load-bearing, not the shuffle) NC2 ring correct at N=4 AND N=16 (correct across scale, not a fixed-N fluke) NC3 ring(N=8) result == naive(N=8) result byte-exact (two independent algorithms agree) expect_exit: 0 license_tier: ORIGINAL