nx_unsloth_determinism_h2h_gate.nx

nx_unsloth_determinism_h2h_gate.nx -- FIRST measured head-to-head on the Unsloth benchmark: DETERMINISM. The documented real source of float LLM non-determinism is BATCH-INVARIANCE: a reduction's result depends on how it is GROUPED/tiled, which changes with batch size. Here the SAME 1024-value sum (realistic non- exact values 1/k) is computed three ways = three reduction orders (sequential / pairwise-tree / group-of-8), the way different batch tilings would. FLOAT (the Unsloth/PyTorch path) gives DIFFERENT bits per grouping; INTEGER (sovereign) gives BIT-IDENTICAL. Honest: the float divergence is small (a few ULPs) -- but it is REAL, it is the documented issue, and it is enough to flip a token when two logits are close; integer is batch-invariant FOR FREE, where float needs special kernels + a 10-40% cost. No hw writes (Rule 26). expect_exit: 0 license_tier: ORIGINAL