nx_flash_byte_exact_gate.nx

nx_flash_byte_exact_gate.nx -- TWO-PASS BYTE-EXACT flash attention: the determinism+memory exceed. The team's one-pass online-softmax flash (nx_flash_attention) is memory-efficient but only EPS-matches naive (eps_q10=100) because the online rescale `acc*exp(m_old-m_new)/SCALE` ROUNDS in fixed-point. This rung proves a TWO-PASS flash -- pass-1 finds the GLOBAL row-max, pass-2 recomputes exp(s-gmax) and accumulates EXACTLY (no rescale) -- is BYTE-EXACT == naive AND block-size-INVARIANT (any tile size -> identical output), while keeping O(block) score memory (not O(n^2)). = byte-exact AND memory-efficient attention -- the combination float FlashAttention CANNOT do (its parallel float reductions are non-deterministic). criteria: 1 two-pass(block=2) == naive BYTE-EXACT (0 mismatches) 2 two-pass(block=1) == naive BYTE-EXACT 3 two-pass(block=4=full) == naive BYTE-EXACT (=> block-size-INVARIANT = the determinism exceed) 4 output is the known non-trivial answer [[17,1],[21,2]] (real attention, not all-zeros) 5 memory win: two-pass score scratch (block) < naive score scratch (n_kv) at byte-exactness expect_exit: 0 license_tier: ORIGINAL