nx_nofloat_linattn.nx

nx_nofloat_linattn.nx -- SOVEREIGN no-float LINEAR ATTENTION recurrent kernel = the foundation of Kimi Delta Attention (KDA, the K3 namesake; operator 2026-07-19 "state of the art"). Linear attention replaces softmax(QK^T)V with a RECURRENT matrix state: S_t = S_{t-1} + k_t (x) v_t (outer product, d x d); o_t = q_t S_t. This is O(1) memory per step (no growing KV cache) -> the 75% KV cut + 6x decode @1M that KDA delivers. Done in PURE INTEGER Q16 (outer product + matvec, accumulate-then-shift) = BIT-EXACT deterministic linear attention -- a float linear-attn stack drifts by accumulation order; ours does not. KDA extends THIS with (a) the delta rule S_t = S_{t-1}(I - b k k^T) + b k v^T and (b) channel-wise gating alpha_t (diagonal forget) -- both follow-ons on this proven kernel. Grounded arXiv 2510.26692 (Kimi Linear). license_tier: ORIGINAL No hw writes (Rule 26).