nx_nofloat_attn_gate.nx

nx_nofloat_attn_gate.nx -- HARD-EVIDENCE gate for the ATTENTION-CORE backward (CAP-NF-ATTN-CORE): a single-head CAUSAL self-attention block backprops end-to-end in PURE INTEGER Q16. This is the heart of CAP-NF-TRAIN-ATTN. New ops proven: matmul (C=A.B), matmul_nt (S=Q.K^T), cmul (1/sqrt(d) scale), causal softmax_rows. A1 matmul gradcheck : C=A.B, loss=mse(C,t); tape grad dL/dA == central finite difference. A2 matmul_nt gradcheck : S=A.B^T (the Q.K^T form); tape grad == finite difference. A3 attention-core gradcheck: full path X->{Q,K,V}=X.W -> S=Q.K^T -> scale -> CAUSAL softmax rows -> O=A.V -> mse; gradcheck dL/dWq, the gradient that flows THROUGH softmax + both matmuls (the real attention backward). D neg-control teeth : a deliberately wrong matmul grad is rejected. B attention LEARNS : with Wq,Wk fixed (attention pattern A fixed), the VALUE path O=A.(X.Wv) is linear -> train Wv from zero to a realizable target; assert loss collapses + Wv converges (the value projection trains). C bit-exact : train twice -> identical integer Wv (determinism is structural for integer). Evidence -> knowledge/status/nofloat_attn.log. Sovereign: imports nx_nofloat_autograd (pure integer) + nx_syscalls. HONEST scope: this is the attention CORE (matmuls + causal softmax). RoPE + output-projection + multi-head are the next sub-rung (CAP-NF-TRAIN-ATTN full). license_tier: ORIGINAL expect_exit: 0