nx_nofloat_multihead_gate.nx

nx_nofloat_multihead_gate.nx -- HARD-EVIDENCE gate for MULTI-HEAD attention in pure integer Q16. The defining transformer feature: split Q/K/V into H heads (slice_cols), run the verified attention core per head (RoPE + scaled causal softmax + A.V), concat the heads (concat_cols), then the output projection. Backprops end-to-end. A1 slice_cols gradcheck : extract cols backward (scatter) == finite differences. A2 concat_cols gradcheck : concat backward (split) == finite differences. A3 multi-head gradcheck wrt Wq : the gradient through BOTH heads (slice+RoPE+softmax+A.V+concat+proj) == FD. D neg-control teeth ; C bit-exact. B multi-head LEARNS : with attention fixed, the output projection Wo is trained (AdamW) to a realizable target; loss collapses -> the multi-head output path trains. Evidence -> knowledge/status/nofloat_multihead.log. Sovereign: nx_nofloat_autograd + nx_syscalls. expect_exit: 0