nx_nofloat_block_gate.nx

nx_nofloat_block_gate.nx -- HARD-EVIDENCE gate for a COMPLETE pre-norm TRANSFORMER BLOCK backprop + train, in PURE INTEGER Q16 (CAP-NF-BLOCK). The block (single-head) is the real Qwen-style shape: H = X + Wo * Attn( RoPE, scaled, causal-softmax )( rmsnorm_rows(X) ) [attention sublayer + residual] Y = H + SwiGLU-FFN( rmsnorm_rows(H) ) [FFN sublayer + residual] SwiGLU-FFN(z) = Wd * ( silu(Wg*z) (*) (Wu*z) ) Composed ENTIRELY from gradcheck-verified ops (rmsnorm_rows, matmul, matmul_nt, rope, cmul, softmax_rows, silu, hadamard, vadd, mse). No new backward math -- this proves the COMPOSITION trains. A1 hadamard gradcheck : loss=mse(a(*)b,t); dL/da == finite diff. A2 rmsnorm_rows gradcheck: per-token norm; dL/dx == finite diff. A3 FULL-BLOCK gradcheck wrt Wd (down-proj, short path): dL/dWd == finite diff. A4 FULL-BLOCK gradcheck wrt Wq (the LONGEST chain: through FFN, residual, attention, RoPE, softmax): measured. B the BLOCK TRAINS : with all weights fixed except the down-proj Wd, Y is LINEAR in Wd (convex MSE) -> train Wd from zero to a realizable teacher target; assert loss collapses + Wd converges. A weight DEEP in a full transformer block is driven by the block's own end-to-end backward. C bit-exact ; D neg-control teeth. Evidence -> knowledge/status/nofloat_block.log. Sovereign: nx_nofloat_autograd + nx_syscalls (pure integer). HONEST scope: single-head; multi-head/GQA + multi-block stack are the remaining COMPOSITION. expect_exit: 0