nx_nofloat_xformer_gate.nx

nx_nofloat_xformer_gate.nx -- HARD-EVIDENCE gate for the TRANSFORMER-SUBLAYER backward ops added to the no-float autograd (softmax / silu / rmsnorm): the nonlinearities of a Qwen block, now DIFFERENTIABLE in pure integer Q16. This is the CAP-NF-TRAIN-XFORMER rung: backprop through the transformer's own ops. A GRADCHECK each op (oracle = mathematics): build loss = mse(OP(x), target), compare the tape's analytic reverse-sweep grad dL/dx_i against a CENTRAL finite difference (L(x+h)-L(x-h))/2h, all Q16. Done for softmax (Jacobian-vector product), silu (x*sigmoid(x) derivative), rmsnorm (normalization Jacobian). D TEETH (neg-control): negate one analytic grad and assert the SAME check now FAILS -> a wrong gradient is provably caught (the agreement in A is real, not vacuous). B A TRANSFORMER ACTIVATION TRAINS: fit y = silu(w*.x) (realizable, w*=1.5) from w=0 by GD through the tape+silu backward. Assert loss collapses >=90% and w converges to w*. Proves the FFN nonlinearity is trainable end-to-end in pure integer. C BIT-EXACT: train twice from zero -> identical integer w (determinism is structural for integer). Evidence -> knowledge/status/nofloat_xformer.log. Sovereign: imports nx_nofloat_autograd (pure integer; NO nx_f32 in the import graph) + nx_syscalls. license_tier: ORIGINAL expect_exit: 0