nx_nofloat_vs_float_gate.nx

nx_nofloat_vs_float_gate.nx -- CROSS-VALIDATION: do we need the float equivalent? Answer: it already EXISTS (nx_autograd, the software-f32 scalar tape), and its right role is a VERIFICATION ORACLE + the exceed COMPETITOR -- NOT a parallel product stack. This gate USES that role: it builds the SAME graph loss = ( relu(w1*x + b1)*w2 + b2 )^2 in BOTH engines -- FLOAT via nx_autograd (ag_*, IEEE-754) and NO-FLOAT via nx_nofloat_autograd (nfa_*, integer Q16) -- and asserts the four parameter gradients AGREE (in milli-units). Two INDEPENDENT autograd implementations (different tapes, different number systems) computing the same gradients is far stronger evidence than finite differences alone. It directly answers the question: the float engine validates the no-float engine; we do not need to BUILD a float trainer (we have it) and do not need it in the PRODUCT (no-float is the reproducibility exceed) -- we need it exactly here, as the oracle. T1..T4: for each of {w1,b1,w2,b2}, |float_grad_milli - nofloat_grad_milli| < 60/1000. The true grads (analytic): w1=3.094, b1=4.125, w2=2.063, b2=2.750 -> both engines land on them. expect_exit: 0 license_tier: ORIGINAL imports: nx_autograd (->ag_*, ->nx_f32_*), nx_nofloat_autograd, nx_syscalls