nx_nofloat_adamw32_gate.nx

nx_nofloat_adamw32_gate.nx -- the REAL win from the R4 scale-up work: a numerically STABLE fixed-point Adam. DIAGNOSIS: the Q16 AdamW (lib nfa_adamw) DIVERGES/STALLS on small gradients because g^2 UNDERFLOWS in Q16 -- e.g. g=0.06 (Q16 3932) -> qmul(g,g)=235 (0.0036) and the (1-b2)*g^2 increment qmul(66,235)=0, so the 2nd moment v never accumulates, the denominator collapses to eps, and the update mhat/eps explodes (or stalls). FIX: accumulate v in Q32 and compute g^2 = g*g at FULL precision (no >>16) so small gradients survive. T1: Q32-moment AdamW CONVERGES on a small-gradient fit (loss -> ~0). T2 (measured exceed): Q32 final loss << Q16 final loss on the SAME problem (the underflow fix is real). Self-contained controlled experiment (n=8 fit, hand-coded grad 2*(w-t), no transformer). Pure integer. expect_exit: 0 Sovereign: nx_nofloat_autograd (for nfa_adamw + nfa_isqrt) + nx_syscalls.