nx_nofloat_adamw_gate.nx

nx_nofloat_adamw_gate.nx -- HARD-EVIDENCE gate for the Q16 AdamW optimizer (nfa_adamw), with a MEASURED exceed over plain SGD. On an ILL-CONDITIONED linear regression (two features of very different scales: col0 ~ O(1), col1 ~ O(1/40)), one global SGD learning-rate either crawls on the small-curvature direction or diverges on the large one. AdamW normalizes each parameter's step by its own running gradient scale, so it converges FAR faster. Pure integer Q16 -> bit-exact/deterministic. T1 AdamW converges : from zero, AdamW drives the SSE loss to ~0 (it works). T2 MEASURED exceed : at the SAME step budget, AdamW's final loss is MUCH lower than tuned SGD's (and AdamW reaches the loss threshold in FAR fewer steps). Numbers printed. T3 SGD really struggles: SGD's final loss on this conditioning stays high (the problem is genuinely hard for one global rate) -- so T2 is a real win, not a strawman. T4 bit-exact : two AdamW runs give IDENTICAL integer weights (determinism). Evidence -> knowledge/status/nofloat_adamw.log. Sovereign: nx_nofloat_autograd + nx_syscalls. expect_exit: 0