nx_nofloat_scale_wide_gate.nx

nx_nofloat_scale_wide_gate.nx -- R4 the genuine SCALE test: real WIDTH scale-up. Prior R4 tests stayed at dm=24/32 and ruled out optimizer/depth/batch/FFN -- but WIDTH (features per token) is distinct from depth and is exactly what producing SHARP 4-way-conditional distributions needs. Maximal tractable model: dm=64, ffn=128, the COMPLETE block (attn+FFN), mini-batch averaged grads (lowest noise). Richer 4-cat grammar. floor = avg(ln2,ln4,ln5,ln5)=1324 milli-nats (ppl 3.76); uniform=2773; dm=32 plateaued ~2135 (ppl 8.5). T1 held-out CE << uniform. T2 held-out CE ~= floor (near-OPTIMAL -> WIDTH-scale closes it -> R4 progresses). If T2 fails too, even 2x width is insufficient at tractable compute -> R4-FULL needs massive (GPU-scale) compute. expect_exit: 0 Sovereign: nx_nofloat_autograd + nx_syscalls. (run in background -- longer compute)