nx_nofloat_scale2_gate.nx

nx_nofloat_scale2_gate.nx -- R4 DEPTH test: does a 2-BLOCK no-float LM reach the richer-grammar floor where the single block plateaued (ppl ~8.5)? Empirical (the lesson: TEST, don't assert). Same richer grammar (DET2->ADJ4->NOUN5->VERB5, vocab 16) + dm=32, but TWO stacked pre-norm attention blocks. SGD lr=0.2 (the best single-block optimizer config). Floor = avg(ln2,ln4,ln5,ln5)=1324 milli-nats (ppl 3.76); uniform=2773. T1 held-out CE << uniform (learned). T2 held-out CE ~= floor (near-OPTIMAL -> R4 lands). If T2 fails too, depth is NOT the lever and the boundary is genuinely optimization/compute (honest stop). expect_exit: 0 Sovereign: nx_nofloat_autograd + nx_syscalls.