nx_nofloat_scale_batch_gate.nx

nx_nofloat_scale_batch_gate.nx -- R4 retry via the ONE untested fundamental lever: MINI-BATCH GRADIENT AVERAGING. All prior richer-grammar training was BATCH-1 (one fresh-random stream/step = very noisy gradients -> SGD oscillates, can't sharpen to the floor). Standard fix: average gradients over a mini-batch before stepping. Same richer 4-cat grammar (DET2->ADJ4->NOUN5->VERB5, vocab 16) + dm=32 that plateaued at CE ~2135 (ppl 8.5); here we accumulate grads over B=16 fresh streams per step (low-noise), then ONE SGD step. floor = avg(ln2,ln4,ln5,ln5)=1324 milli-nats (ppl 3.76); uniform = ln(16)=2773. T1 held-out CE << uniform (learned). T2 held-out CE ~= floor (near-OPTIMAL -> batch-averaging cracked it). If T2 fails too, batch size is NOT the lever and the tractable training levers are genuinely exhausted (honest). expect_exit: 0 Sovereign: nx_nofloat_autograd + nx_syscalls.