nx_nofloat_scale_ffn_gate.nx

nx_nofloat_scale_ffn_gate.nx -- R4 FINAL test: the COMPLETE transformer block (attention + FFN) on the richer grammar. All prior scale attempts were ATTENTION-ONLY (no FFN) -- but a real transformer block has an FFN, the per-token nonlinear map that can turn a token's category into a SHARP output distribution. Honesty requires testing the complete architecture before concluding the floor is unreachable. Strongest single shot = FFN + mini-batch averaging (lowest-noise training). Richer 4-cat grammar (vocab 16), dm=32, ffn=64. floor = avg(ln2,ln4,ln5,ln5)=1324 milli-nats (ppl 3.76); uniform = ln(16)=2773; attention-only plateaued ~2135. T1 held-out CE << uniform. T2 held-out CE ~= floor (near-OPTIMAL -> the FFN was the missing piece -> R4 lands). If T2 fails too, the tractable levers (optimizer x6 + complete architecture) are EXHAUSTED -> R4-FULL genuinely needs compute-scale (operator-gated), the honest end of the road. expect_exit: 0 Sovereign: nofloat_autograd.