nx_nofloat_prose_gate.nx

nx_nofloat_prose_gate.nx -- CAP-NF-PROSE: the no-float char-LM trains on REAL ENGLISH PROSE (words + spaces, not the alphabet) at a LARGER scale (dm=24/ffn=48). Real prose is genuinely harder than a periodic sequence: the space char precedes many different words, so the model MUST use causal-attention history to disambiguate. Trains the full pre-norm transformer block (all weights) with clipped SGD. Honest: an INCREMENTAL step into CAP-AI-FRONTIER (real corpus + scale), not a claim to finish it. T1 started untrained (CE > 0) ; T2 CE drops substantially (>= 60%) ; T3 next-char accuracy >> chance (1/V) + prints the model's predicted text vs target (visible evidence it learned the prose). Lean: 2 clm_fwd call sites (train-loop + accuracy/print). Sovereign: nx_nofloat_autograd + nx_syscalls. expect_exit: 0