nx_nofloat_landing_gate.nx

nx_nofloat_landing_gate.nx -- THE MEASURED LANDING (roadmap R4 increment; finding 7 = perplexity is the standard cheap held-out metric, nfs_perplexity.raw). A no-float LM is "landed" on a domain when its HELD-OUT per-token cross-entropy reaches the domain's INFORMATION FLOOR (the irreducible entropy) and is far below the uniform baseline. Domain = the grammar [ARTICLE,NOUN,VERB] (fresh-random members), whose intrinsic per-token entropy = avg(ln2, ln3, ln3) ~= 0.964 nats (perplexity ~= 2.62): an article position has 2 equally likely members, noun/verb positions have 3 -- the members are RANDOM so NO model can beat that floor; the uniform-over-vocab baseline is ln(8) = 2.079 nats (perplexity 8). We train on fresh-random streams (so the model learns the GRAMMAR, not sequences) and measure CE on NEVER-SEEN streams. T1 held-out CE << uniform ln(8) (the model beat the baseline = it learned). T2 held-out CE ~= the information floor (near-OPTIMAL = landed: it reached the domain's entropy limit). Pure integer Q16. Sovereign: nx_nofloat_autograd + nx_syscalls. expect_exit: 0