nx_nofloat_grammar_gate.nx

nx_nofloat_grammar_gate.nx -- CAP-NF-GRAMMAR: the AFFORDABLE "landing" step the sovereign researcher identified (knowledge/research/2026-06-23-nofloat-affordable-land-roadmap.md, R1, grounded in TinyStories arXiv 2305.07759 + Textbooks arXiv 2306.11644: a tiny model is COHERENT iff the DOMAIN is constrained). Instead of memorizing one fixed sentence, the model learns a small GRAMMAR and GENERATES coherent NOVEL sentences. Grammar: a stream of [ARTICLE, NOUN, VERB] triples. Vocab 8: articles {the=0,a=1}, nouns {cat=2,dog=3,bird=4}, verbs {ran=5,sat=6,ate=7}. EVERY training stream is fresh-random members -> the model cannot memorize sentences; it must learn the CATEGORY GRAMMAR (article->noun->verb->article ...). T1 GENERATES coherent grammar: from a seed, free argmax-decode produces a stream whose every token is in the grammatically-correct category for its position (>= 90%, vs ~33% chance) = coherent NOVEL generation. T2 learned the grammar on HELD-OUT streams: teacher-forced next-token category accuracy >= 90% (>> chance). Pure integer Q16, attention-only block, 3 clm_fwd sites in separate helpers (nx_cc-lean). Sovereign. HONEST: a CONSTRAINED grammar domain (the affordable landing per the research), not open prose. expect_exit: 0