nx_nofloat_generalize_gate.nx

nx_nofloat_generalize_gate.nx -- HELD-OUT RULE GENERALIZATION in pure integer Q16 (CAP-NF-GENERALIZE). The real test of learning vs memorization: train the no-float char-LM on a periodic sequence "abcabc..." (it must learn the first-order RULE a->b, b->c, c->a), then evaluate on data it NEVER trained on: T1 a HELD-OUT string "cabcab..." (SAME rule, DIFFERENT string) -> high next-char accuracy (>> chance 1/3). Succeeding on a SHIFTED string proves the model learned the CHAR-rule, not position-by-position memorization. T2 (teeth) a CONTROL "acbacb..." (DIFFERENT rule a->c,c->b,b->a) -> ~chance accuracy. Proves the model applies the SPECIFIC learned rule (it is not outputting constants or trivially-correct). Lean (nx_cc .s ceiling): exactly 2 clm_fwd call sites -- do_train (helper) + eval_acc (helper, called twice). HONEST: first-order periodic rule on a tiny task = a real but modest generalization (not deep in-context learning). Sovereign: nx_nofloat_autograd + nx_syscalls. expect_exit: 0