nx_nofloat_bpe_lm_gate.nx

nx_nofloat_bpe_lm_gate.nx -- CAPSTONE: the WHOLE sovereign efficient text pipeline wired END-TO-END, pure no-float Q16. Composes R2 (learned BPE, nx_nofloat_bpe_gate) + R1 (the no-float transformer LM): text -> learn K BPE merges -> compress to sub-word tokens -> remap to a compact vocab -> the no-float LM trains to model the token stream -> autoregressive argmax generation -> decode tokens back through the merges to bytes -> the ORIGINAL text. T1: the LM learns the BPE-token stream (argmax generation reproduces every token). T2: END-TO-END LOSSLESS -- the decoded generation equals the original text, byte-exact (the full tokenize->train->generate->detokenize loop is correct). T3: EFFICIENCY -- the BPE token stream is shorter than the char stream (fewer tokens to model). Sovereign: nx_nofloat_autograd + nx_syscalls. Reuses verified pieces (BPE merge-learn, clm_fwd attention LM, fixed-T argmax decode, recursive merge-expand). expect_exit: 0