nx_nofloat_olmoe_forward_gate.nx

nx_nofloat_olmoe_forward_gate.nx -- MoE rung 5 (2026-07-15): a COMPLETE OLMoE forward -- 16 stacked real layers + token embedding + final norm + LM head -> a REAL next-token from the real 64-expert model, in deterministic no-float. Lazy per-layer weight load (attn set into REUSED buffers; experts lazy) over the zero-copy mmap. Composes every eaten-debt piece: full-quant dequant, exact scale, real rope base 1e4, sys_map_file, nolmoe_attn/nolmoe_moe (rung 4). Teeth: T1 TOKENIZER round-trip: encode(prompt) -> decode == prompt (our BPE self-consistent on OLMoE's gpt2 vocab) T2 NON-DEGENERATE logits: argmax logit strictly > min logit (the head discriminates, not all-equal) T3 DETERMINISM: the full forward's predicted id + top logit byte-identical on repeat T4 ROUTING ALIVE AT DEPTH: >=2 distinct expert-masks seen across the 16 layers' last-token routing T5 PREDICTION real: a valid non-negative vocab id, decoded + printed (honest: OUR segmentation, uniform Q16 -- HF-canonical BPE + q24 precision + H2H = rung 6, deliberately not claimed here) Requires the OLMoE gguf. Heavy (~1-3 min: 16 layers x lazy dequant + 50304-row head). Return from main. No hw writes (Rule 26). expect_exit: 0 license_tier: ORIGINAL