nx_nofloat_moe_real_gate.nx

nx_nofloat_moe_real_gate.nx -- MoE rung 3 (2026-07-15): REAL MoE WEIGHTS, LAZILY. Loads a real sparse-MoE gguf (OLMoE-1B-7B: 64 experts, top-8, per-expert ff=1024, D=2048 -- the canonical small real MoE) and runs ONE real MoE-FFN through our integer machinery with LAZY PER-SELECTED-EXPERT dequant -- the design the operator's environment correction settled (dequant-once-everything is the WRONG shape for MoE; sparsity means only K experts' weights are touched per token, so fetch exactly those). Composes the proven pieces: nac_read_config (+ the MoE metadata keys), nx_gguf 3-D tensor info, dequant_row / dequant_to_q16 windows, mm_out_in / fx_exp / silu / qmul. HONEST SCOPE: this rung proves REAL-WEIGHT ROUTING + LAZY FETCH + THE MIX MECHANISM on blk.0 -- NOT a faithful OLMoE model output (full arch = QK-norm etc., a later rung). T1 MoE METADATA read from the model itself: expert_count=64, expert_used_count=8 (olmoe.* keys) T2 3-D expert tensors located, dims EXACT (gate/up [D,ff,E]; down [ff,D,E]; router [D,E] 2-D) T3 LAZY-SLICE LIAR-KILLER: bulk expert-slice dequant (type-aware block offset) == per-row dequant_row on sampled rows x sampled experts x gate+down, EXACT (0 mismatches) -- kills offset-math bugs T4 REAL ROUTING: top-8 = 8 distinct ids; deterministic repeat; a second input selects a DIFFERENT set (the real router discriminates -- the neg-control against constant routing) T5 REAL MIX: K experts lazily fetched (counter == 8), output dense-nonzero, byte-identical repeat Requires /home/elderwesto/nx_stage/nx_moe_model.gguf. ~4.4GB read; slices ~400MB. Return from main. No hw writes (Rule 26). expect_exit: 0 license_tier: ORIGINAL