nx_nofloat_q4k_gate.nx

nx_nofloat_q4k_gate.nx -- the REFEREE for nx_nofloat_q4k (LM4 resident-quant decode, /compare/engineshift 2026-09-02). In-process teeth that need no model: a hand-built Q4_K super-block matrix with d=1.0, dmin=0, every scale 1 and every min 0 makes each dequantised weight EXACTLY its 4-bit code, and the activation is chosen so the dynamic i16 pack is exact, so nq_mm's Q16 output has ONE correct value that an independent i64 reference computes (both shift regimes, and a full-width 18944 row for the overflow headroom); the route decision is a pure function tested on every branch (Q4_K aligned -> in place; Q6_K -> i16; a non-256 width -> i16, never a silent in-place read); the i16 route must equal mm_pool_i8 on the same inputs; the scale must never let max|x| wrap an i16 lane. Neg-control: one post-shift off must NOT match the reference. T0 (gv_need the Q4_K_M fixture) censuses the fixture's tensor types so the model-level serve gate can only claim the fused path where Q4_K tensors exist -- fixture-reached-the-condition. nx_nofloat_q4k_gate [model.gguf] default /home/elderwesto/nx_stage/nx_15b_model.gguf license_tier: ORIGINAL. Writes nothing. No hw writes.