nx_nofloat_arch_moe_gate.nx

nx_nofloat_arch_moe_gate.nx -- gates the arch-config debt eaten 2026-07-15: EXACT attention scale + REAL rope base from metadata + fx_ln_int, across the fleet's 3 real models. Teeth: T1 fx_ln_int KATs: ln(1e6)~905419, ln(1e4)~603597, ln(2)~45426, ln(1)=0. Tol +-20 Q16-ulps = the Q16 fixed-point floor (~2e-4 rel), NOT slack: after the rope /np division + exp it is negligible, and llama.cpp computes these in f32 with comparable relative error. Exactness would need a Q24/Q32 internal accumulate -- a real but zero-value-for-rope rung, deliberately not spent here. T2 EXACT scale: hd=64 -> 8192 (IDENTITY, unchanged), hd=128 -> 5792 (was 5957 = ~2.8% high), hd=96 exact T3 Qwen2.5-0.5B config UNCHANGED: base=1e6, scale=8192 (identity -- serve/gate reproduce byte-for-byte) T4 Qwen2.5-1.5B: base=1e6, hd=128 -> scale 5792 (the correction lands on the 1.5B's real head_dim) T5 OLMoE: base=1e4 (NOT 1e6 -- the hardcoded default would have garbled OLMoE rope), ln(base) sane Reads real ggufs (arch config only, no forward) -- fast. expect_exit: 0 license_tier: ORIGINAL