nx_nofloat_weightconv_gate.nx

nx_nofloat_weightconv_gate.nx -- the BRIDGE: load REAL BF16 model weights -> no-float Q16 integer. To run a frontier open model (GLM-5 etc.) on the sovereign no-float stack we must convert its BF16/FP8 weights to our integer Q format -- and to keep the WEIGHTS themselves bit-exact reproducible, the conversion is PURE INTEGER (no float touches them). BF16 = top 16 bits of an f32: sign(1) exp(8) mant(7). value = (-1)^s * 2^(exp-127) * (1 + mant/128); Q16 = value*2^16 = (-1)^s * (128+mant) << (exp-118). Decoded with bit-shifts only -> DETERMINISTIC. Verified: synthetic BF16 (1.0/2.0/0.5/1.5/-1.0) convert EXACTLY, then a REAL BF16 tensor is read from a safetensors on disk (bounded read) + converted + sane. No hw writes (Rule 26). expect_exit: 0 license_tier: ORIGINAL