Serve SurjoLabs/Surjo-50m-SFT-Only (int8, 2T) on 7860

deploy c66f5aa: decode push

Space: AVX2 int8 restored as default (zmm opt-in); log kernel variant

Space: AVX512 kernels (zmm fp32/int8, parallel VNNI-Q8) + 2T cap + NEON int4 fix

Fresh Surjo-50m bench on current code (all 4 quants) + HW inventory

CISM remote autobench: full source + fleet runner, serve results on 7860