rows=2048 K=17408  (Qwen3.8-27B down_proj intermediate = 17408)
seed 0:
  FUSED      int8 mismatch          6 / 35651584 ( 0.000017%)  max step 1   scale: 1 rows differ, max rel 4.339e-07
  EAGER*     int8 mismatch    1879890 / 35651584 ( 5.272949%)  max step 1   scale: 2048 rows differ, max rel 3.163e-03
  torch-fp32 int8 mismatch          6 / 35651584 ( 0.000017%)  max step 1   scale: 1 rows differ, max rel 4.339e-07
  FRAG       bytes differing vs shipped frag: 6 / 35651584
  -> null control fired: eager misses 1879890 where fused misses 6 (313315x)
seed 1:
  FUSED      int8 mismatch          4 / 35651584 ( 0.000011%)  max step 1   scale: 0 rows differ, max rel 0.000e+00
  EAGER*     int8 mismatch    1892952 / 35651584 ( 5.309587%)  max step 1   scale: 2048 rows differ, max rel 3.350e-03
  torch-fp32 int8 mismatch          4 / 35651584 ( 0.000011%)  max step 1   scale: 0 rows differ, max rel 0.000e+00
  FRAG       bytes differing vs shipped frag: 4 / 35651584
  -> null control fired: eager misses 1892952 where fused misses 4 (473238x)

verdict: fused is inside the torch-fp32 band
