
M=512: activation reaching the quant boundary vs silu_mul_reference: 84/8912896 fp16 bits differ (0.000942%)
       NULL CONTROL, the eager chain:              7909531/8912896 fp16 bits differ (88.742548%)
       FUSED PROLOGUE int8 vs served-activation int8: 10/8912896 (0.000112%), max step 1, scale max rel 0.00e+00
       NULL CONTROL int8 (eager glue):                513239/8912896 (5.758386%)  -> 51324x the fused miss

M=2048: activation reaching the quant boundary vs silu_mul_reference: 307/35651584 fp16 bits differ (0.000861%)
       NULL CONTROL, the eager chain:              31631941/35651584 fp16 bits differ (88.725205%)
       FUSED PROLOGUE int8 vs served-activation int8: 29/35651584 (0.000081%), max step 1, scale max rel 0.00e+00
       NULL CONTROL int8 (eager glue):                2054906/35651584 (5.763856%)  -> 70859x the fused miss

PASS: the fused prologue reproduces the served activation's int8 on the real module with the real checkpoint's suh; the eager chain does not.
