=== v10: row gate inside the op, box100 sole tenant, 0 errors, 14 dynamo graphs ===
[i8p] shipped row threshold stamped onto 64 MLPs: [280]
  [0] code           REF     16384tok  wall=  4.96s  finish=length
  [0] code           INT8F   16384tok  wall=  3.38s  finish=stop
  [0] code           INT8P   16384tok  wall=  3.46s  finish=stop
  [1] code           REF     16384tok  wall=  4.74s  finish=stop
  [1] code           INT8F   16384tok  wall=  3.46s  finish=stop
  [1] code           INT8P   16384tok  wall=  3.78s  finish=length
  [2] wiki           REF     16384tok  wall=  4.95s  finish=length
  [2] wiki           INT8F   16384tok  wall=  3.68s  finish=length
  [2] wiki           INT8P   16384tok  wall=  3.79s  finish=length
  [3] math           REF     16384tok  wall=  4.96s  finish=length
  [3] math           INT8F   16384tok  wall=  3.67s  finish=length
  [3] math           INT8P   16384tok  wall=  3.79s  finish=length
  [4] longctx        REF     16384tok  wall=  4.96s  finish=length
  [4] longctx        INT8F   16384tok  wall=  3.69s  finish=length
  [4] longctx        INT8P   16384tok  wall=  3.79s  finish=length
  [5] sci            REF     16384tok  wall=  4.97s  finish=length
  [5] sci            INT8F   16384tok  wall=  3.68s  finish=length
  [5] sci            INT8P   16384tok  wall=  3.80s  finish=length
[i8p] prologue calls=1600  fallbacks: not_serveable=0 lora=0 pad=0 act=0 noncontig=0
[i8p] small-M calls routed to the shipped leg A (correct, not a fault): 2688
[i8p] dynamo unique graphs at exit: 14

medians: REF 4.960  INT8F 3.675  INT8P 3.790   REF/INT8F 1.350x  REF/INT8P 1.309x  INT8P vs INT8F +3.1%

=== per-kernel accounting, one profiled 16384-tok prompt, same prompt both arms ===
kernel                                       F n     F ms   P n     P ms  delta ms
exl3_i8_gemm_kernel                         1600  1859.08  1600  1859.01     -0.07
had_hf_r_128_kernel                         1600   261.77  1600   261.19     -0.58
a_prep_kernel<2,32> (K=5120)                1344   114.20  1344   114.01     -0.19
a_prep_kernel<5,32> (K=17408 down_proj)      256    60.39   256   109.58    +49.19
glue ..._mul_silu_10                         192    89.47     0     0.00    -89.47
glue ..._mul_silu_2                           64    29.00     0     0.00    -29.00
cast ..._mlp_down_10                           0     0.00   384   118.22   +118.22
cast ..._mlp_down_2                            0     0.00   128    39.28    +39.28

total device time INT8F 3361.5 -> INT8P 3436.7 = +75.2 ms (+2.2%); wall +3.1%
glue removed -118.47 | casts added +157.50 | prologue a_prep +49.19 (same 256 calls)
NOTE: the cast kernels run at EXACTLY 2x the call count of the glue they replace (384 vs 192, 128 vs 64)
