tools/bench/bench_http.py --concurrency 8 --reps 4 --max-tokens 512, two boots per arm
step rate = tok_s / (1 + K*accept), K=7. Zero degenerate rows in any cell.

arm                   n    tok/s    sd   accept   steps/s    sd
OFF replicated        8    57.51  3.73   0.3122    18.039 0.545
ON base               8    56.33  3.16   0.3069    17.887 0.457
ON +B                 8    57.85  3.24   0.3112    18.190 0.467
ON +B +C              8    56.78  2.52   0.2992    18.347 0.350

paired per (boot index, rep) -- the harness rotates prompts by rep index,
so the same prompt set faces each arm in the same cell:

  ON +B            vs ON base         : +1.73% +-2.78 sd, sem 0.98, 5/8 positive
  ON +B +C         vs ON +B           : +0.90% +-2.45 sd, sem 0.86, 7/8 positive
  ON +B +C         vs ON base         : +2.59% +-1.41 sd, sem 0.50, 8/8 positive
  ON base          vs OFF replicated  : -0.81% +-2.00 sd, sem 0.71, 2/8 positive
  ON +B +C         vs OFF replicated  : +1.74% +-1.36 sd, sem 0.48, 7/8 positive
