== rounds/memladder-crossover-rotated-2026-09-17/roundB-main-393ec39ba.jsonl  budget=128  metric=cpu  288 ok leg(s), 0 refused
   arm_order=rotating-by-rep
   load_before min/median/max 3.8/5.6/10.8   load_after 4.0/5.7/10.8

-- rotation, as banked --
   arm    pos0  pos1  pos2
   auto     32    32    32
   fold     32    32    32
   force    32    32    32
   BALANCED: every arm in every position equally often (counts [32])

-- paired per-rep difference, fold minus force (positive = force is cheaper) --
        m  reps   fold med  force med  med diff     med %     sign  p
      160    24       9.22      10.48    -1.250   -13.56%    0/24   0.0000
      192    24      10.37      10.58    -0.210    -2.03%    0/24   0.0000
      224    24      11.52      10.68    +0.830    +7.21%   24/24   0.0000
      256    24      12.66      10.79    +1.870   +14.77%   24/24   0.0000

-- crossover --
   fold wins (significant):  [160, 192]
   ties (sign test n.s.):    none
   force wins (significant): [224, 256]
   => crossover is between m=192 and m=224

-- position effect (arm level and rep level divided out; what running Nth is worth) --
        m          pos0           pos1           pos2
      160        +0.00%         +0.00%         -0.05%   pos0-pos1=-0.00%
      192        -0.01%         -0.00%         -0.01%   pos0-pos1=-0.01%
      224        +0.01%         -0.02%         +0.00%   pos0-pos1=+0.04%
      256        +0.03%         -0.05%         -0.01%   pos0-pos1=+0.07%
   The levels are centred, so the READABLE quantity is the GAP: pos0-pos1 is what
   the first slot cost the arm that held it, and is the term a fixed order banks.
   (a FIXED-order round carries pos0-minus-pos1 as a systematic term it cannot see:
    on this driver until 17 Sep 2026 that term sat entirely on `fold`, which ran first)