verify rows: 7551   drafted/row: 7.00   accepted/row: 2.162   committed/step-row: 3.162

accepted-length distribution (accepted drafts per row)
   0     2068   27.4%  ################
   1     1737   23.0%  #############
   2     1236   16.4%  #########
   3      784   10.4%  ######
   4      494    6.5%  ###
   5      309    4.1%  ##
   6      215    2.8%  #
   7      708    9.4%  #####

first rejection position (6843 rows rejected something, 90.6%)
  slot  0     2068   30.2% of rejections
  slot  1     1737   25.4% of rejections
  slot  2     1236   18.1% of rejections
  slot  3      784   11.5% of rejections
  slot  4      494    7.2% of rejections
  slot  5      309    4.5% of rejections
  slot  6      215    3.1% of rejections

the four cases, over 23171 TESTED slots (slots past a rejection are excluded)
  drafter right, draw kept it      -> no loss     13612   58.7%
  drafter right, draw REJECTED it  -> SAMPLING loss      813    3.5%
  drafter wrong, draw rejected it  -> DRAFTER loss     6030   26.0%
  drafter wrong, draw kept it      -> no loss      2716   11.7%
  -> of all rejections, 11.9% were tokens the drafter got RIGHT (sampling) and 88.1% were drafter misses
  NOTE: 342 of 7551 rows are GREEDY (temperature 0); the writer emits no probabilities for them, so they contribute agreement bits only and cannot show a sampling loss.

per-slot decomposition  (hit = draft token IS target argmax)
  hit%(all) is over every DRAFTED slot: unbiased drafter quality.
  hit%(tested) is over slots the prefix reached: conditioned on surviving,
  so it runs HIGH at depth and its n shrinks. Read the trend off hit%(all).

slot   n_all hit%(all)  tested hit%(tst)   E[p'top1|hit]    E[p'draft|miss]  alpha_obs
   0    7551     61.2     7551     61.2     0.8336+-0.0064     0.0994+-0.0043     0.726
   1    7551     53.3     5483     59.3     0.8331+-0.0079     0.0875+-0.0048     0.683
   2    7551     48.0     3746     59.8     0.8473+-0.0095     0.0848+-0.0058     0.670
   3    7551     43.5     2510     63.8     0.8797+-0.0102     0.0757+-0.0071     0.688
   4    7551     40.6     1726     67.1     0.9028+-0.0110     0.0804+-0.0099     0.714
   5    7551     36.9     1232     70.7     0.9091+-0.0127     0.0855+-0.0131     0.749
   6    7551     34.2      923     73.9     0.9339+-0.0128     0.0778+-0.0151     0.767

reading: hit%(all) falling with slot = the DRAFTER degrades with depth.
         hit%(all) flat and high but alpha_obs low = SAMPLING rejects tokens
         the drafter got right; the bucket table above says so directly.

value of one more accepted token per row: +31.6% tok/s at the measured accept length (3.16 committed), 40 us/token commit cost assumed
