dubai-02-wave2: the pre-registered headline with R2 on (as published) and off

R2 on (reproduces the committed q1_surface_overlap.csv byte for byte):
  40 questions, 115 surface pairs compared, 5 excluded (G4 chatgpt-gemini, G4 chatgpt-ai_mode, G4 gemini-ai_mode, T3 chatgpt-ai_mode, T3 gemini-ai_mode)
  below the mean of their two split-halves: 104 of 115 = 90.4%
  median ratio of cross-surface overlap to that mean: 0.652
  threshold: replicates at share >= 70% and ratio <= 0.80; fails at share < 50% or ratio >= 0.90
  verdict: REPLICATES
R2 off (decided returns the entity key unchanged):
  40 questions, 115 surface pairs compared, 5 excluded (G4 chatgpt-gemini, G4 chatgpt-ai_mode, G4 gemini-ai_mode, T3 chatgpt-ai_mode, T3 gemini-ai_mode)
  below the mean of their two split-halves: 105 of 115 = 91.3%
  median ratio of cross-surface overlap to that mean: 0.652
  threshold: replicates at share >= 70% and ratio <= 0.80; fails at share < 50% or ratio >= 0.90
  verdict: REPLICATES

firms named anywhere: 188 with R2, 192 without
firms R2 adds to runs: 2, in 232 run credits; of them never named by their own name in any run: 2
firms R2 removes from runs: 6, in 260 run credits

Note: the pre-commitment (524e5d9, left as pushed) says R2 "moves 260 run credits". That mixes two
directions: R2 adds 2 firms to 232 run credits and removes 6 from 260, as above. The computation it
fixed is unaffected.
