Phase Attribution

Average findings caught per run, split by pipeline phase (bar width = volume). Which phase — TEA, Dev, or Reviewer — actually catches the defects?