The runner that was larger than the job needed

An eight-core runner assigned to a job that is network-bound, at four times the cost for no gain.

the audit, per job:

  job              runner   peak CPU   verdict
  test             8-core     780%     correct
  assets           8-core     190%     over-provisioned
  docker build     8-core     120%     network-bound
  lint             8-core      95%     one core
  deploy           8-core      12%     absurd

four of five moved back to the standard runner.
runner minutes: unchanged. cost: -38%.

A larger runner was applied to every job when it was applied to one, which is how a per-job decision becomes a default. The peak CPU figure is the whole diagnostic and it is available from the job summary — five minutes of reading against a bill that had been growing for a year.