Reliability

This combines !1877 (closed) and !1882 (closed) and fills the gaps in between.

• I did not migrate data from showup to would_reinvite because that would be mostly guesswork
• Reliability of dropouts is still planned but not yet implemented
• The new scoring mechanism
  • bad: >=50% studies studies would not reinvite (exception: must be at least 2 studies)
  • medium: >=25% studies would not reinvite