• New: a history of what FaceID actually reported — with the face image it used. New
  HISTORY tab. Every published recognition is kept with its crop, score, camera, zones
  and the embedding behind it.
• This exists because of a concrete failure. Asked to verify a suspected misidentification,
  I pulled the Frigate snapshots for the events in question — and one of them showed a
  different person than the one that had been reported. The recognition turned out to be
  right: Frigate replaces an event's snapshot while the event runs, and the picture it
  serves later is usually a different moment. Reconstructing the truth took the recording,
  frame by frame, for a single event. That is not a workable answer to "was this wrong?".
• "was wrong" does the analysis for you. One click computes which reference photo
  carried that match, what the score would have been without it, and who would have won
  instead. If a single photo is responsible — its similarity far above every other photo of
  that person — FaceID offers to set exactly that one aside. Reversible, with the reason
  recorded, like every other set-aside.
• If no single photo is responsible, it says so instead of suggesting a fix that would not
  help: the two people are then genuinely close, and the answer is more references, not
  fewer.
• Kept to the last history_keep entries (default 200, 0 disables). A crop is a few KB.
• Unknown reports are recorded too — they are published messages like any other, and
  "why was I reported as a stranger?" is the same question.