v0.6.6

0.6.6 — 2026-07-25

• match_top_k is now adjustable in Settings ("Photos averaged per match"). It was
  config-file only, yet it turns out to be the single biggest lever on recognition — and
  the sensible value depends on your gallery.
• Measured on a real 128-photo gallery (leave-one-out, so with ground truth), lowering it
  from 3 to 1 nearly doubled correct recognitions (36 → 65) with zero misassignments,
  and widened the margin to the runner-up. The reason: the score averages the k best
  matching photos, so a person whose references cover many different angles gets dragged
  down by her own less similar photos — punishing exactly the well-covered people.
• The practical probe now excludes self-hits: an event whose face is already in the
  gallery scores ~1.0 and measures nothing. It flattered small k badly — after excluding
  them on a k-independent basis (highest single similarity, not the k-mean), the honest
  comparison on one identical test set is 90% / 93% / 100% recognised for k = 3 / 2 / 1.
• The imbalance concern behind top-k was checked, not assumed: strangers (ignore anchors)
  peaked at 0.19 against a 0.50 threshold at every k, and no wrong person ever crossed
  the threshold. Measure your own with scripts/measure-recognition.py before changing.