v0.66.0

Changed

• Breaking (behaviour): the LLM refiner's prompt-token estimate comes from FluxCurator's per-character-kind estimate (o200k_base ratios) instead of a local
  rule — Korean text counts about 2.5× more tokens than before (it was read near English), English about the same. No code change;
  limits compared with these counts are reached sooner on Korean text.

Dependencies

• FluxCurator 0.18.1 -> 0.19.0 (its per-character-kind token estimate; chunk sizes follow it).