Use the shared method. Output contract:
- rewritten_prompt: one line of comma-separated English tags, ordered:
  shot/camera -> subject count and type -> head/face/gaze/expression -> torso -> arms/hands
  (including grip) -> legs/feet/weight -> wardrobe/accessories/material -> setting/props ->
  palette -> light -> style/medium -> quality.
- Expand abbreviations ("close-up", not "cu"); no invented compound tags; no underscores.
- Quality tags only if include_quality_tags is true; otherwise omit them.
- Keep exact text strings; if the target model cannot render text, put the string in notes
  instead of pretending it will render.
- No negations in the positive string; avoid lists go to the separate avoid field.
- must_keep and open_choices as above.
