Read and translate Japanese manga lettering into Korean.
The image marks requested regions with numbered rectangles. Use those exact IDs.
Read only the writing inside each rectangle, using the full page for context.
Never invent text from the illustration or borrow another region's writing.

Classify text_kind as one of dialogue, narration, title, nameplate, sign,
packaging, sfx, other, uncertain. Translate readable dialogue, narration,
titles, character nameplates and signs even when colored or hand-lettered.
Preserve product names and decorative text printed on packaging, even if large.
Readable warnings and instructions on products are signs to translate.
Preserve sound effects/onomatopoeia, non-text, publisher/legal credits and
unreadable text. If you cannot read the exact writing, preserve it.

For each translatable region, FIRST transcribe the exact Japanese into
ja_source, preserving every word, name, number and negation. THEN translate
that transcription into natural Korean in ko_translation. Do not substitute
similar Japanese words. Preserve meaning, commands, numbers, names, quotations
and speaker relations. Korean must contain no Japanese kana or explanations.
For preserved regions use no_translate=true, ja_source=null,
ko_translation=null. For translated regions use no_translate=false.

Return only a JSON object with the translations property. Each element must
be exactly [id, text_kind, ja_source, Korean_translation_or_null]. Transcribe
the Japanese first, then translate that exact transcription. For preserved
regions, return null for both ja_source and Korean_translation_or_null.
Include every requested integer id once, in numeric order. No dictionaries,
booleans, markdown or extra fields.
