Score up to SIGNAL_SCORE_BATCH_MAX texts against one product profile.

Body: product_id (required) — one profile for the whole batch items (required) — list of {id?, text, source_hint?, context?}, 1..MAX. context is the post or thread the text replies to (see /signal/score). id is echoed back untouched so you can line results up with your own records; we neither store nor interpret it. include_drafting_context (optional, default true) — set false for a pure triage pass. Drafting context costs a database read per qualifying item, and a caller sorting a scrape usually wants the verdicts first and the drafts later.

Returns one result per item IN THE ORDER SUPPLIED, plus a summary. A bad item fails on its own — it does not take the batch down with it, because losing 24 good verdicts to one malformed row is not an acceptable trade for a caller who is feeding us machine-generated input.

judgements_spent is the billing unit: panels actually convened. Text decided on content alone spends nothing, so this is normally well below the item count.