Annotation ids the AI runner must answer. PR7b hands EXACTLY these to the runner; every
other annotation either has an answer above or is listed in skippedFieldIds.
Non-contact annotations carrying no prompt and no sub-field prompts, whatever their
generated_by_ai flag. There is no question to answer, so no answer row is written and they
are NOT sent to the AI runner — the user fills them by hand, as in v1. Surfaced so the
partition stays total and auditable.
Answers ready to persist — one per contact and per manual/static annotation.