Export Operator Miss Boxes

Stage 1 of the operator miss-annotation export (design round-0 Q4, round-2 R1.1): NDJSON with two record types, no fold decided here. record_type="box" carries each active operator-drawn annotation. Trailing record_type="camera_history" records carry source_camera_id, the box-time site_id/parent_id snapshot and first_boxed_at for every ever-boxed camera (one record per site it was boxed under), including those with only deleted boxes or no surviving camera/event. All timestamps are UTC ISO-8601. Read the COMPLETE export before assigning folds; history preserves permanent camera restrictions, and only box records are training labels. Owner-only via the router-level dependency above -- the offline fine-tune tooling (stage 2, not in this repo) reads this, resolves train vs. eval-nomination per camera against camera_folds.json on S3, and extracts frames from each box's clip_s3_key at frame_timestamp_ms.