Add production_vertical rows; shard the paired dataset

Add paired ablation samples (one row per question + its reformulations)

initial commit