Restructure testset: rename output fields, add golden_trajectories, drop annotation labels

Add 800-query evaluation set (expert-annotated trajectories) and dataset card

initial commit