Drop the first export: its model_sha carried a run commit, not the eval code commit

Add benchmarks.json: benchmark metadata beside the results, so the UI stays dumb

User simulation leaderboard: results