Sandboxed agent evals with Harbor

Harbor's native Braintrust plugin syncs task cases, verifier rewards, and agent trajectories into a project as each job finishes.