docs(evidence): add multi-agent eval, long-horizon trace, and constrained eval logs

feat(eval): add benchmark harness and reward integrity artifacts