Score an agent-harness project across instructions, state, verification, scope, and lifecycle, then recommend improvements.