Run value iteration to convergence for a finite MDP, then extract the greedy optimal policy with stable tie handling.
