Switch to real putnam_small benchmark + statement-match guard

Leaderboard: score JSONL submissions by token-level proof reduction