When a 0.4-Point Lead Means Nothing: Reading Open-Model Leaderboard Margins Like a Statistician