When “Reasoning Mode” Backfires: Why More Thinking Can Make AI Less Reliable