Many Ways to Be Fake: Benchmarking Fake News Detection Under Strategy-Driven AI Generation — ThinkLLM