Enhancing Robustness of LLM-Synthetic Text Detectors for Academic Writing: A Comprehensive Analysis