Adaptive Agentic Attacks on LLM Vulnerability Detectors via Adversarial Comments