Standard vs. Modular Sampling: Best Practices for Reliable LLM Unlearning