This paper describes our system for Subtask 1 of the SMM4H HeaRD 2026 Task 2, which is an LLM-based system for binary insomnia classification from MIMIC-III clinical notes using OpenAI GPT-5.2 with chain-of-thought (CoT) prompting. Our approach implements three strategies: baseline fixed 8-shot prompting, dynamic retrieval using semantic embeddings, and self-consistency voting. The system applies rule-based criteria combining symptom patterns (difficulty sleeping and daytime impairment) with medication indicators (primary and secondary insomnia medications).Our best configuration (Self-Consistency Voting) achieved 95.67% weighted F1 on validation and 82.35% F1 on the official test set , outperforming the Baseline (81.25% F1). Notably, our test F1-score of 82.35% substantially exceeded the task mean (68.05%) and median (70.37%) across all participating teams. Key contributions include explicit comorbidity exclusion prompting, context-aware nursing note handling, logical constraint enforcement for prediction consistency, and a comparative analysis demonstrating that self-consistency improves recall at moderate computational cost.