We present a retrieval-augmented few-shot system for the MedSynth Dial2Note shared task at SMM4H-HEARD 2026, placing 3rd on the official leaderboard (0.51 avg). Across 28 configurations, we find that retrieval design (hybrid BM25 + medical-domain dense fused via RRF) and prompt presentation format (few-shot examples as conversation turns) are the primary quality drivers, while model scale has surprisingly limited impact: Llama 3.2:3B, Llama 3.1:8B and GPT-4o mini remain within a narrow band on our locally computed scores. Our final submission used GPT-4o mini with k=3 few-shot examples retrieved by RRF over BioLORD-2023 embeddings. We report a full ablation, including negative results, to show where the gains come from and where further engineering stops paying off.