Scaling Recurrence-aware Foundation Models for Clinical Records via Next-Visit Prediction — ThinkLLM