When Thoughts Meet Facts: Reusable Reasoning for Long-Context LMs