Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents