Create a diagnostic tool that visualizes spectral edge dynamics during model training to detect grokking early. This will give researchers a new window into internal model learning states that standard probing misses.