Rebuild canonical learning-rate balancing evidence

Update logbook: Repro - Balancing Learning Rates Across Layers: Exact Two-Step Dynamics and Optimal Scaling in Linear Neural Networks

initial commit