Variance Is Not Importance: Structural Analysis of Transformer Compressibility Across Model Scales