Characterizing Model Behavior Under Synthetic Data Training: An Empirical Study Across Scales and Mixing Ratios