ICL Suite Shows LLMs Lack Monotonic Generalization Gains During Pretraining