The Code That Never Ran Modeling Attacks on Speculative Evaluation