From Foresight to Forethought: VLM-In-the-Loop Policy Steering via Latent Alignment