VLK: Learning Humanoid Loco-Manipulation from Synthetic Interactions in Reconstructed Scenes — ThinkLLM