[CVPR2026] V-Attack: Targeting Disentangled Value Features for Controllable Adversarial Attacks on LVLMs