Provably Robust Adaptation for Language-Empowered Foundation Models