Phoenix-MWA: Learning Generalizable World–Action Models with Structural and Semantic Representations