Shortcuts and Boundaries: Verifiable RL Misalignment and Contextual Safety Learning | ArxivCSExplorer