A Policy Algebra for Trust-Preserving Agentic AI Execution