SWE-MISBEHAVIOR: Benchmarking and Mitigating Boundary Violations in Software Coding Agents