The Geometry of Self-Verification in a Task-Specific Reasoning Model