Reflections on Trusting Trust, Revisited: Poisoning Self-Modifying AI Coding