Emergent misalignment, LLM value systems, recursive critiques, no canonical SAE features, refusal cones, REINFORCE attacks, and the case for theoretical understanding