Anthropic’s Claude Fable guardrails apology shows why AI companies need clearer safety rules, better user signals, and less mystery around model limits.