Exposing the LLM Code Trust Gap in AI IDEs