Cross-model consensus in AI code review: why different LLMs catch different bugs (technical writeup)