ORIC: Benchmarking Object Recognition under Contextual Incongruity in Large Vision-Language Models