Toward Supporting Perceptual Complementarity in Human-AI Collaboration via Reflection on Unobservables