Understanding crowd mobility is critical for many applications. In this paper, we propose CrowdMirage, a WiFi positioning-based crowd mobility digital twin for smart campuses. Specifically, we first design an end-to-end human mobility trace extraction pipeline from the comprehensive but noisy WiFi connection logs on a university campus. We then design two predictive and simulative models for the crowd flow prediction and simulation tasks, respectively. Considering the particularity of on-campus mobility, we propose a cross-grained crowd flow prediction model to forecast crowd flow at both building and floor levels. For crowd flow simulation, we design a conditional generative model based on conditional diffusion to simulate the crowd flow under given mobility-related contexts that are systematically identified. We evaluate CrowdMirage on two-year WiFi connection logs collected at our university. The results show that CrowdMirage achieves superior performance in both crowd flow prediction and simulation tasks. Our case studies show that CrowdMirage cannot only accurately forecast cross-grained crowd flow across different cases, but also simulate interpretable crowd flow under previously unseen conditions.