Multi-Agent Path Finding (MAPF) is a fundamental problem in multi-agent systems, with broad applications in warehouse logistics, robotics, and autonomous driving. Ensuring that MAPF solvers consistently return a solution whenever one exists and deliver high-quality plans is critical for real-world deployment. However, MAPF algorithms remain insufficiently tested. The high-dimensional, tightly coupled input space arising from map topology and agent configurations makes systematic testing extremely challenging. This paper introduces MAPFFuzz, a metamorphic fuzzing framework for systematically testing MAPF solvers to detect both completeness failures and optimality failures. MAPFFuzz uses semantics-aware mutation strategies that jointly perturb both maps and agent configurations, coupled with a lightweight feasibility verification mechanism to ensure valid scenarios. Additionally, it incorporates diversity-guided seed selection and leverages metamorphic relations as oracles to assess solution quality. We evaluate MAPFFuzz on four recent MAPF solvers: PIBT, HCA, PIBT+, and LaCAM*. Results show that MAPFFuzz can detect completeness failures and more than 400 cases with degree of suboptimality exceeding 25% with 100 agents on 128x128 map cases per solver. MAPFFuzz also identifies 37 scenarios unsolvable by LaCAM* even under a 30-second parallel optimization setting. These findings demonstrate that MAPFFuzz can reveal both completeness failures and significant solution degradations, even in state-of-the-art algorithms under generous time budgets.