We treat the scheduling of a single server in a finite-buffer capacity, multi-class, make-to-order production system subject to inventory holding costs, set-up times, and customer rejection costs. We employ theoretical and numerical analysis of a Markov decision process model to investigate the structure of optimal policies and the performance of heuristic policies. We establish the monotonicity of optimal performance with respect to the system parameters. Based on our insights, we provide a heuristic policy called the Capacitated Modified Index Rule (CMIR) for capacitated scheduling with customer loss penalties. The CMIR heuristic can easily be precomputed and stored for real-time control. Numerical benchmarking with respect to the optimal performance as well as an existing heuristic suggests that CMIR is very effective.