We study limited-coordination scheduling in a wireless downlink network with multiple base stations, each serving a collection of users over shared channel resources. When neighboring base stations simultaneously schedule users on the same channel resource, collisions occur due to interference, leading to loss of throughput. Full coordination to avoid this problem requires each base station to have complete "instantaneous" channel-state information for all its own users, as well as the ability to communicate on the same timescale as channel fluctuations with neighboring base stations. As such a scheme is impractical, if not impossible, to implement, we consider the setting where each base station has only limited instantaneous channel-state information for its own users, and can communicate with other base stations with a significant lag from the channel state variations to coordinate scheduling decisions. In this setting, we first characterize the throughput capacity of the system. A key insight is that sharing delayed queue-length information enables coordination on a slow timescale among the base stations, and this permits each base station to use limited and local channel-state along with global delayed queue-state to stabilize its users' packet queues. Based on this, we develop a distributed, queue-aware scheduling (and information exchange) algorithm that is provably throughput-optimal. Finally, we study the effect of inter-base-station coordination delay on the system packet delay performance under the throughput-optimal algorithm.