Coalition is an important mean of multi-robot systems to collaborate on common tasks. An effective and adaptive coalition strategy is essential for the online performance in dynamic and unknown environments. In this work, the problem of territory defense by large-scale heterogeneous robotic teams is considered. The tasks include surveillance, capture of dynamic targets, and perimeter defense over valuable resources. Since each robot can choose among many tasks, it remains a challenging problem to coordinate jointly these robots such that the overall utility is maximized. This work proposes a generic coalition strategy called K-serial stable coalition algorithm (KS-COAL). Different from centralized approaches, it is distributed and anytime, meaning that only local communication is required and a K-serial Nash-stable solution is ensured. Furthermore, to accelerate adaptation to dynamic targets and resource distribution that are only perceived online, a heterogeneous graph attention network (HGAN)-based heuristic is learned to select more appropriate parameters and promising initial solutions during local optimization. Compared with manual heuristics or end-to-end predictors, it is shown to both improve online adaptability and retain the quality guarantee. The proposed methods are validated rigorously via large-scale simulations with hundreds of robots, against several strong baselines including GreedyNE and FastMaxSum.