The current implementation of virtual shards leaves a lot to be desired when used in big corporate applications.
A better way to model shards, is to identify a virtual shard by a pairing of "some name" + "an index" (example: Clients, 11). The name plays a role similar to namespaces. The index is always zero based and valid only within the namespace. There can be many shard namespaces with unique names. A configuration file (possibly XML) will map each virtual shard to a single physical shard. It's a simple mapping strategy. The mapping can allow overlaps of virtual shards into a single physical shard. This type of naming can make things more clear and easier to manage on the long-run.
The big advantage to virtual shards over physical shards is that virtual shards can be easily persisted to file or database and retrieved again and still remain valid. Any change to the shard structure is fully controlled by the shard configuration file, without modifications to the persisted representation of the shard. This way the code can evolve separately from the physical shard architecture.
With this kind of virtual shard, top-of-tree classes will map to a single shard namespace. The object will reside in any of the virtual shard indexes associated with the shard namespace. The resolution strategy becomes very simple because the set of shard indexes will be returned.