implementation would simply regenerate all resources, of all subscribed types, for each client, on any configuration change. However, this scales poorly. As a result, we have many levels of optimizations to avoid doing this work.

First, we have a concept of a `Full` push. Only `Full` pushes will recompute `PushContext` on change; otherwise this is skipped and the last `PushContext` is re-used. Note: even when `Full`, we try to copy as much from the previous `PushContext` as possible. For example,...