But in any case, none of that is an argument for building a protocol that forces the problems that result from that on all platforms that don't have such restrictions, at best that is an argument in favor of workarounds to make things work as well as possible on such platforms.
Also, I would think that the long-term solution to this problem should be mobile link layer protocols that allow the mobile station to receive incoming packets with little delay while still conserving power. Stuff like high-powered, high SNR alert channels that allow the mobile station to shut down everything apart from a simple, low power, receiver, that only powers on for a few microseconds every few hundred milliseconds to listen for a wakeup signal from the network, so that incoming packets can be delivered with a latency of a few hundred milliseconds at any time. And then you are stuck with a stupid poll-based protocol that's adapted to the limitations of some ancient technology as a replacement for an even more ancient protocol that didn't have that limitation.
The platform argument is roughly that if they hold open one connection, it's cheaper than each app holding open their own connection. And maybe, if you're optimistic, the platform will be better at figuring out ping intervals for that connection that adapt to broken Nat devices per network (ie: on some networks, wake up once a minute to keep the connection active, and on reasonable networks wake up once every 30 minutes)