hmm, i don't know where you get this idea from, but to me, the code of Pd's networking infrastructure looks, as if a) all incoming traffic was polled for in the main thread b) all output of this traffic (to the Pd-patch) was propery protected by sys_lock()
I am honestly not that familiar with this low-level aspect of pd. What I am however aware of is that the poll function responsible for receiving message (meaning when it arrives at the socket) inside netreceive immediately dispatches it to pd and its polling is linked not to pd's main loop but to the moment when the socket has received a packet which appears to be a part of any basic network socket example code available online and as far as I can see, I did not notice any relationship between its polling intervals and Pd's main loop.
FWIW, The only difference in my implementation is that the received messages are forwarded to pd via clock_delay(0) which has fixed all such occurrences inside L2Ork which is quite network traffic heavy, and apparently has resolved problems for others as well.
Best wishes,
Ico