[amitchone]

The acceptor thread used to write newly accepted sockets straight into the array the handler thread was polling, and because Zephyr's poll checks that array twice, when we refilled a slit mid-poll, the handler mistakes an idle client as readable and could block on it indefinitely...

Now the acceptor hands sockets over through a queue, only the handler touches these and client sockets are non-blocking.

Ultimately, after many hours of uptime we would occasionally see refused connections because there were no sockets left, even though, theoretically, "nothing" had used that many sockets

This also rounds out the keepalive fix which, after even more testing, led to this discovery and fix

[garethpotter]

I think this is OK. I suppose it might be worth having Claude try to beef up the client/server tests a bit.