We should buffer all packets of a transaction and only when all have arrived, should we start processing them with the client thread blocked.
The way it's now, any network lag during transactions will directly translate to loss of frames. Fixing this might even make #66 imperceptible.