I noticed the problem when the deadlock threshold that was added recently started getting triggered at very high injection rates (post-saturation) with the synthetic network traffic.

It seemed odd to me that a NI could not inject for 10000 cycles, however congested the network may be, as round-robin should ensure that eventually every VC at every input port should get a chance to leave the router (in the absence of a true deadlock).
I didn't measure how the overall stats change now, but the deadlock counter does not trigger now even if I (try to) inject 1-flit every cycle showing that this fix works.