It's a design problem, but I'd be interested in suggestions for how to rework the mod to avoid the issue.
You can try tuning the I/O points to reduce the problem. There is a small knob on each I/O point to adjust the threshold for sending items. You can reduce that from 10 to e.g. 1, but it is then more likely to oversend.
The whole mod uses circuits, and it uses Lua only to create the circuits and not per-tick, for performance. For each requested item type, the circuit simulates "heat" transfer. Each router has a "temperature" for the item, and at each tick the requests make the network "hotter" at that point, items traveling in the network make it "colder", some of the "heat" diffuses from each router to each of its neighbors, and some of it leaks away. This sets up a temperature gradient across the network. The I/O terminals send stuff when the item's "heat" is more than some threshold, and the network routes items along the temperature gradient. It overall mostly works, even at scale, but there are some undesirable behaviors, like oversending or items looping around when demand is weak.
All the math is fixed precision, so small amounts of heat get rounded to nothing, and if they got rounded down to e.g. 1 instead of 0, then the network still wouldn't know where to send stuff because there's no gradient. It's even worse in a grid than in a chain, because the heat spreads out slowly across the grid. The leakage term contributes to this, but it's required to prevent heat from building up forever when there are no items of that type available.
If you have a big network and need a high flow rate, you have to request a lot anyway, because if you need e.g. 1/s and it'll take 2 minutes to cross your base, then that's already minimum 120 items in motion. But with the falloff, you'd probably have to request ~500 items depending on how many routers are in between.
I've thought about changing the network to separately track which direction requests came from, in addition to the temperature gradient. This would let it work with requests too small to create a stable gradient, but the requests would still need to propagate well through the network.
