The tests assume a mesh network with known connections, the node doesn't,
so unneeded packet leakage isn't really as controllable as I'd hoped.
What's definitely important is that the system doesn't cache entries for
long enough that it starts continuously sending packets the wrong way
no matter what.
In practice it is unlikely that data that's too big for the reliability layer won't need splitting
anyway by the application at some point, and the libraries are already too open to OOM DoS.
