Design adaptive n23 algorithm that works
In #4488 (moved) we put together a patch based on Mashael and Kevin's work on N23. It includes the static N23 algorithm and also an adaptive algorithm that lowers N3 when cells spend too long in connection outbufs.
Except Kevin's original adaptive algorithm a) tracks cell time in a really sketchy way, and b) was shown in the defenestrator paper to be worse than the static one.
We can solve "a" at least by using our cell statistics: see cell_waiting_time in connection_or_flush_from_first_active_circuit().
Once somebody writes a patch to do what the adaptive n23 algorithm aimed to do, but with our cell statistics, this will turn into a research problem to see if there are particular tuning values that work well.