Did not read the code yet, so maybe there is something to simulate congestion packet loss.
TCP-over-TCP is not as bad as some documents describe. It works surprisingly well in practice, especially with modern congestion control algorithms (BBR). For traditional algorithms that rely on packet loss, DSVPN has the ability to couple the inner and outer congestion controllers, by setting the BUFFERBLOAT_CONTROL macro to 1 (this requires more CPU cycles, though).
I had similar experience lately whenever using the mobile internet in poorer coverage cases (eg. a high speed train Brussels-Paris or Brussels-London) which result in frequent and big variations of latency. Of course that all is decoupled congestion control.
If anyone compared the performance of this solution on poor connections with and without BUFFERBLOAT_CONTROL, would be very very interesting to know the results!