The previous value caused statistics to occasionally timeout spuriously.

We're closing and reopening the connection each time, we should be
renegotiating instead.

That would be racy.

Don't take packet loss into account, just clamp at 384kbps.

Otherwise we never recover from low throughput.