I believe Tor and I2P should not be conflated. Timing attacks are more difficult against I2P. With Tor, you have exit nodes which make it easy to see one of the endpoints for the purpose of correlation. It can still be done without visible exits, but this is a bit harder.

There has been some talk about introducing random delays to harden the mix network layer. Since that has various drawbacks, how about adding cover traffic instead? It is wasteful but should still allow low enough latencies while mitigating timing analysis. This is another thing that would not work well with an exit node based system.