And sure as hell it is impossible to develop a mixnet that will generate Camouflage traffic
It would have to generate traffic in equal amounts for every flow, which would halve network speed to give an attacker a 50/50 chance of guessing the correct flow. Those fake flows would also have to be carried to something that looks like a reasonable endpoint as well.
PRISM-level metadata collection makes it trivial to see which computer sent the original 682-byte request (recurse as necessary until the 800 byte request starts at the "sender") as well as which computer the multi-megabyte response was sent to (recurse as necessary until the multi megabyte response returns to the requesting computer). Camouflage traffic can't fix this on its own, it's easy to exclude the data that wasn't requested from the analysis.
I think that Tor's best bet while maintaining performance at this point would be to round all packets up to the nearest MTU (lets say 1400 to account for PPPoE, VPNs, and other layers on ethernet), so every request and response becomes a multiple of 1400 bytes, would make most tracking rely on packet timing. The next step would be to introduce packet delays at each hop, but that will slow the already slow network down.