I'm not sure we can agree that the dogfooding is preferable. The penalty paid in the stdlib would have to be paid time and time again by users -- as it stands, Jigsaw is not something I can conveniently opt-in (or opt-out) -- it's very clear that the opt-out switches are temporary hacks, and even those don't make porting trivial. As someone who used to be very active in the Python community, I think the 2/3 comparison is very apt.

The problem with that is that nobody is backporting fancy compiler improvements to JDK8; I can't get my faster JDK without also buying into a module system I didn't want and also broke my code.

It is certainly not clear to me that stronger visibility restrictions are a priori desirable when they break dynamic hosted languages. Reflection as a side-stepping mechanism is already easy to detect and clearly opt-in by the reflector, and you've had the option of opting out for ages with Security Managers. All Jigsaw buys you is another level of granularity in that visibility (not just class, subclass, package) and breaking the workarounds for when you actually really mean it.