Now you have my attention - "Or Node embedded in a phone or network switch where it is performing routing logic for multiple connections, but in a single process and you’re not far off". I would be interested in seeing what the performance impact is on ARM architectures since the above are strong use cases.
Does an uncaught exception still kill the process? Because I'm really growing to like node, but this is a very strange design decision. Yes, you can mitigate it with forever or pm2, but you really shouldn't have to.
Is this the C++ accessible side of vm.createContext() and vm.runInContext()? Seems like it would help with sandboxing.
Is the long-term goal thread safety? Having a native threading mechanism in Node would definitely be helpful.
Thread safety in this case I think would mean running multiple instances of node embedded in your app, each in their own thread and with their own event loops.