He's not saying that. What he is saying is that are simple, non-production quality implementation in Scala is much more amenable to experimentation than a sophisticated, production-quality implementation in C++ that weighs in at 300x the LOC.
But a simple non-production quality implementation in C++ would also be amenable to experimentation and not have the bootstrapping issues as well as provide an easier starting point to incorporate more of the existing optimizations as desired.
Probably not, because JVM users are much more likely to be more proficient in Scala/Java than in C++.
That is true, but the author of Metascala wanted to write it in Scala. Other people are free to write a simple C++ implementation of the JVM themselves.
I'm sure it's not complete, but I also wouldn't be surprised if 99% of what's in HotSpot is optimization tweaks and performance boosts which aren't essential to the JLS.
It's also unclear to me if the architectural assumptions made in GC & VM engineering 20 years ago still apply today. I'm not an expert in this field, so take with a grain of salt... but I do wonder if a VM engineered from the ground-up today targeting recent ISAs would look substantially different.
It's mostly GC and JIT and a huge amount of corner case interactions between them. And also the many supported instruction sets.