If you want to know about how well we can implement the semantics of an existing language then look at JRuby+Truffle - it passes more Ruby language tests than any other alternative Ruby implementation. The only stumbling block is co-routines as the JVM doesn't have these and they're hard to implement efficiently with other constructs.
That's a pretty critical feature of Lua that somewhat defines it(along with it's low overhead + embeddability, both of with you won't get with the JVM).
It seems like engineers always want to build One Tool to Solve Them All(tm) and yet there are always tradeoffs to be made. The reason 90% of this industry is still employed is because we don't have these one size fits all problems that we'd so love to solve.
SubstrateVM solves the embeddability part.
For context we use to run VM + tuneable data + game logic code in a 400kb block allocated to Lua on the PSP. I'd love to hear how SubstrateVM compares. Lua has a really rich history of being embedding in some pretty small targets.
Some of this is just simply a design trade-off. E.g., in JRuby+Truffle, we've implemented much of the core library in Ruby. This has allowed us to achieve a high level of language compatibility in a fairly short period of time -- implementing 3,000 core library methods in Java would be quite the undertaking. As a consequence, the static binary must include those full Ruby sources (compressed, but they're not in an optimized bytecode format).
Having said that, if we wanted to optimize for size over a restricted subset of the language (think mruby), that would be straightforward to do.