Consider the case of an 'interpreter per thread' in an application server. In lua this is implemented by creating a lua context per thread, which is your memory foot print. The size here limits the number of concurrent requests the server can handle.

So, in addition to my other comment, I think there's a major barrier to wide adoption to any of the 'mini' scripting language implementations, and that is a lot of people who use python would expect your implementation to act like the mainline. Here I'm alluding to the trouble pypy has gaining acceptance, because so many users expect all the c libraries to just work.

I also contemplated going this route, not to get the language I want, but to get the runtime instrumentation I want. The conclusion I came to is that it would almost be better to create a new language than to always be explaining "no, that functionality doesn't work because of these design decisions". The language may in fact be a lot like python, or a subset python as you say, but just don't call it that!

Edit; created a dupe somehow...