Yeah, examples of 'small' interpreters (including 'small' self-interpreters) have always made me uneasy when they depend on a powerful runtime environment. E.g., in the limit, you can create a language where the 0-byte program is a self-interpreter, and all other programs are interpreted as Python or whatever. You can then point to "look how small the description is!" when it's really just sleight of hand. To put it as a hot take, bootstrapping is very overrated.

Of course, every human definition will similarly hide complexity to some extent, so there's no real way to win.