I want a language with an internal/core calculus of GRTT with a specific grading that I believe will provide both optimal evaluation and prompt resource release. You aren't going to get that by building on top of another language.

Of course you could do it as a library for another language, but the few values you could lift/lower and the necessary analysis do to bidirectional type inference/checking mean that you are writing a compile phase whether you call it that or not. In theory, you might save on the tokenizer/parser, but those are trivial to write. You might be able to reuse parts of the API/ABI, which could be an advantage if the underlying language has a good, stable one that can reflect your linking/calling/passing restrictions; but that's not likely.