Thanks for the feedback!
I didn't think about this tradeoff between simplicity and power for max recursion depth. So there's no reason to think that the proposal currently chooses the best compromise! And indeed, a completely hardcoded limit is probably bad because users who really need more depth will be stuck.
Not sure if you meant #2 to be a global limit, or per-function, but a configurable global limit sounds like a pretty good point in the tradeoff space to me. Looking to other languages for precedent, I find that a few C++ compilers (clang and gcc) have a default limit of 512 for constexpr max recusion depth, and a compiler flag that allows you to change the limit if you really need to.
Something dynamic like #3 sounds a bit too powerful to me. It would add a totally new concept to the language (polynomial bounds), and I expect that most practical compile-time evaluated functions will be small and simple enough to not need it. Users can easily deal with the few large and variable-depth functions by occasionally bumping up a global limit.
(And I strongly suspect that the proposed programming model is turing complete, so #1 is indeed undecidable.)