there's no need to actually have the numbers all at once, i.e. no need for the vector at all.

even with SSO eliminating heap usage, using to_string is much more work than necessary, when a simple loop and modulus operation works fine.

small string optimisation means that this should (probably?) never allocate on the heap.
I dont think this answer is best practice, but this comment thread starter is about as mislead as they thought the answer was.