You added the assumption that data.size() fits into an int. Which it does, but regardless, baking more assumptions into the code is a very clear regression.

Why don't you get the size as an int and then test that ? You can still have both checks and a nicer error code this way.

Even though it is unlikely to matter in practice, it is a good idea to have one check for error condition then detail in there.

if (uint64_t(size) > MAX_SCRIPT_ELEMENT_SIZE) {
  return set_error(serror, size < 0 ? SCRIPT_ERR_NEGATIVE_NUMBER : SCRIPT_ERR_PUSH_SIZE);
}

This is not an error in itself. This needs to be reformulated.

Maybe unexpected negative number ? Because negative numbers are not errors per se.