parser to utf8mb3 conversion hides errors

This shows in a test case for MDEV-22558

set names utf8mb3;

--error ER_INVALID_CHARACTER_STRING

create table t1 (x int comment 'a’');

set names utf8mb4

--echo # no error ???

create table t1 (x int comment 'a’');

In the second case the comment is valid 'a?' string.

This happens because all metadata are internally always utf8mb3, stored as utf8mb3, shown in INFORMATION_SCHEMA as utf8mb3.

So if the connection charset is utf8mb3, the parser gives it to the server as is. Later validate_comment_length() validates that the comment is well formed and finds that it is not.

If the connection charset is not utf8mb3, the parser uses TEXT_STRING_sys rule that converts the string to utf8mb3 with THD::convert_string() and the latter ignores charset conversion errors. The server and validate_comment_length() gets a perfectly well-formed string "a?".