which is both unexpected and different from the Vim behaviour.

Indent was triggered once by the '!' check in insert_execute(), and
inserting the char was correctly skipped, but then triggered again in
insert_check() (provided that cindent was not being ignored after manual
indentation, i.e. `can_cindent == true`).

While this is the smallest fix, another solution would be to remove
VimState#check and instead move that to *_enter()/-_execute(), since the
control flow is pretty unnecessarily convoluted as is. That would also
have the benefit of differing less from the Vim source code.

Using sleep(10) to wait for typeahead to finish is flaky, especially on
macOS, where legacy/global_spec.lua has failed several times.

N/A patches for version.c:

vim-patch:9.0.0135: comment about tabpage line above the wrong code

Problem:    Comment about tabpage line above the wrong code.
Solution:   Move the comment. (closes vim/vim#10836)
https://github.com/vim/vim/commit/0b0ccbbfb014e1c9682b86a7a41ff2837b0b8047

Remove check for MOD_MASK_META as it is for <T- which never appears in TUI.
Make small changes to docs.