Test passes on 5.3.0 but fails on 5.4.0~beta2

[edwintorok]

It reliably fails like this on 5.4.0~beta2

 TEST  test_man
       Invoke with --test-help[=FMT] to test --help[=FMT] interactively
       Test plain text manpage
       File "/var/home/edwin/git/cmdliner/test/test_man.ml", line 112, characters 35-3477:
--- expected
+++ found
@@ -1,6 +1,5 @@
 NAME
-       man_test - UTF-8 test: 🐫 íöüóőúűéáăîâșț
-       ÍÜÓŐÚŰÉÁĂÎÂȘȚ 雙峰駱駝
+       man_test - UTF-8 test: 🐫 íöüóőúűéáăîâșț ÍÜÓŐÚŰÉÁĂÎÂȘȚ 雙峰駱駝
 
 SYNOPSIS
        man_test [OPTION]…

       Test groff manpage
       1 snapshot is incorrect. Run with --correct to correct it.
 FAIL  1/2 test failed in 1.9ms

And it reliably passes on 5.3.0.

Tried with 2 different terminal emulators (ghostty and konsole).
Tried with 2 different shells (bash and zsh).
The behaviour is the same in all 4: works with 5.3.0, fails with 5.4.0~beta2.

A quick peek at the changes in 5.4.0 shows that this might be related, especially given the Unicode char there (do they compute different widths?):

+- #13570, #13794: Format, add an out_width function to Format device for
+   approximating unicode width.
+  (Florian Angeletti, review by Nicolás Ojeda Bär, Daniel Bünzli,
+   and Gabriel Scherer)

This is going to be a bit annoying to fix, perhaps based on the compiler version the test should expect 2 different outcomes?
A better fix might be to override out_width on OCaml 5.4 and set it to be compatible with OCaml <5.4 behaviour, although that'd need conditional compilation based on the OCaml version then.

[edwintorok]

At least I think this should be marked as a breaking change in the OCaml 5.4 release notes, since it affects testcase outputs (even if the new behaviour is more correct), I'll raise that on the OCaml bugtracker.

[dbuenzli]

A quick peek at the changes in 5.4.0 shows that this might be related, especially given the Unicode char there (do they compute different widths?):

Yes starting from 5.4 computed width are 1 for scalar values so it becomes more correct in a lot of cases, especially if your text is in NFC.

This is going to be a bit annoying to fix, perhaps based on the compiler version the test should expect 2 different outcomes?

That, or devise a tester that ignores newlines (though that would a bit defeat the purpose).

Perhaps simply do nothing about and develop with the latest OCaml version.