Basically the build was never running this target (unless maybe in some
edge cases as could be demonstrated by the Arch repository build, though
it doesn't look they did anything particular; this is how we discovered
the bug #6447 as this was not run on our own local or CI builds).
Reading the docs of run_target(), it may actually have been normal.
run_target() looks to be only about creating top-level targets (which
one could run with `ninja validate_menus` in our case for instance), not
actually running the command (i.e. badly named function).
For the command to be actually run on a normal build, let's use a
custom_target() as proposed by Paolo Bonzini. After testing, the xmllint
validation is properly run on a normal build and dependency works fine
with both the source XML and generated XML files (touching these files
trigger a rebuild).
The output of xmllint is stored in some dummy file, which is only useful
to prevent re-running the command at each build even though source XML
were unchanged (so it's more of a flag file).
See discussion in #6447.
19e545bc1f