Patrick Steinhardt [off-list ref] writes:

quoted

I do not think "prefer host Git" is necessarily a good idea; falling
back to use host Git is perfectly fine, of course.

Why is that, though?

Mostly because your "differences in features supported by just-built
one and what happens to be on $PATH can cause problems" cuts both
ways, and as a general principle to work around such issues, using
just-built one is a better discipline.  The features the build
infrastructure ("self-checks" being discussed is a part of it) of a
particular version of Git source depends on are more likely to be
found in the binary that will be built from the matching source,
than what happens to be on $PATH that may be from a year-old version.

As you said, you'd need to accomodate need for those who are
initially porting or building Git on a host without an installed
one.  If we were doing a cross build where just-built on would not
be executable on the host, whatever version on $PATH (or in
/usr/bin) may have to be used, but then in such a case you would not
be testing on host anyway.  For these two reasons, it is a given
that one of the choices has to be to use just-built one.  We can
safely give lower precedence to the host tool.

The only one-and-half practical reasons we may want to fall back to
whatever happens to be on $PATH are:

 - just-built one is so broken that even the simple use by the build
   infrastructure (like "self-checks") does not work (but then it
   becomes "if it is so broken, fix it before even thinking about
   running tests", and that is why I count it as half a reason), or

 - we are bisecting down to an ancient version, and just-built one
   from such a version may not understand the current repository.

so I do think it is an excellent workaround to fall back to a
version of Git with unknown vintage that happens to be on $PATH,
than failing and stopping by relying only on just-built one.

We already use host Git in other parts of our build
infra, and the options we pass to git-diff(1) have been around for ages:

It only argues for "trying host one, if available, before just-built
one does not hurt for this particular case".  But I was interested
in laying out a more general principle we can follow in similar
situations in the future.