Well at least today (F39) /usr/libexec/git-core/ seems full of symlinks

The way that the Docker plugin works is that it mounts the image with libguestfs (using FUSE), and then runs tar on the result. However, FUSE, by default, makes up the inodes without regard to the underlying inodes, so tars hardlink detection in tar can't work.

There would be two fixes for this:

A) pass the use_ino option to the mount call, Something like:

diff --git a/imagefactory_plugins/Docker/Docker.py b/imagefactory_plugins/Docker/Docker.py
index 68e12c3..3b44603 100644
--- a/imagefactory_plugins/Docker/Docker.py
+++ b/imagefactory_plugins/Docker/Docker.py
@@ -317,7 +317,7 @@ class Docker(object):
         try:
             tempdir = tempfile.mkdtemp(dir=storagedir)
             self.log.debug("Mounting input image locally at (%s)" % (tempdir))
-            guestfs_handle.mount_local(tempdir)
+            guestfs_handle.mount_local(tempdir, options="use_ino")
             def _run_guestmount(g):
                 g.mount_local_run()
             self.log.debug("Launching mount_local_run thread")

(not tested since I'm having trouble getting imagefactory to work...)

B) Skip the guestmount and use guestfs_tar_out_opts() which would simplify the code quite a bit (no need to start a thread) and likely perform considerably better. The computation of layer size would take a bit of adoption, but I think this is some old obsolete thing - if you use skopeo to copy to 'docker-archive:' you don't get that field.

Perhaps your one-line patch/workaround could be tested for Fedora Rawhide?
I opened https://pagure.io/releng/issue/11735 to see.

Turned my pseudo-patch into a a pull request. It's still not actually tested for an image build, but I created a test program to verify the syntax and make sure that it has the intended result.