I don't think there's coverage for many of the subtle cases, and I
found it hard to understand what the code is doing by reading it. The
test takes 40s to run on a laptop, or 9s with --chg.
I have yet to find a description of what the files field is supposed
to be for merges. I thought it could be one of:
- the files added/modified/removed relative to p1 (wouldn't seem useful, but hg diff -c -r mergerev has this behavior)
- the files with filelog nodes not in either parent (i.e., what is
needed to create a bundle out of a commit)
- the files added/removed/modified files by merge itself 
It's clearly not 1, because file contents merges are symmetric. It's
clearly not 2 because removed files and exec bit changes are
listed. It's also not 3 but I think it's intended to be 3 and the
differences are bugs.
Assuming 3, the test shows that, for merges, the list of files both
overapproximates and underapproximates. All the cases involve file
changes not in the filelog but in the manifest (existence of file
at revision, exec bit and file vs symlink).
I didn't look at all underapproximations, but they looked minor. The
two overapproximations are problematic though because they both cause
potentially long lists of files when merging cleanly.
 even what it means for the merge commit itself to change a file is
not completely trivial. A file in the merge being the same as in one
of the parent is too lax as it would consider that merges change
nothing when they revert all the changes done on one side. The
criteria used in the test and in the next commit for "merge didn't
touch a file" is:
- the parents and the merge all have the same file
- or, one parent didn't touch the file and the other parent contains the same file as the merge