I agree this is dirty and I will be very happy to improve this. I can think of folllowing improvements

First is this:

@@ -837,10 +838,14 @@ class cgpacker(object):
         size = 0
 
         clstate, deltas = self._generatechangelog(cl, clnodes)
-        for delta in deltas:
-            for chunk in _revisiondeltatochunks(delta, self._builddeltaheader):
-                size += len(chunk)
-                yield chunk
+        if changelog:
+            for delta in deltas:
+                for chunk in _revisiondeltatochunks(delta, self._builddeltaheader):
+                    size += len(chunk)
+                    yield chunk
+        else:
+            for delta in deltas:
+                pass
 
         close = closechunk()
         size += len(close)

The second improvement is to add a part argument instead of changelog where by default, part = ('changelog', 'manifests', 'treemanifests', 'filelog') and caller can pass a value of parts. The call in narrow will pass ('manifests', 'treemanifests', 'filelog') and in future we can reduce that to ('treemanifests', 'filelogs').

In D4638#70613, @indygreg wrote:

I think the code as written makes sense: I'd rather have 1 loop over deltas than 2.
Regarding the other potential improvement, the reality of changegroups is that both they and bundle2 are container formats. In the context of bundle2, I think it makes more sense to send "delta groups" as individual bundle2 parts and to do away with changegroups completely. My guess is we haven't done this because the changegroup code has always existed, is non-trivial, and "just works." In other words, there hasn't been a pressing need to split up the changegroup code, so nobody has done it.

Yep, just because changegroup exists, and it was not very much hard to prevent sending the changegroup data, I prevented myself working on the "delta groups" idea. However, in future, if we want to come up with a new format due to some reasons, individual "delta groups" sounds a promising idea.