Would it make sense to combine this into one hg invocation? I think it wouldn't be that hard (hmm, not as trivial as I thought, but.. not the worst thing ever? :))

allheads=(${(f)"$(_hg_cmd log -r 'sort(head(), -rev)' --template '{ifcontains(rev, revset(\".\"), \"T\", \"F\")}:{rev}:{branch}\n')"})

# Extract current revision from allheads, remove "[TF]:" prefix from each item
selfentry=(${${allheads:#F*}/#T:/})
heads=(${${allheads:#T*}/#F:/})

# Remove "{rev}:" prefix from each item in heads to get just branch names
branches=(${(S)heads/#*:/})

# Remove branches that match the one we're currently on
selfbranch=(${(S)selfentry/#*:/})
branches=(${branches:#$selfbranch})

It probably would make sense, but you don't need to exclude current branch, because hg merge my-current-branch works when you have more than one branch head (except when you stand on the newest head, unfortunately: "abort: merging with a working directory ancestor has no effect").

Now it's only one call to hg. All possible branches included, even if hg merge that-branch would complain (see my comment above), because in completions it's better to give extra suggestions than to miss some.

But this "revision numbers + branch names" got me thinking, maybe it should also suggest hashes, since some people prefer them? Specifically, {node|shortest}, so that they don't take too much space.