Wouldn't it be simpler to use the re {m,n} notation to grab the appropriate number of arguments?  

In ArgumentParser _get_nargs_pattern we could add:

+        # n to m arguments, nargs is re like {n,m}
+        elif is_mnrep(nargs):
+            nargs_pattern = '([-A]%s)'%nargs

        # all others should be integers

where is_mnrep() tests that the nargs string looks like '{m,n}' (or '{,n}' or '{m,}').  The resulting nargs_pattern will look like

    '([-A]{m,n})'

In _format_args() a similar addition would be:

+        elif is_mnrep(action.nargs):
+            result = '%s%s' % (get_metavar(1)[0], action.nargs)

   'FOO{2,5}'

It would take more code to generate

   FOO FOO [FOO [FOO [FOO]]]

A matching programmer input would be:

   parser.add_argument('Foo', nargs='{2,5}')

though it wouldn't be much work to also translate a tuple.

   parser.add_argument('Foo', nargs=(2,5))

I'll try to test this implementation against wm's tests.