I've been playing around with this. My implementation is basically the naive:

def __getitem__(self, value):
    return self.group(value)

I have the following tests passing:

    def test_match_getitem(self):
        pat = re.compile('(?:(?P<a1>a)|(?P<b2>b))(?P<c3>c)?')

        m = pat.match('a')
        self.assertEqual(m['a1'], 'a')
        self.assertEqual(m['b2'], None)
        self.assertEqual(m['c3'], None)
        self.assertEqual(m[0], 'a')
        self.assertEqual(m[1], 'a')
        self.assertEqual(m[2], None)
        self.assertEqual(m[3], None)
        with self.assertRaises(IndexError):
            m['X']

        m = pat.match('ac')
        self.assertEqual(m['a1'], 'a')
        self.assertEqual(m['b2'], None)
        self.assertEqual(m['c3'], 'c')
        self.assertEqual(m[0], 'ac')
        self.assertEqual(m[1], 'a')
        self.assertEqual(m[2], None)
        self.assertEqual(m[3], 'c')
        with self.assertRaises(IndexError):
            m['X']

        # Cannot assign.
        with self.assertRaises(TypeError):
             m[0] = 1

        # No len().
        self.assertRaises(TypeError, len, m)

But because I'm just calling group(), you'll notice a few oddities. Namely:

- indexing with integers works, as would group(i). See the discussion of omitting len().
- for defined but missing group names, you get None instead of KeyError
- for invalid group names, you get IndexError instead of KeyError

I can't decide if these are good (because they're consistent with group()), or bad (because they're surprising. I'm interested in your opinions.

I'll attach the patch when I'm at a better computer.