maybe it would be a good start to include the examples at the end of RFC2047 into the regression tests? These examples at least support the case that a '?' may immediately follow an encoded string:

encoded form                                displayed as
(=?ISO-8859-1?Q?a?= =?ISO-8859-1?Q?b?=)     (ab)

when trying this in python 2.7:

>>> decode_header ('(=?ISO-8859-1?Q?a?= =?ISO-8859-1?Q?b?=)')
[('(', None), ('a', 'iso-8859-1'), ('=?ISO-8859-1?Q?b?=)', None)]

this fails. So I consider this a bug.

Note that although RFC2047 is vague concerning the interpretation if two encoded strings could follow each other without a whitespace, these *are* seen in the wild and *are* interpreted correctly by the mailers I've tested: mutt, thunderbird, exchange in various versions, even lotus notes seems to get this right. So I guess python should be "liberal in what you accept" and parse something like 
'(=?ISO-8859-1?Q?a?==?ISO-8859-1?Q?b?=)'
into
[ ('(', None)
, ('a', 'iso-8859-1')
, ('b', 'iso-8859-1')
, (')', None)
]