Here is what we know right now:

1. If you use a keyed checksum with RC4 keys and an authenticator subkey
in a TGS request, AD 2003 verifies the checksum using the subkey.  It
turns out that RFC 4120 doesn't specify what key to use for AP-REQ
checksums, but Heimdal and MIT use the TGS session key.  RFC 4757
(Microsoft's own informational RFC about RC4-HMAC) says to use the TGS
session key, so MS is in conflict with its own documentation if not with
the binding standards.

What we don't yet know for sure is whether this problem affects AES.  We
need to find that out to know the appropriate scope of the fix.  If the
problem affects only RC4, then the appropriate answer is probably "don't
use keyed checksums with RC4, it hurts."  If the problem affects AES as
well, then it gets more involved.

2. RFC 4757 erroneously documents a key usage of 8 for a TGS-REP
encrypted part authenticated with a subkey; the value used by MS is
actually 9.  Unfortunately, Heimdal and MIT both implement what is
documented.  This means you can't interoperate with both {Heimdal or MIT
1.6} and AD with RC4 TGS subkeys using a single key usage value.  It's
easy enough to try both when decrypting the response, however.

Sam has committed a change to switch from 8 to 9, fixing TGS RC4 subkey
interoperability with MS but breaking it with Heimdal and MIT 1.6.  We
will need to amend this to try both usage values.