William Mead has been working on LMTP code for me. He's produced implementations for version 3.0 and for version 2.2, with these tests applied after RCPT TO in the LMTP conversation:
1. message will be rejected if the list name is not known.
2. message will be accepted if the sender matches "accept_these_nonmembers".
3. message will be accepted if "generic_nonmember_action" is not reject.
4. message will be accepted if the sender is a list member.
5. if we get this far, the message will be rejected - the sender is a non-member of a closed list.
We could also reject other members if they're moderated, for example. However, we've adopted the view that it is relatively safe to generate a bounce message for someone who is a member of the list.
William has completely reimplemented the SMTPD code in Python, to support ENHANCEDSTATUSCODES, because the LMTP RFC requires that - even though the examples in the RFC don't show them being used! However, the code doesn't implement PIPELINING - also required by LMTP - because the underlying ASYNCHAT/ASYNCORE architecture doesn't seem to support it. We discovered that advertising PIPELINING causes the test smtp client to fail, but we've not even thought about how to fix that - LMTP clients which are re-implementations of SMTP clients might just live with the fact that PIPELINING isn't advertised.
William's code is at https://code.launchpad.net/~wilunix
The LMTP queue runner allows us to run Mailman on a server that's unrelated to the main MTA. With an Exim MTA, you could use a recipient callout to verify that the sender is permitted to post to the list, before accepting the message for deliver. This means that rejecting an unwanted message should not create collateral spam.