Due to repeated use of StringIO as a way to "look ahead" into subparts 
while checking that multipart boundaries are unique, memory consumption 
during email.generator.Generator.flatten() can be up to 3 times the 
original message size.

I implemented a subclass of email.generator.Generator that works around 
this using email.message.Message.walk() to check message headers and 
string (final) payloads for the boundary without duplicating their 
contents into a StringIO.

It assumes that the boundary only ever might be duplicated in a single 
part's headers or in a single part's payload when that part's payload is 
a string.  IOW, it assumes that the boundary will not be duplicated by 
some combination of all the parts' and recursive subparts' headers and 
string payloads.

If this assumption is safe, then this implementation should work.  If 
this assumption is not safe, then perhaps a different boundary format 
can be used which will make this assumption safe?

You can find my implementation at http://gitorious.org/rpatterson-
imappipe/rpatterson-
imappipe/blobs/master/rpatterson/imappipe/generator.py

The email module has several problems. RDM is working on overhauling the email module for 3.2. Existing issues may not get individual attention.

When I first looked at this issue, it appeared to me reading the code that the assumption was not considered safe.  I'm hoping I can implement something akin to your algorithm in email6 (which, unfortunately, won't make 3.2), but I'm going to have to give it a lot of thought to make sure I'm not making any unsafe assumptions.  So while I have this issue on my list, it won't be addressed in 3.2.