AFAIK the only reason upconversion would ever break content-md5
my guess is that it breaks content-md5 for the case where
bare CR or bare LF appears in the canonical form, and this
same character is coincidentally the line ending for the stored
form of the message at some point following upconversion. the MD5
computation at the recipient would see the line ending as CR LF
rather than just bare CR or bare LF as in the original.
basically, converting a body part that is end-to-end transparent
for all octet values to one which is not end-to-end transparent
for all octet values can cause MD5 to fail. no surprise there.
Keith