If message text is base64-encoded (or uuencoded), bogofilter cannot parse such message. Parsed tokens will be undecoded garbage.
The message have following format
Content-Type: text/plain; charset="windows-1251"
The message is attached.
After a quick debugging, it look like a big logic problem with lexx parser which prefetches lines in advance. When parsers detects "end of message header" event, next line of message was already fetched and buffered by lexx. Since bogofilter still was in "header" mode at the point of this fetch, line was buffered as is, without base64 decoding.