I'm even more puzzled now. Following Joe's link I also landed on the rfc to which I was refering, which made me take a closer look at it.
That rfc starts by stating the outcome is 16 bytes, and then ends with those results you can test against which are all 32 bytes. All in the same document!
I tried finding in the C code where the jump to 32 bytes is made. It could be something silly like appending a 16 bytes MD5 based on the first 16 bytes MD5 of the original string. But as I wrote before, I'm not good at reading C and couldn't find a single clue.
I need a C master-brain here! ;-)
