It looks like KG 2.5 puts more than one StartOffset in the exth metadata:
In Doitsu's example, the TOC, then the Start.
So, yeah, if what's in there isn't bogus (and it doesn't look like it), and the K3 chokes on this (or only honors the first entry: you start on the TOC instead of the 'proper' start), probably a K3 bug. (Calibre only puts a single entry in there: the start).
On Mobi Unpack's side, it was doing some stuff with it for KF8 files (pop'ing the list to get the last one), so you ended up with only half of it in the OPF meta. And it was indeed building guide entries from those, without a title, even if the proper guide extraction was successful. (On the other hand, it later tries to do this again a dozen of lines after having deleted it o_O).
Here's a small patch against MU 0.54. Take with a grain of salt, it's 7 AM, it was mostly to try to understand what's going on, and why I had an incomplete header dump (saw the dual StartOffset with Calibre but not MU), and I'm a bit shaky on the whole OPF thing to begin with
I'll check tomorrow on my Touch how it handles those files with dual StartOffsets...
Also check if KG 2.5 stops at two entries (TOC & Start), or if will happily create one for every reference in the guide...
FWIW, I dimly remember seeing those kind of double StartOffsets on stuff straight from Amazon, so a KDP pass probably doesn't change a thing
EDIT: Updated patch in #41