So I actually decided to try the splitting feature, to see how much space it would save on a Kindle loaded with converted ePub books. The answer turned out to be about 11.5% on my corpus of 1277 books. But that's not what this post is about…
I compared a KF8 stripped by KindleUnpack with one generated from the same original file by Amazon's Personal Document Service, and found that, aside from some minor changes in the metadata (including addition of the atv:kin:1 tag that they harvest and upload to track documents), it does something different with DATP sections near the end of the document. There are two in the original file and the KindleUnpack KF8, but one is removed in the Amazon KF8, and it's put in a slightly different location in the file. I wonder whether this matters. Do you have any idea what this section is for? It looks like a table of offsets.