@michaelcahill, it occurred to me we should be able to skip the same prefix bytes on the child page we skip on the parent page, that is, there's no reason to clear the low/high prefix byte counts when we descend the tree. It's not going to make a performance difference, I don't think, because the first 2 comparisons on the child page would reset the values, but it still seems like the clean thing to do.

It's a little scary caching information across a page switch (!!?), but I can't think of any reason it's not going to work, if it didn't work, it seems to me there's tree corruption.

Wanted to get a review, though, just in case there's some edge case I'm not thinking of, or there's a race when switching pages that could mess with this?

Ref WT-595.