Michael, can you please review the changes in btree/bt_curnext.c? Specifically, I'm now recording the page's write-generation number as part of cursor.next/prev operations. I can't think of any reason that's not OK to do, but it makes me want a second opinion. (Technically, I don't even think the read has any ordering constraints, as the cursor has the page pretty well "pinned", but I did use WT_ORDERED_READ.

Everything else is pretty straight-forward, and it passes the test suite, once you agree on that chunk of code, I'll merge it (or you can).