After looking at the data leftover after the failing run for WT-5916, we see an anomaly in the data files.  Each data file contains a range of keys (string-ified integers) like "1000000000" through "1000118532".  A program reading this could individually do WT_CURSOR->search for each key sequentially, or it could search to the first key and do WT_CURSOR->next(), and it should give the same results.  However, we see "extra" keys doing next() that we do not see with explicit calls to search.