An odd statistical anomaly was seen in WiredTiger where a reconciliation was reported to take 96 hours. The usual suspect in this is a long checkpoint > 96 hours, but we didn't see that here.

We should investigate this to understand potential causes:

1. Is it a statistics implementation bug? e.g. Concurrency issue.
2. Is it an accounting bug? Are we starting the timer before we enter a lock?

There may be other causes.