dynamodb-lock-client | general purpose distributed locking library built on top | Application Framework library

The race you are suggesting is very surprising, because it is exactly what DynamoDB claims its conditional updates avoids. So either Amazon have a serious bug in their implementation (which would be surprising, but not impossible), or the race is actually different than what you described in your question.

In your timeline you didn't say how your code resets "StartedRefreshingAt" to nothing. Does the same UpdateTable operation which writes the results of the work back to the table also deletes the StartedRefreshingAt attribute? Because if it's a separate write, it's theoretically possible (even if not common) for the two writes to be reordered. If StartedRefreshingAt is deleted first, at that moment the second process can start its own work - before the first process's results were written - so the problem you described can happen.

Another thing you didn't say is how your processing reads the work from the item. If you accidentally used eventual consistency for the read, instead of strong consistency, it is possible that execution 2 actually did start after execution 1 was finished, but when it read the work it needs to do - it read again the old value and not what execution 1 wrote - so execution 2 ended up repeating 1's work instead of doing new work.

I don't know if either of these guesses makes sense because I don't know the details of your application, but I think the possibility that DynamoDB consistency simply doesn't work as promised is the last guess I would make.