"The repair of taking the highest value of two inconsistent might cause getting higher values?"

I guess so:

From Sylvain Lebresne: '"Now what the code does to "repair" in that case is to pick the higher of the two value it has. But honestly that's random, there's a 50/50 chance that it will pick the right value."

So this repair has 50% chance to lead into an over-count.

"Maybe even much higher values if it takes place multiple times?"

I don't understand this bug very well. So I can't really answer this. However, I never had over-counts that were so big that their were unacceptable maybe because I got a few (from 0 to 30) of these errors every day out of between hundred of thousandths and a few millions increments every day.