Why is only the final logged value counted towards min or max of a metric?

I wanted to find the run with the best validation accuracy, however, I noticed that to calculate the maximum only the last logged values are used.

You can see this on the screenshot below - it says the run with the maximum validation accuracy is run ‘21-…’ (the one that stopped earlier), even though run ‘18-…’ had a higher value one epoch after the other ones end. The problem is that one epoch later run 18 dropped in accuracy again, so it doesn’t have ‘final maximum’ accuracy…

This is a bit problematic for me - I’m saving my models after each epoch, so I don’t really care only about the last model - precisely to prevent such a problem where the accuracy would suddenly drop at the end.

I think this is a bit similar to this post: Can I plot the value of a metric at a single step?

Is this a feature, or is this a bug?