I agree that the approach is probably too pessimistic. It essentially takes the view that all uncertainty is downside.

I would be reluctant to fudge the calculation of the mean. Which aspect of the data has more "information" to it? The mean or the confidence? That should guide you about how to combine the two.

To my eye the least invasive treatment would be to compute the means and the standard errors, and then use the standard error to break ties when sorting by the means. Obviously you need the actual values in the population for this (to compute the standard deviation). For the other ones, maybe you can stick in a "default" standard error (don't know what you're scoring, so who knows).