Nice article which describes how the tail ratio (and other criteria) can help assess the likelihood of future success of trading algos:
https://blog.quantopian.com/using-machine-learning-to-predict-out-of-sample-performance-of-trading-algorithms/

Has anyone ever heard of an add-on to NT7 that would allow to optimize algos based on the tail ratio ?

Interesting. "Mean" is overly effected by outliers, and things like sharpe or standard deviation assume normal, symmetrical distributions. How many times have you heard about cutting your losses and letting your losers run. Well that's the opposite of a symmetrical distribution so it would make sense that something like 'tail ratio' is a good predictor. Of course the fact that “user_backtest_days” ranks so highly (I assume as a negative predictor) probably means that most systems are overfit, or in other words 'bad' systems. I doubt you would get the same results if your universe of designed systems were better.

Sorry doesn't answer your question.

So, "tail ratio" is bigger when returns are positively skewed? As in, a few big wins and many small losses? Is that what it means?

If so, then higher tail ratio would probably mean that you're using a trend following strategy.

So this post is saying that trend following strategies have better out of sample results than mean reversion strategies?

Interesting analysis but possibly flawed unless they chose different periods for all the backtest, if trend following did better during the subsequent period it would tend to favor those strategies. You need to mix all the dates up so that all the OOS periods are different.