My question now is, which of the coeffecients (1, 2 or 3) gives me the true change of a one-unit change in a predictor variable
I've got several parts to my answer. The first is the most straightforward. You said you're using MinMaxScaler, so I'm guessing that you're using SciKit-Learn and doing this with Python. That package can make a scaler object that already has a method for undoing the scaling called inverse_transform(). There's more information over at stack exchange here or you can read the documentation here. I believe that your manual attempt at an inverse_transform simply didn't perform the right operations, and it is almost always safer to use pre-built well-tested functions from the kit anyway. It has the added benefit that if you'd like to see exactly how it works, it's open source--just read the code under the source link on the documentation site.
Part II may contain some things you already know, but here it is anyway. A logistic regression is non-linear, which means that the effect one-unit change in the predictor differs depending on the value of your predictor. The reason that we're allowed to make blanket statements in linear regression model interpretations, such as "for each 1 unit increase in $x$, $y$ tends to increase by such-and-such on average" is because a linear regression model has fixed slope coefficients. In other words, the first derivative with regard to any predictor is a constant, so the impact of a one-unit change is constant.
That is not the case in a logistic regression. As such, logistic regressions are typically used to predict the chance that a certain observation will fall into a certain category. The coefficients in a logistic regression are not regular slope coefficients that can be interpreted as simple unit-changes as in linear regression, they are logged-odds ratios. This makes logistic regressions much less intuitive to interpret. That much you already alluded to in the question.
However, the reason for performing the transform to begin with is unclear. If you had serious multicollinearity, sure. If the spread of the values was huge, maybe it makes sense. If there were several different units, like kilos, pounds, and inches all being used from different systems that measure different features, again, it might make sense. But I'm not convinced you really need to normalize your variables to begin with in this situation. I'd recommend trying the following instead:
- Do not standardize your variables before you fit the model
- Exponentiate the coefficients you get after fitting the model. This will convert them to odds instead of logged-odds. If you want, you could further convert them to probabilities to make interpretation even easier. The formula is $$probability=odds/(1+odds)$$
- Interpret it as "for each [insert unit] increase in $x$, there is a such-and-such increase in the odds (or probability) of $y$ occurring.
how can I rank the importance of my features?
I'm not sure if you need a formal test of this. If you've got the odds or probabilities, you can use your best judgement to see which one is most impactful based on the magnitude of the coefficients and how many levels they can realistically take.
Can I even compare real-valued features with categorical variables?
Sure, but I wouldn't normalize them first. You can compare different types of variables to each other, just bear in mind the meaning of the different types. If increasing the distance from the goal by 1 meter decreases the probability of making the shot by 1% and having good weather instead of bad increases the probability of making the shot by 2%, that doesn't mean that weather is more impactful--you either have good or bad, so 2% is the max increase, whereas distance could keep increasing substantially and add up. That's just an example, but you can just use common sense like this and explain your reasoning when you're comparing them.