Thanks @jd-foster! Unfortunately I know about this, and while it is the “perfect solution” from a performance point of view, it is not really usable for large scale models - without significant “manual” work (that I was hoping to overcome with some other approach).
Consider the case of a variable (with my coefficient) being added to another expression that is then added - together with other expressions - to a constraint. It’s hard to find out - afterwards - what the contribution of my initial coefficient was (given the fact, that the same variable could occur in multiple of those expressions)… 
Also, “knowing the variable” it is not immediately clear in which constraints it ended up.