Sensitivity propagation when one "independent" variable depends on another are problematic.
In the case where we have the graph
graph TD;
A-->x;
x-->y;
and we want to compute \frac{\partial A}{\partial x}
and \frac{\partial A}{\partial y}
then we end up with a \frac{\partial x}{\partial y}
which I haven't previously considered.
This shows up in a few applications involving PoKiTT where, for example, the temperature (an independent variable) depends on mass fractions (also independent variables).