It’s easy to figure out : don’t use modules that produce halos, they are inherently broken (masking and blurring in Lab cannot work). Not every bit of darktable is clean, I’m afraid, but you don’t have to use all the soft to get good pictures.
The layman’s explanation is “use this to do that” in this context, because it’s really all about maths and the link between what you do on RGB code values and what it means for photons in real-life. Remember most of image processing is just simulating numerically what use to be done with glass filters and light sources back then. Then, it depends whether you blend scene-linear data, or display-linear/gamma encoded data.
For example, the overlay blending mode expects display-linear data (where 0.5 = 50 % means middle grey) to work properly, whereas the multiply mode does not make much sense in display-linear (overlay is actually a tweaked multiply mode with a threshold at 0.5 and a sigmoid above). Soft-light is another variation on the same theme. They are often used to blend frequency-separated layers back :
- blur(image) = low-frequency,
- image - low-frequency = high-frequency,
then work separately on both (increase contrast on HF to get more perceptual sharpness, or decrease it on LF to smoothen skins without loosing texture), and merge them back with overlay when you are done (or soft-light for a more subtle effect). That’s exactly how the sharpen or high-pass modules work in darktable, and in most other softs.
Another example, multiplying a (scene-linear) image by a solid color layer simulates a real-life color filter (in the days of film, they liked to use yellow filters on their lenses while shooting B&W to get darker skies). On the opposite, dividing by a solid color allows to remove unwanted color cast (pick the color you want to remove).
So, I don’t think you can avoid the maths overhead… Sorry.