To get the most out of compositing, you’ll need to carefully plan every render that will provide inputs to the process. You want every channel/layer of information to be produced separately. (Which the EXR format is capable of storing in just one file). Alpha channels should not be premultiplied.
Do not use lossy-compression. Do not use “image” formats (especially not JPG). Floating-point numbers, big sizes. Terabyte hard drives are dirt-cheap now. Splurge. Save every version of everything. Scrupulously keep a written log.
So the result of “all that rendering” at this point is not “an image.” Rather, it’s a bunch of masks and maps. Color maps, specularity, opacity (alpha), reflections, shadows, z-depth, surface normals, velocity (maybe). Every one of them clean and distinct.
“The image is then made in the digital Darkroom.” Just like in the days of yore. Just like a mix-down in a recording studio. A network of nodes (a “noodle” in some circles) takes these inputs, and separately filters them and adjusts them and blends them, and uses one to modify another, and so on and on (the sky’s the limit…) until … voila! a finished picture appears. (Simple “layers” are not enough: your system undoubtedly supports both metaphors. Use “noodles,” by any other name.)
And here’s the sweet part: once all that laborious rendering is finished, it doesn’t have to be repeated. You can adjust the node network and see the result in real time. If you find that a particular piece of it is hosed, you can selectively fix that piece. It looks daunting at first but there’s no other way to do it.