Given that one of the best attributes of Lightroom's Develop module is the ability to apply edits to a raw file, the most desirable approach would be to be able to maintain that nondestructive workflow when using features based on machine-learning like denoise, fine details, and super resolution. In other words, for those functions to perform their magic without demosaicing the raw image and emitting a linear DNG.
I don't know whether that is a feasible objective. I suspect nobody external to Adobe does. But I often find I want to do some additional tweaking to the derivative file either because the enhancement process has affected my prior edits or because I simply have changed my mind about one or more of the earlier parameters. So it's either necessary to start over or to wind up with two independent sets of edits: one for the raw file and one for the linear DNG.
Having said that, I don't think it is inherently impossible for the current machine-learning denoise process to work on a demosaiced file. If I demosaic a file with DxO Pure Raw, I can apply Lightroom denoise to DxO's linear DNG. Since the DxO product provides its own noise reduction, I have doubts about how useful it is to subsequently invoke Lightroom's denoise. However, the fact that it's possible at least suggests that denoise is not confined to operating on raw image data.
For now, when I need both denoise and increased resolution, I use Lightroom's "Edit in" dialog to launch Topaz Gigapixel. For most images, my experience has been that Gigapixel produces enlargements that are roughly of the same quality as those produced with Lightroom's super resolution. Viewed at 1:1, sometimes one approach or the other seems marginally better—but for most purposes the differences can be ignored.
