I don’t see how the Denoise process has rendered anything more than any other tool in Lr. It’s still a raw file till you export it or chose to edit it in Photoshop at least I believe it is.
On reflection, I shouldn't have used the word "rendered." The file produced by the previous implementation of Denoise was a scene-referred linear DNG: demosaiced, but without color, tonal, or gamma corrections applied. So for subsequent post-processing, it was mostly possible to treat it as a raw file.
The advantage of that approach is the one Andrew pointed out earlier in this thread. The derivative file was independent of the Lightroom catalog. It could be backed-up separately from the raw file—and optionally edited in a program other than Lightroom without losing any Lightroom adjustments that had been applied before it was created.
The advantage of the current approach is that it preserves the functionality of a raw workflow. All the edits performed in Lightroom are recorded in the catalog data and can be retrospectively modified—at the expense of a little catalog bloat.
Disclaimer: I have no information about the internals of Adobe software so what follows is conjecture.It seems that some and possibly all of the Lightroom functions that are based on machine-learning are performed on a demosaiced file. The demosaiced data may be stored in an independent linear DNG or in the Lightroom catalog. I suspect the reason Adobe changed the behavior of Denoise, Raw Details, and Super Resolution was that storing the information for those modifications in the catalog was consistent with the "non-destructive" Lightroom design philosophy of using a database for a list of edits while leaving the raw file untouched. The illusion of a pure raw workflow isn't perfect, however, since in the current version of Lightroom, you can't apply Super Resolution after Denoise or vice-versa.
