Hi, Bart,
I'm still having trouble imagining how correction for geometric distortion applied to the sensel field (the raw data) would be more effective than correction for geometric distortion applied to the pixel field (the developed image) - with regard to geometric distortion per se.
While geometric distortion in general may currently be complicated to correct before demosaicing, think about a specific 'geometric' (or rather projection) distortion, Lateral Chromatic Aberration. The three colorbands have what seems like a different radial magnification (on top of pincushion and/or barrel distortion). If we first re-align the colors, we will get a much more accurate color interpolation for the 2 missing color bands in each sensel in the demosaicing phase of the Raw conversion, and a higher sharpness.
But perhaps the demosaicing will work better on a sensel field that has had the geometric distortion already removed. (I can't yet imagine how, but of course these things are very complicated.)
It is indeed complicated, which means that research is ongoing and may still produce improvements. Of course there is also the trade-off between speed and diminishing returns in quality improvement. As processors get more powerful there will be an opportunity to use more complex algorithms which were deemed too slow for practical use although the quality was higher.
I can slightly imagine (if I stretch my imaginer) a better result in the matter of correction for vignetting operating on the sensel field.[/QUOTE]
Vignetting correction is mostly a matter of simplifying calculations with linear gamma image data. One basically normalizes a flat-field exposure (at a given aperture and focus distance for a given sensor) to 1.0 for the brightest area, usually in the center for non-T/S lenses, and divides the image data by the flat field. As long as it is done in linear gamma, it's easy, and it is more accurate than having to temporarily invert the gamma with rounding errors, divide, and re-introduce the gamma. The colors/saturation will also shift less when the corner brightness is boosted by perhaps a stop or more, before demosaicing.
All sorts of image data math is easier and more accurate in linear gamma (e.g. resampling), and some calculations (also noise reduction) on the Raw Bayer CFA data will also improve the demosaicing quality.
Cheers,
Bart