Does it use just some generic manufacturing info or can it take into account some programmed calibration data stored in the lens, as predicted by Roger Cicala in “This Lens Is Soft” and Other Facts?

Yes. It uses both. In comments to some of his more recent blog posts Roger has discussed what happens after a lens' optical alignment is adjusted as part of a repair and how that makes the calibration information stored in each lens at the factory obsolete. This applies to not only Canon lenses, but many other makers as well. Particularly smaller mirrorless cameras tend to automatically apply lens corrections to allow lenses to be smaller and lighter by not correcting everything optically.

Can it automatically detect e.g. the exact strength of chromatic aberration using some sophisticated algorithms, directly from the RAW image?

There are manual adjustments available to fine tune CA correction when not using DLO with DPP 4. When DLO is enabled, those options are greyed out and not available. One would assume DLO applies CA correction based on the measured characteristics of the lens.

Can DLO fully overcome the effects of diffraction or is diffraction still an issue with small aperture diameters?

Nothing can fully overcome the effects of diffraction, particularly when one is approaching the diffraction cutoff frequency (not to be confused with the much wider diffraction limited aperture when diffraction begins to be perceptible when viewing an image at 100%). The ability to reduce the effects of diffraction at moderate apertures has been described as pretty remarkable by numbers of well known photographers.

A post at Fred Miranda discussed this shortly after it was released concurrently with the 5D Mark III and the example pics, intentionally shot at f/16 to test how well DLO deals with diffraction, are fairly impressive, especially given that on the forum post they have been re-sampled (by tinypic).

When it was first introduced back in 2012, the downside to DLO is that applying it took some intensive processing and the size of the RAW file was doubled. If you later turned DLO off for that image, the RAW file returned to the original size. This is no longer the case, as applying DLO seems to take much less processing power and it does not significantly increase the size of the raw file.

If applying DLO increases the amount of detail in the photo, it can increase the size of a converted JPEG. I tried it with a 31.4MB .cr2 file from a 5D Mark IV. The image was shot at f/6.3, so it was still below the DLA of f/8.6 for the 5D mark IV. With quality compression settings at "9" and all other settings the same, converted JPEGs were:

• 7,878 KB without DLO turned on.
• 8,501 KB with DLO turned on.

Pixel peeping at 100% there appears to be a bit more detail in parts of the image with DLO enabled than not, but the difference is not very significant. I might try to shoot something at around f/16 and f/22 and compare the difference for such images. If so I'll update this answer to reflect the results.

Is the DLO on the camera the same as it's on the DPP software, or does the DPP software use the additional processing power available on a computer when compared to that of a camera?

Only Canon knows for sure, but the effect of using it seems to be very similar if not exactly the same. I've not used it much in-camera because it slows down processing and doubles the size of the raw files written to the memory card. Since a lot of what I do is sports/action/performing arts for which a fast handling camera is paramount and because I almost always save raw files, it's not worth the speed penalty to do it in camera for me. I rarely feel the need to use it when processing raw files with DPP 4, but again, the type of photography I do most of the time doesn't really require it. The more basic CA, Peripheral Illumination, and Lens Distortion corrections offered by DPP 4 are usually enough for my purposes.