Thomas McGlynn writes:
One that comes to mind is whether use of this
keyword should be recommended only for the primary header of a FITS
file. If not then a file may not be associated with a unique dataset
id.

Is that a problem, or actually a desirable characteristic? I could imagine
a single FITS file containing data from several origins in different
extensions (e.g. M31 images from IRAS, ROSAT, and the CfA CO(1-0) survey)
where different DS_IDENT would be apropriate. Or should such a merged
dataset be given a new DS_IDENT and forget about the original ones? At
first sight at least, retaining the original pedigree sounds like
a good thing.