I do not agree. It would mean *each* application has to normalize because it cannot rely on the kernel. It has huge security implications (two file names with the same name in NFC, so visually impossible to distinguish, but two different string of code points).
Couldn't this cause problems if copying two files to a floppy on a system NOT normalising the data (e.g. a customised kernel) with file names that would, when normalised, be identical and then accessing the floppy on a system that DOES normalise the data? Then the second system might think that the two files have the same file name, and wouldn't know which one you're referring to.
Example:
You make two files on system A: "e-acute" and "e combining-acute". You move the files to system B, which supports normalising, and request file "e-acute". System B normalises that to "e combining-acute", and might point to the wrong file. System B thinks that the name of both files is "e combining-acute", so even if typing "e comibining-acute" it might sometimes return "e-acute".
Stefan
