> [...], but the way to do that is _not_ to make
> all labels 'sticky' when first they are read from disk!

FWIW, that was actually never the case.  I was also very careful to
only cache the info in cases where the driver could sense a disk
change.

The *only* time this is a problem is when you change the label behind
the kernel's back (so to speak).  I'm not sure I consider this even
vaguely a problem -- especially since you have to do an ioctl to
enable writing of the sector with the label anyway (well, on most
platforms).  Perhaps this operation could just clear the cache or
explicitly reread the label...