It turns out that some chips (for example the Agilent ones) have a
several features that do not get it in the annotation package.
For example, hgug4112a (human whole genome array) and has 43931
features, while the annotation package knows about 41000. The missing
ones include the control probes as well as some truly obscure probes
that are almost unannotatable.
If it is in the 41k and no data is found, then mget will return an NA,
so the "ifnotfound=NA" argument gives a proper behavior (except that no
name is attached to it).
It is a bit frustrating to have to filter those away before calling
mget, or a function which calls it.
Or maybe that would be considered a problem with the annotation
packages? I still think that having mget calling stop whenever an
unknown probe is used is pretty harsh (an NA and a warning would be more
reasonable, I think).