Stem's DescriptorReader misses 10% of descriptors in tarballs

When I parse server descriptors in a metrics tarball and in extracted form using DescriptorReader, I get different numbers of descriptors:

from stem.descriptor.reader import DescriptorReader

descriptors = 0
with DescriptorReader('server-descriptors-2012-12.tar') as reader:
    for descriptor in reader:
        descriptors += 1
print "%d descriptors in tarball." % (descriptors, )
descriptors = 0
with DescriptorReader('server-descriptors-2012-12/') as reader:
    for descriptor in reader:
        descriptors += 1
print "%d descriptors in extracted directory." % (descriptors, )

250048 descriptors in tarball. 279042 descriptors in extracted directory.

What happens to the rest?

This bug means that most of #7828 (closed) must be re-run. :(