It was correctly identifying reads whose “paired” flag was set, but then
it was incorrectly skipping over the read-pair that follows such an
incorrect read.

Also added a test for this behavior.