Interesting... All the data around that blip is valid. However there appears to be an extra 2 byte header, with no data throwing off the offsets for the remaining data. The Transformer is designed to recover from this, and appears to do so, but not until after a record with a 2 byte shift is written to the log.
So, it seems it is in Shadow Logger that it is allowing a header to be written on an empty array for some reason. But the FRD transformer could / should also have a filter to catch the bogus record.
Either way, for this instance, there are 2 fixes, I shall put in both.