Simple answer: crapy software!
The software uses a circular buffer, and instead of nulling old images just overwrites them with new images once they become available. If there is a data gap, the old images remain in the buffer instead of being deleted. If the next DSN pass includes data from the onboard tape recorder (which can be several weeks old), the software sometimes gets confused and associates the header with a wrong image file.
I should emphasize that this affects only the quicklook data (quick and dirty near real-time data that are created on the fly for our web pages). The final level 0 science quality data in the SOHO science archive are correct.