Ocelot seems to “break” content that Memsource puts into CDATA blocks by turning certain characters into their HTML entities. Now it seems that this is not a huge deal for Memsource (we cannot open the resulting XLIFF file in their desktop editor, but it imports OK into Memsource and so far we did not see any issues with this), but just to be on the safe side, would it be possible for Ocelot to ignore CDATA sections?
The CDATA is part of a Memsource specific <m:in-ctx-preview-skel> element which gives the ability to have a preview of the file.
This would probably need to be fixed in Okapi. There is already an option added in M33 that tweaks how to handle CDATA that appears in source/target content, but I don't think there's an option to just ignore the CDATA.