In some cases, published content with Content Translation leaks cx-segment metadata elements into the published content. For example, in this article published with version 2 of Content Translation (on Feb 5 2019):
<ref> <span data-segmentid="129" class="cx-segment"><span title="A volte può capitare che un link presente su Wikipedia non sia più raggiungibile...
This abuse filter can be useful to get example and check if the issue persists over time.
These were created long after the similar issue T113137 was resolved, so it's not a caching issue. I am marking this as Regression, even though the thing that caused it in this article might be different from what was addressed in T113137.