In T295834, we stripped all annotations from template arguments. While this is valid for the Translate extension, which doesn't rely on Parsoid output to process the annotations, this is sufficient.
If other extensions use the annotation mechanism in a way that involves the Parsoid output, we need to identify a mechanism for recording these stripped annotation ranges so the extension implementing the annotation can extract them for its own processing. Stuffing them in the data-mw object in some form is one solution. A spec needs to be proposed and implemented.