Unlike other engines, Apertium supports only plain text translation. Since it does not support HTML markup the process of translation involves removing the markup and applying it again to the translated contents. Reapplying the markup to a translated text is a process that cannot be 100% accurate and sometimes references and links may be applied to the wrong place.
Since for this particular case the target location seems quite obvious, it may be a good example to improve the algorithm. But I'd defer to @santhosh who knows the internals better to decide if this is a useful example to improve the algorithm or not really worth it.