I recently built a small Python library to try getting time management right [1]. Exactly because of the first part of your comment, I concluded that the only way to apply a time delta in "calendar" units is to provide the starting point. It was fun developing variable-length time spans :) I however did not address leap seconds.
You are very right that future calendar arithmetic is undefined. I guess that the only viable approach is to assume that it works based on what we know today, and to treat future changes as unpredictable events (as if earth would slow its rotation). Otherwise, we should just stop using calendar arithmetic, but in many fields this is just unfeasible...