Training the query encoder from answer supervision is the historical detail I would carry into the next part. It makes passage usefulness observable through the answer objective, but an answer string appearing in a passage can still be attached to the wrong entity or relation.
For the hostel-fee example, a useful control would include two notices with the same date for different fees. Checking which notice gets retrieved and which source supports the answer would show why retrieval training and source attribution remained separate engineering problems as these systems evolved.