Does anybody know why it wasn't implemented in a backwards compatible manner?
One could wrap a whole merkle-tree with an additional extension tree, that just adds the new hashes. That way both kinds hashes could be used to traverse all data. The new hashes could be used to check the consistency, the old hashes would still be there to use in UIs or old release documentation. The downside being that you introduce more nodes in the overall data-structure which will have to be supported basically forever. And if SHA-256 is to week a third layer would need to be introduced. But the point is, it could be done. Albeit it would loose some of the elegance of the data structures involved.