Not OP but good question. The diffs end up being semantic and not merely a difference of triples. This is due at least in part to the open world nature of RDF and concepts such as reasoning and materialization.
We also support a bi-modal interpretation of ontologies - if you put them in a graph of type "schema" they are interpreted in a closed world way, in a graph of type "inference" they are interpreted as being open world.
In our experience the inference stuff is super cool (for example we use a property chain rule in our system database to allow arbitrary nesting of authority domains) but 99% of the effort is trying to make sure that your data and schema line up and for this, you really want to be operating in a closed world regime.
Quick note, it seemed like it supports only closed-world ontologies.