I agree with Don and Dave. I would think that a small number of generic callbacks is more likely to be both useful and easy to understand for OpenACS developers as a whole.
Callback implementors would have to accept and process data about the event that has been triggered (the state change). But the result is less callback implementations, which surely will make maintenance easier. Besides, a disincentive to use a callback to solve any particular problem is probably a good thing.
The performance concerns (generic callbacks being called more often) are valid. But I think the guiding principle should be governed first by ease of maintenance. Perfomance concerns should be considered in specific cases and with supporting data. So as we are only starting to introduce callbacks into core I would suggest implementing them the way we want first and assessing the impact on performance as it becomes apparent on real sites.
