IMHO associating keywords with acs_objects instead of cr_items is absolutely the right answer. I agree, the fact several packages have their own keyword/category implementations is a good indicator that the service needs to be generalised - so that, for example, my breakdown of bboard posts by "sport" is equivalent to my breakdown of news by "sport".
The question of whether all content should be stored in the content repository goes back to my first post.
If you regard the content repository as a convenient group of services, then it is no longer a question of whether content is in the content repository - it is a question of which services do you want to provide to this group of content.
Currently if I want my news to go through workflow, be searchable and have keywords assigned but I don't want the overhead versioning, I have no choices. If I could pull any combination of the distinct services together to form my own content repository then I would have more flexibility.
I know that this flexibility would come with a cost in terms of system complexity. I believe, for example that the abstraction of services is a major factor in the performance and learning curve problems present in ACS JAVA. I still believe however that categorisation is a service that should be system wide - and that its functionality should be made available to end users more often ;)
