I have abandoned Datasets as a way of organizing a GDB. I feel the performance and usability are awful. I have all feature classes together at the "root" of the GDB, and organized by using a prefix, either by domain (publicworks_*, hydro_*, etc.), or by department. But prefer by domain. Users can easilly browse the entire GDB alphabetically, and zoom in to the domain/dept. they're looking for. It seems counter intuitive but this really seems to work faster, at the software level as well as for the users.
Also, having several GDBs is an awkward solution for me as I imagine there will be a huge increase in connection files for the users, and connections to the databases in the server(s), since arcmap and arcgis server will require new connections for each database referenced in a mxd. Also, I am not sure how the rdbms server can optimize its cache mechanism when the data is divided into several databases. Maybe SQLServer has this solved, I am not sure.
To find and load data, we are using a more friendly approach for the users, by pre-producing lyr files. On 90% of the use-cases users don't have to connect and look for the correct feature class, and choose an appropriate symbology and/or definition queries. They just load a certain lyr with an easy to identify name. Actually we have an addon in ArcMap that shows a tree of themes. The users double click one theme in the tree and the corresponding lyr is loaded.
Getting back to dividing data according to editing rights, I feel that arcgis would benefit from supporting inheritance in Feature Classes. Many times editing responsabilites on a single FC are geographically defined. This makes it really difficult to have a good solution, that is not a maintenance nightmare. Having spatial inheritance would be a really cool addition to the GDB capabilities.