1) There is no requirement for a specific spatial reference and the tool will do the reprojection if needed. Even when you download/import the OSM data there is now requirement to store the data in WGS84 in the geodatabase. You can specify
a different spatial reference in the geoprocessing environment and the tools will respect that.
2) You can use either one. When you are working against *_osm_* feature classes make sure you are removing "supporting elements" from the display. You can apply a definition query with "supportingElement = 'no'".
You can do both but you'll have a nicer editing experience if you go through the themed layers - IMHO.
3) Yes, you can continue to use the same mxd/geodatabase. However keep in mind that the upload tool only uploads your changes and reconciles them with the OSM server and brings the local geodatabase up to the reconciled state. It doesn't request or get updates
to other features made by other users. Eventually you would want to do a full sync with the OSM server again. On how you do it depends on the size of OSM data you are working against. If you are working within your neighborhood or a suburb then re-downloading
whole extent might be faster. If you are working against the state of Colorado you probably want to use the Load Diff File tool which reloads only the deltas against your extent.
4) We do have a workflow outlined on how to do it at
http://esriosmeditor.codeplex.com/wikipage?title=Contribute%20Your%20Dataset%20to%20OSM&referringTitle=Documentation. This workflow is not set in stone as there are a lot of moving parts and we are always curious about feedback from users what worked
and what didn't work for them from the workflow perspective.