Processing is different depending on the state of whatever object you are processing (a cube, the whole db, a dimension)

Have a look here: http://msdn.microsoft.com/en-us/library/ms175634.aspx

My guess is that in your case processing the cube alone is not invoking a full process of all of the dimensions where there are duplicate keys at source.

Have a look at your dimension design - the error configuration in 2008 onwards defaults to "Custom" which does not allow duplicate keys in dimensions. You can change this to "Default" which allows this and confusingly isn't the actual default, or you can use composite keys in your key attribute key columns to avoid duplicate keys altogether.