Here's a basic example: Suppose we want to show via BPMN or BPEL that we want to call services A and B to fulfil a business requirement. I would have (for BPEL) a resulting orchestration R that involves the sequence of operations from A and B. When I move to my SCA editor, I want to have a composite corresponding to this orchestration. This composite C will contain inside the services R, A and B with wires between R - A and R - B. The service offered by C is in fact the service offered by R and we would have the service of R promoted to the composite C (wire between C's service and R's service). This is an architectural choice of representing processes that might not be what everybody needs, but it may be a good pattern for SCA.
Obviously the two diagrams (BPEL and SCA) would convey different semantics, each focused on their own usage domain. However, using the intermediate meta-model, we could transport information between the two editors. If we only had the WSDL-specific stuff in there, we would lose the interactions therefore having to duplicate work.
In other words, the "steps" in the intermediate model would mean different things to different editors (real steps for BPMN/BPEL), hints for composition in SCA, hints for assembly in JBI etc.