Peter,
I haven't used scaling much, but here is my answer. First, I would suggest that you start by doing the same experiment that I just did - download a very good model (like the Bell Eagle Eye), select all the components in the Geom Component Browser, go to the Gen(eral) tab, make sure the Scale Loc is highlighted in yellow, type in a .50 (50%) Scale Factor, and click accept. You should see that all the components perfectly scale so that everything stays in the same relative location and you've created an exact 50% replica.
If you now uncheck the Scale Loc box on the Gen tab (just below the Scale Factor label) then you notice that all the components scale separately and each component is not disconnected from the others.
But I believe the result of scaling depends on how the model was built. The Eagle Eye was built with apropriate parent child relationships between the components (to enable rotating nacelles and having the other components move with the nacelles). But if appropriate parent child relationships are not used, surely aspects of scaling could result in 'problems' at different scales. But if good modeling practices are used, with parent child relationship, then scaling works perfectly.
I hope this answers your question. Please let us know if this works for you, and it now makes sense.
Mark