Original User: whuber
The result looks correct to me, Dan, insofar as one can discern any details in this rendering. (Hillshading the result can be extremely helpful.) This is definitely not a Euclidean pattern, which would be a series of circles concentric with the destination point shown. What's happening, though, is that CostDistance is perfectly willing--and able--to take shortcuts around ridges when those can reduce the total cost of travel. To verify that this is what's going on, compute the backlink grid and use that to track least-cost paths from several characteristic points within the region.
The possibility of a least-cost path completely ignoring flow channels shows the CostDistance approach to this problem is fundamentally flawed. To track storm runoff you need either to use a flow accumulation calculation or a path distance calculation.