Thanks Eric. Very useful. I forgot about the Zonal Geometry tool �?? that's a much better solution. I'd also remembered how to do the export to an older version.
I had problems using the in_memory workspace for rasters in ModelBuilder, although it works for Python scripts.
Is there any chance you can explain why it's more effective to use the Radius2 parameter than clipping out the raster? Does the R2 parameter apply a dynamic mask to the layer or something? I presumed that an issue with using the R2 parameter is that you're limited to using the units of the SR of the input raster, which doesn't make for a very versatile tool, and much elevation data isn't projected so it'd be tricky to accurately calculate a distance in DD for each feature. Is that not true? I can see that the raster clip operation is probably expensive and best avoided, but I don't see a reliable way of getting a consistent distance otherwise. Or can you work around that by using the output coordinate system environment setting?