The accuracy tests added for SpectralCoord in #10185 are currently limited by the precision of the output files from Starlink (currently we test down to 20m/s). It would be good to create more stringent tests of the accuracy of SpectralCoord transformations to different velocity frames. Just to be clear, I believe the implementation in SpectralCoord is more accurate than 20m/s, but this is currently a limitation in the accuracy testing.