Hisanori Furusawa and Sogo Mineo report that most of the WCSs produced by meas_mosaic in hscPipe 6.x (all of those that correspond to sensors with rotations relative to the focal plane) are incorrect - the SIP coefficients are dropped after rotation. This is seen both in the headers and in the binary table.
After a quick look at the code, I suspect this is only superficial, but still confusing: we rotate a WCS by prepending its mappings with a rotation using AST, which yields a WCS that cannot be represented as TAN-SIP without loss of precision.
First piece of this ticket is to check that hypothesis. The second (assuming the hypothesis holds) is to come up with a way to reduce that confusion.