A bit more digging into each individual shape in this shapefile seems to indicate that the struggle here relates to the number of parts within each individual MULTIPOLYGON. So row 2 has whereas Shetland, row 244 has 549 parts row 387, has 521 and both lag in similar ways.
I'm a bit out of my depth here in terms of identifying faults within individual polygons, but it seems to be the case that (as suggested some specific shapes are tripping this up, by my testing of the first four, geo_codes: W06000016 and E07000138 (1 and 4) are fine, whereas S12000013 (2) is causing problems.