I'm fairly sure that this is down to leaving the 'Accepted error' field blank.
Add in an accepted error and all is OK.
My specification did say how to handle the values when there is no accepted error and we clearly need this as a usability aid to authors. Anyone like me would know never to do an exact equality on 'real' numbers using a digital computer. And I guess that's why the qtype passed all my tests. But the specification for accepted error states:
"Nothing specified means answer must be the same – within reason for real numbers e.g.answer*1e-6" and I confirm that this isn't working.
Jamie and I will pursue this as a matter of urgency, though the record will be on an OU bug tracker. We'll post back here when a fix is in place.