"During the heat of a Planning Poker session, it can be very tempting to decide to shorten some of the discussion by using multiple estimates instead of a single value. For example, if testers and programmers are having a hard time coming up with a single estimate for the full work, someone may suggest putting “programmer points” and “tester points” on each product backlog item. This is almost always a mistake."
After reading the full post I think the key idea is that you can't really trust composite estimates, made by several contributors, working on what is ostensibly the same job. Any estimate and re-estimate must be down to the judgement of the person who signs up for the item, even though it is implicit that they will necessarily need to involve other specialisms. As I see it estimation shouldn't be a back-door to task atomisation. The correct reading of estimation is that it offers a process of sharing knowledge and, in conjunction with the standup meeting, shared support. Ulimately we want a person who 'takes it on' also takes responsibility to make sure it gets done done, not 3 or 4 people all pointing fingers at the other to blame for a problem.