This is why, rather than reading text, I think people should describe these problems with code. Unambiguous code. With test cases. Then everybody can inspect the unambiguous code rather than having to parse human text.
Indeed you should switch if you think there's any chance at all that Monty always opens doors with goats. (Unless of course you think that there's also a chance that Monty always reveals cars whenever he can, but that wouldn't make for very good TV.)
Additionally, to open the door with the prize would terminate the game, so to open the door with the goat is the only action that makes sense given what we know.
Or you know, somebody could just write a computer program that described the rules unambiguously so everybody could inspect them and not have to make reasonable interpretations.
After many years, I finally understood where I made the mistake. It's not stated explicitly in the wording of the problem, but if you point this out explicitly, nearly everybody changes their mind:
When Monty Hall opens the door, he has to pick one of them, and he has knowledge of what's behind it.
The way the problem was described, I didn't appreciate that.
It wasn't until one of the groups that wrote in to support Vos Savant explicitly said "since some of the students in the class were skeptical, we wrote a simulation of the problem, which output the probabilities you would expect if Vos Savant was right." I inspect the source code, and viola- instantly understood my false assumption about the problem description.
It's an open problem whether the Monty Hall problem should explicitly state the implication of which door Monty opens. To me, I think the whole point is that it leaves one of the consequences unstated, and the reader is expected to make that indirect inference.