Good question. I think it is worth noting that this behavior is exhibited because nested transactions do not truly exist in SQL Server.
I have to disagree with you, Jason. Nested transactions do exist, and they work exactly as I think common sense dictates.
Transactions are primarily intended to implement the A and I of the ACID properties: atomicity and isolation. Nessting fringes with that goal, and I can easily imagine that the people on the ANSI committee have considered not allowing nested transactions. But that would invalidate many common use cases. If I write a stored procedure that does multiple things but should be considered as a single unit of work, I use BEGIN TRAN and COMMIT TRAN (or ROLLBACK TRAN in case of error) in the procedure code. But what if I next have to implement a stored procedure that is also considered a single unit of work, but that includes the first stored procedure? I use BEGIN TRAN/COMMIT TRAN in the outer procedure; I call the inner procedure, and there we have the nesting. Prohibiting nesting would requiere me to duplicate the code, which I obviously don't want.
Support for nested transactions is unavoidable. But what are the "most correct" semantics? Consider the example above. The inner stored procedure implements actions 2.1 and 2.2, as a single unit of work. The outer stored procedure implements actions 1, 2.1, 2,2, and 3, also as a single unit of work. And it does so by caling the inner procedure.
So what should a nested COMMIT do? It can not really commit the changes of actions 2.1 and 2.2. After all, action 3 might still fail, and in that case the whole outer procedure needs to be rolled back, including actions 2.1 and 2.2. Otherwise, the outer procedure would not be atomic. So the only thing the COMMIT can do is to decrease the nesting level counter; actually committing the data has to be postponed until all nesteed transactions have finished.
Conversely, a ROLLBACK in the nested transaction should roll back ALL open transactions. If it would only rollback the effects of the nested transaction, the outer procedure could continue to perform action 3 and commit. IN that case, actions 1 and 3 are committed, but actions 2.1 and 2.2 are not. The procedure is no longer ACID. Making any rollback, regardless of nesting level, roll back ALL open work prevents that problem.
This question is actually almost similar to the one two days ago. The only real difference is the use of the extra keywork WORK (that is not required).