I must be missing something fundamental here.
In a 'stacked normal form 2 person zero-sum game' we can solve for the NE strategies and payoffs recursively. If we have access to the full tree we can solve for the leaf nodes using linear programming, and since a 2 person game will have a unique NE payoff we can pass this up and repeat this process back to the root node.
From my reading about CRM it seems we must also have a complete tree (although realistically it may be an abstraction of the actual extensive form tree) and at each step of the algorithm we perform a similarly recursive forward-and-backward pass through the tree. However, CRM is a convergent algorithm that requires many such iterations, where as the linear-programming method will solve the game precisely in one pass (although it is more intensive at each node).
So why would we ever use CRM? Surely I am missing something. In both cases we have chance nodes, but that does not fundamentally change the above considerations. Is it pertaining to the concept of 'information'? I am exclusively talking about 'stacked normal form 2 person zero-sum games' here.