The first thing that came to mind was "Newton," which used stroke-recognition for recognizing letters. I doubt it's actually computing letters on the fly, I go along with your thought that you precompute vectors for each word in the dictionary. My thought is that the vectors are probably very simple. They may start with a specific letter, but after that I'm guessing they are very forgiving of inaccuracies. A rough direction and distance is probably enough to disambiguate words from each other.

I'm guessing it precomputes a tree for the dictionary. With each stroke, it drills down the tree along the most likely path. If it runs out of nodes, it backtracks and tries a more likely path somewhere,possibly using a precomputed list of vectors most likely to be mistaken for each other.

That gives On for the best case.