What about bridges and tunnels? It seems like an ideal (or idealistic) solution is to clean up the source data to include intersections when there are intersections so they don't have to be inferred. Is that impractical or would it explode the size of the dataset?
Bridges and tunnels are still represented as linear features in most road datasets. In good ones, there's an attribute to mark it as a bridge or tunnel, and you can choose to use them or not.
Including intersections wouldn't significantly change the size of the dataset, but that type of clean-up is best done as a pre-analysis step. There's no good reason to do it in the "raw" data.
Furthermore, it's dependent on the projection/etc that you do it in. The intersection of two lines in lat/long space is not at the same point as the intersection of two lines in another lat/long space with a different datum, and it's not the same as the intersection in a projected coordinate system.
Ideally, you'd only include the intersection if you've actually measured it at that location. It's best to leave the observations as observations and try not to alter them too much.
At any rate, by the time you're analyzing the data, you've already made the type of assumptions about cartesian vs. spherical vs. real space that I'm referring to above. Therefore, it's usually a good idea to defer cleanup that's specific to a particular operation (this one is) until you're doing that operation.
