Of course, changes to the network would mean that they would have to be re-evaluated, but still.
Not necessarily. I mean, the trip time could be kept separate to the pathing (because most of the time they’re just using a stored value to decide whether to path by car or not) and although it would mean that they might choose the train even though their usual route is clear after an alteration to the network, or there’s a newer, faster route, this would only last a few trips until they tried the car again so it’s not really a big deal.
What I’m basically envisioning is that every X trips they would run the normal routine of deciding whether to go by car or whether public transport is faster but then the actual travel time would be recorded and they’d then use that value for X trips before running the normal routine again, setting off in their car, the time being recorded again and repeat. This decision on whether to take the car or not necessarily comes before they do their pathing, so alterations to their route would not necessarily need to be detected and recalculated immediately, they would just have a delayed effect. Now, if they decide that car is faster because of the stored value, but then they fail to find a path (because a critical road has been demolished for example) that could just cause the normal routine to run immediately again.