I believe that a predictive model would be the best thing you could create. There are so many variables from tire wear and temperature to road conditions and slope that having a perfectly accurate "gauge" for traction wouldn't be possible. Computer learning techniques could read as many sensors as you can pack on the car and develop a safe "handling envelope" for the car, but it would need to be constantly measuring and adjusting to keep itself accurate.
You could tune the predictive model to be somewhat conservative so you would almost always be safe, but there's always a chance for an unknown wind gust or slippery road kill to send you sliding. And since the system is conservative, your overall performance would suffer.
One issue is that static friction is always higher than kinetic friction. That means that once the tire starts to slide, it will continue to slide unless a force is reduced. That means that in order to really find the limit, you have to loose traction, start slipping, and then change the forces acting on the tire until the slide stops. Modern Electronic Stability Programme/Control (ESP/EPC) systems do just that, and they do a pretty great job. They can allow the car to operate just beyond the safe limits and adjust the forces to bring the car back under control faster than a human.
What you want to do is actually very similar to testing the strength of a wooden board. The only way to do that is to break the board. Unless you break it, you don't know the limit. But, once that board is broken your measurement is just an estimation because that exact, unique board no longer exists. You can measure 100 boards and get a pretty good model of when a board will break, but there will always be very hard boards, and very weak ones that defy your model.
So, in summary, the only way to measure adhesion is to push it beyond the limit and determine when adhesion was lost. You can then use that as a good predictor of when adhesion would be lost in the future, but it will never be 100% accurate because of the number of variables involved.