Just in case anyone's interested I found that it was possible with certain types of collisions to get an object moving with such great angular momentum that the simutlation eventually clamped the value and turned off physics for the object. This looks like the object just freaks out and fires off screen after a couple collisions or so.
Yes, also you get bunch of LCP error messages usually at this stage. That damping is usually something that occurs with geoms that do not have well concentrated mass, e.g. long narrow boxes tend to spin off widly. I assume you get similar problems if you play too much with the inertia matrix.
Actually the "right" way to reduce this problem is this:
void dBodySetFiniteRotationMode (dBodyID, int mode);
This function controls the way a body's orientation is updated at each time step. The mode argument can be:
* 0: An ``infitesimal'' orientation update is used. This is fast to compute, but it can occasionally cause inaccuracies for bodies that are rotating at high speed, especially when those bodies are joined to other bodies. This is the default for every new body that is created.
* 1: A ``finite'' orientation update is used. This is more costly to compute, but will be more accurate for high speed rotations. Note however that high speed rotations can result in many types of error in a simulation, and this mode will only fix one of those sources of error.
See also dBodySetFiniteRotationAxis.
In any case, I read a little on the ODE mailing lists to see how to do linear and angular damping and they simpy suggested a very small impulse in the opposite direction of each velocity. So I ended up with this slightly lame method to apply damping wich you can call per step:
This is something that most simulations would need, unless you want your simulation to run in void (space) without any friction like air has, so I'd say your damping codebit is usefull for all users.