yes, i know the problems that arise with stacking and accurate collisions detection. i was considering deliberatly lowering accuracy in cases where contact with low relative velocities is found, and use an impulse based method as descibed in this forum the past week.

if youd want to do it exactly, i indeed think youd have to integratein steps, and change the impulse every time a contact breaks or forms on the vert-tri level. but this would get very very complex with for example a round shape resting and rotating in a concave bowl, and be very complex to code aswell.

however, i think stacking is doable with a hackish yet good method like impulse based and the right simplifications.

i also wonder what research has been done in this field. all papers i find are always geared towards games/not-too-realistic simulation. interest in physically accurate simulation or avoiding degenerate cases usually isnt too high. they expect the programmer to taylor and tweak the algorithm to his specific needs, and trust him to put in some hacks like limiting rotation speed.

anyway, i think speed is going to be fine, if the right global algorithm is used. my very quick and unoptimized sweep&prune can handle a dense cloud of 1000 objects in 3ms. after this youre left with 500 potential contacts at most, which will rapidly fall with each level of the test.