but I do not go by those rules.
this is how I would resolve the kind of problem you describe.
imagine the model is standing on it two feet, and all the stability problems are resolved.
let us call this state idle.
In this state, the center of mass can be anywhere inside the support polygon an dthe model should not fall. again assuming the support polygon is stable.
Now let us say it is going to transition to stand on one foot.
if you take the step the way you said, yes will fall, but that's not how will work.
to make that transition the state machine makes a transition to an intermediate state.
The goal of this state if to displace the center of mass until it is in both support polygons,
the two feet support and the target foot support polygon.
as soon as that condition is met, the state signal itself to transition to the stance on one foot state.
if what have now idle and the stance working (which we see now although with noise)
the only part that is an unknown is the transition from state idle to one foot stance.
but this state is just the union of two states that are already working.
this has nothing to do with solver, the only thing the solver has to guarantee is that predicted position are reproduced with some degree of fidelity. and that I am confident is happing.
when the stance state start to swing the leg on any direction, the balance system will try to shift the center of mass atomically to try to keep it inside the support polygon.
how fast the reaction can be is something to be seen, but it is most certainly work within some limits we just need to find out what that limit is.
This is not different than a real person doing the same thing and losing the balance.
where is my beer?