I have actualy proved it to myself.
Lets break 3 quaternion characteristics to 3 base axises and 3 angular speeds scalars.
This composes a single rotation matrix against a base matrix (the identity rotation matrix I).
We compose a rotation matrix H by s degress around y.
We now perform I*H=D ; D*H=C ; C*H=E ;......
thus we are rotating by s degrees in a time step (second), rotating by quaternion that costructed H matrix, - the quaternion (0,1,0,s);
We now have to extend rotating by this quaternion along with one other quternion , quaternion (1,0,0,2s).
We construct H matrix from previous quaternion (0,1,0,s) and construct G matrix by the second quaternion (1,0,0,2s).
We now perform I*H*G=D ; D*H*G=C ; C*H*G=E ;..... . and we can reduce to J=H*G thus i*J=D ; D*J=C ; C*J=E ; ....
This means that whatever amount of descibing queternions adds up to describe tendence, they cen be incomposed to a single rotation matrix that transforms after itself the object timestep by timestep. A strange J/s unit.
But this yields problem of order of composited quaternion rotations. But I gess this is the responsibility of expressing last added quaternion corectly, rather than physical dilema.