Hi Evgeniy,
the what you call 'next' geometry can not or should not be much different to the geometry before, since if the convergence criteria are met, the gradient norm and the maximum displacement are small. Changes can thus happen only within the accuracy of the given convergence criteria.
I think the main reason why an additional gradient step is not done, is because this data is not needed by (almost) any subsequent calculation or post-processing tool. So it simply saves time.
The question which data is most reasonable to be stored in which file or data group is probably not a trivial one, though. What TM does is to make sure that coordinates, orbitals and energy are consistent. Those are needed for all subsequent jobs, and since in TM every job is a restart job (so to say), it is important to always have the right starting point for the next step. The gradients, on the other hand, are only needed for a few other properties, e.g. for frequency calculations when the gradients are not exactly zero. That's why the coordinates are stored in addition to the gradients for each step of an optimization in the gradient file - frequency jobs will use (for the terms which include the gradients) the correct matching coordinates.
But - as usual - you can easily adapt Turbomole such that it does what you want. jobex is a script which can be modified such that the convergence is checked after the gradient step and not after the energy step (search for the appearance of checkconv).
Regards,
Uwe