Currently we can compute the proof only from the initial witness from which gnark solver computes the full solution. But for fuzzing/testing purposes it would be good if we could compute the proof directly from the solution, even if the proof would be incorrect.
See also #1474 and #1513 (reply in thread)