A lower bound for the completed hard objective's gradient at the initial point.
Frozen L07: either a query is outside and has full scaled norm, or it and
an optimizer both lie in the radius-four ball, where convexity converts the
S5-C value gap into the exact 1/128 gradient bound.