Alexey A. Popov
Probably there is some misunderstanding. Using geometry optimized with the smaller basis set for computations with extended basis sets is a standard reasonable way. The problem with your computation was that somehow the code has found a gradient from the previous in your input file (maybe you just copied it accidentally?) - at least, it seems so after checking your output. For this reason, it did not compute the gradient again, it just took the gradient from the previous run - and since in the previous run you really optimized the structure, gradient was "zero".