I rerun the geometry optimization with settings that should make the NWChem job compatible with Gaussian (i.e., GRID nodisk lebedev 90 14 ssf euler in NWChem and ultrafine grid in Gaussian also for the CPHF).
Results:
The final stationary point is still a second order saddle point for NWChem.
Instead, Gaussian09 characterizes the same geometry as a minimum.