So, even though you've included the more accurately calculated Hessian, the energy surface is such that the optimisation routine keeps optimising to that saddle-point. I think this can happen when your convergence criteria aren't strict enough - if the imaginary eigenvalue is small, the energy surface at your saddle point behaves quite similarly to that at a genuine local minimum.
The screwer (also known as 'vibrations') tool allows you to distort the geometry along any calculated vibrational mode, including the one corresponding to your imaginary eigenvalue. If you distort it along that one, it could move the geometry out of the 'trap' it's in on the energy surface and allow it to optimise to a proper minimum. The tool is fairly self-explanatory to use - just run it in the directory with the control file and select a mode and a 'temperature' used to determine how much distortion you want; then the distorted coordinates are put in a $newcoord key in the control file. (You then have to copy them into the actual $coord block before it'll use them, as far as I can tell. The documentation on the tool is virtually non-existent.)
Be aware, though, with groups like methyls, I think you can get hindered rotor problems - where the energy surface along the rotational motion isn't sufficiently deep to make the harmonic approximation valid.