Many thanks to @fgomez, I guess we finally could find the reason why this happens.
Here is the screenshot of the Material Explorer and I purposely include the longer wavelength response here:
As you can see, gold becomes very lossy at higher wavelength. For such a broad pulse, even if you set your wavelength to be from 1um to 10 um, injected pulse includes higher wavelengths component. On the other hand, the energy in your simulation region is calculated by summing over all the wavelength which directly effects the early termination (set by auto shutoff level). As a result the longer wavelength causes a fast decay on the energy and auto shutoff to trigger earlier. Thus your simulation does not run to completion (where all the frequency components decays completely) and you see transmissions larger than 1 at some wavelength.
Therefore, a careful attention to the injected pulse behavior is required. You can either change the pulse bandwidth or uncheck the optimize for short pulses from source Edit tab. Unchecking the optimize for short pulses will use a slightly longer pulse in time domain (narrower bandwidth) and will exclude the longer wavelength that causes the problem.
Please let me know if you still have questions.