Unfortunately it is not really possible to find the true net amplitude of an analysis frame other than by resynthesising it. Bins are not truly independent (may reinforce/cancel mutually), and in principle the contribution of the analysis and resynthesis windows should also be taken into account. Any computation to find the level would probably take longer than the already efficient resynthesis.
One solution is indeed to pre-normalise the input, best to a level below digital peak, e.g. -3dB or -6dB, process, and hope. The alternative is to resynthesise to a floating point sound file (on the command line, prepend -f to the outfile name), which will preserve over-range samples without clipping, and then rescale as required.