Fwiw one of the main features of FGX seems to be soft clipping specifically so you can raise the overall program level, so I would warrant that reconstruction meters aren't going to tell the whole truth.
The MP3 format uses 32bit FLT values to recall not pcm data, but 'frames' from a lookup table that are 1. of a given size (specified by the encoder for CBR and via the preset values for VBR/Lame) and 2. vary in length. So what happens upon converstion from an input format is not 100% predictable, and I have no idea how saturated peaks are represented if they are hitting full scale.
I would bounce to 16bit INT with the usual dither that you'd use, and bring that into an mp3 encoder externally. I've had more reliable results this way, as you can check the 16bit (or 24bit) waveform before and after mp3 conversion. If nothing else this will at least tell you what is encoded into the PCM output stream and what the mp3 encoder is reacting to.