In the ITU G.993.2 document, that SNRGAP term only appears in the ATTNDR formula for use when the modem is in "loop diagnostic mode". So it's for when the modem is in a special diagnostic mode, not actually trained up. Normally the ATTNDR wouldn't need to be calculated that way.
With the basic ATTNDR method, the modem is free to choose the amount of FEC data, interleaving depth and other framing parameters for maximum bandwidth (within the maximum delay). Whereas for the actual line rate, it will have to set the amount of FEC data, interleaving depth and other framing parameters to achieve the required minimum INP value specified.
Last time this was discussed, I think I had difficultly explaining or convincing people about the "coding gain" aspect, that FEC+interleaving could be used to give a higher net data rate than without. I'll try explaining it again.
The modem connects at the best speed it can constrained by having to meet the error rate (and target SNRM, but SNRM is defined relative to the error rate).
So when you add error correction capabilities, that would lower the expected error rate, therefore allow a greater bandwidth to be achieved for the same error rate. That's the coding gain.
Of course the FEC data takes up some of the bandwidth, which has to be taken into account to give the net data rate. But it's possible to gain more than you lose.
It was easier to see on ADSL2, when switching on FEC+interleaving can give you a higher net data rate than without.