My lower bound is that mutational load contributes 10% of the variance in IQ. I call that small. Independently, I propose that there should be room for 50 standard deviations in improvement. Although it's not clear what more would even mean. Surely linearity would break down. What I mean by the possibility of "50 standard deviations" is 20 disjoint sets of changes, each of which would accomplish 2.5 standard deviations.

If the typical gene is deleterious and contributes 1/N of a standard deviation, then there is room for N standard deviations of improvement above the mean. Of course there is a mixture of genes of different effect sizes. I expect genes of both effect sizes 1/10 and 1/100. Say, half of each. That gives room for 55 standard deviations of improvement.

If variation came from positive genes, an additive model would suggest much more room for improvement, but such genes would be much less likely to combine well than correcting mutations to the wild type.