Roughly 1% transfer chance is achievable.
Suppose you have a gene that is useful if you get one copy, but fatal if you get two (think sickle-cell). Now if you breed with a normal person,
1/2 the offspring gets the trait and
1/2 is normal. If you breed with another mutant, again
1/2 of the offspring gets the trait (from either parent),
1/4 is normal and
1/4 is dead.
Now suppose you have
n such genes, in different chromosomes (independent, so the probabilities multiply), which are all required for the needed trait. Go out with normal person,
2^-n of the offspring are mutants and
2^-n normal, the rest are passive carriers of some of the required genes.
Go out with another mutant, the probability for the offspring to live is
3^n*2^-2n and the probability to get a mutant is
7 you get
(3/4)^n. So 7 in 8 children in a mutant couple are dead or disabled, no surprise they'll keep separated.
The hard part is making them 1 in 10M and not 1 in 128 though. Longevity and vigor are attractive. Maybe you should limit their fertile age somehow.
Fun is also likely to begin when longevity genetics are discovered and normal children of long-lived people start looking for a couple who has complement set of altered genes. The chances for longevity of the child are still 1 in 128 though.
Now if you tell that "longevity is acquired through getting gene1, gene2, gene3..." this is just boring. I would suggest thinking of actual traits, e.g.
Disclaimer here. I'm not a biologist, need one to refine the following.
1) Cancer resistance. 1x = no cancer, 2x = cell division is slowed down, no hair, no nails, deficient bone marrow leads to anemia and death.
2) Brain tissue regeneration. 1x = very nice, 2x = brain keeps growing, fills cranium, and dies from obstruction. One can show superior intelligence in the meantime.
3) Very potent immune system. 1x = excellent health, 2x = autoimmune disease of choice.
EDIT Maybe you should think of unpleasant but tolerable side-effects of the 7 genes that cancel out if one gets jackpot. This will explain 1/10M distribution. Or it won't, not sure here.