The main thing is diminishing returns. There's a few categories:
Actual cap. Trait gets bigger, but effected mechanic is limited.
Sex Time
Fetal growth time
Crit Chance
conception chance
Knockback resist
Soft cap (numbers get bigger, but nothing in the game can actually differentiate)
Damage
Crit multiplier
Leaving only health that can, uncapped, deliver some marginal effect, but even that effect ultimately gets very small.
Development wise, what might be better overall is passing down averages plus or minus a fixed value/percent rather than sums. Spit balling pseudocode, something like round((.9+.2*rand())*($mom.trait+$dad.trait)/2) + randint(-1,1). That's an average, scaled between .9 and 1.1, rounded, then possibly increased or decreased by 1. That should keep breeding meaningful for much longer.