Short answer

Your implementation is not the canonical normalization used in the Pacejka combined formula.
Because you compute G0 with the same Bxa that depends on k, the denominator shrinks when k (or RBX2) increases, so the ratio can increase — which is the behavior you observed.
In the canonical / intended normalization the denominator is evaluated at a fixed reference value (effectively k = 0 baseline), so increasing k (or RBX2) makes the numerator move toward 1 while the denominator stays fixed — therefore Gx decreases (i.e. combined correction reduces longitudinal capability).

Why that happens (one-paragraph)

You have Bxa = (RBX1 + RBX3*yi*yi) * cos(atan(RBX2*k)). Larger RBX2*k → cos(atan(...)) falls → Bxa falls.

Both numerator and denominator in your Gx = Num(Bxa*As)/Den(Bxa*Shxa) depend on Bxa. Because As (ai+Shxa) is usually much larger than Shxa, the numerator changes faster toward cos(0)=1 than the denominator does. Thus the numerator/denominator ratio can grow, even though the underlying Bxa fell. That’s an implementation artifact from using the same B in numerator and denominator.

What Pacejka expects (practical canonical normalization)

The usual approach is to normalize by a baseline evaluated at the reference shift (so the denominator is effectively fixed). In practice you evaluate the denominator using the baseline B at k=0 (i.e. cos(atan(RBX2*0)) == 1), or explicitly evaluate at the reference slip shift Sh but using the baseline B (not current Bxa).