Rather embarrassing that TI would publish that. Alas, such are appnotes...
If they had increased R_G to, say, 100k or more, and R_F proportionally, it would be about right.
Less gain is required if the RC values are staged. Performance is in fact very poor when the R and C are equal values. Instead, the R values should increase, and C decrease proportionally, along the chain. This keeps the poles closer together, increasing phase shift and gain in the transition band. (Of course, this needs an even larger R_G still, which is kind of annoying.)
Also, hahahaha, the first one uses a TLV2471 (single), and the second one uses a TLV2471 (single) with a part of a TLV2474 (quad). Nice copypasta work they did there.
Anyway, for adding up impedances, no, you cannot take the magnitude, you must work in complex numbers the whole way through. This results in a rather complicated 3rd order equation for the circuit, but that's just the nature of the beast. Plug it into a symbolic math solver and get your results that way.
Tim