I use the parameter

TwoPunctures::TP_epsilon                = 1e-6

to ensure that the NaN never appears, no matter where I place the BHs.  This replaces a denominator

sqrt (pow (r2_plus, 2))

with

sqrt (pow (r2_plus, 2) + pow (TP_epsilon, 4))

See

https://bitbucket.org/einsteintoolkit/einsteininitialdata/src/master/TwoPunctures/src/Equations.c#lines-91

so adds a fixed source of error to your evolution, but the effect of this is negligible.