Halofit algorithm in CLASS vs CAMB

[nnedelko]

I've noticed that the CLASS implementation of Halofit has a "sanity check" procedure for setting the xlogr1 and xlogr2 bounds before looking for sigma(R)=1 (lines 214-312 of https://github.com/lesgourg/class_public/blob/master/external/Halofit/halofit.c). Meanwhile, the CAMB implementation has xlogr1 and xlogr2 simply hard-coded to -2.0 and 3.5 respectively. If I do the same in CLASS, I sometimes get "integration didn't converge" errors even though the iteration limit in CLASS is 10000 vs CAMB's 3000. What's the reason for this difference in the algorithm?

(this is important to me because I'm interfacing with the DES Y3 3x2pt pipeline which requires P(k,z) going as far as z=4.0, and I found Halofit behaviour at large z to be rather erratic)