The keyword “ensure_bt_convergence” reduces the step size until one can be taken in cartesians. However, the first event in our IRC algorithm is a step in the direction of a Hessian eigenvector that is 1/2 the length of the requested IRC step size. So this keyword is probably not helping you get what you need.
On the other hand, if commenting out ‘fix_bend_axes’ allows convergence then I believe you are OK. This keyword fixes the meaning of quasi-linear bends. In some cases, it is needed to get a successful back-transformation but in others I’ve recently found it prevents a precisely chosen step. The latter is not generally critical for stationary point searches.
It’s a good idea to test with SCF if you can. Also, with printing turned up, the Hessian after its transformation to internal coordinates will be printed in the output. You could visually inspect that to see if it is reasonable. That is, do you see larger diagonal elements appropriate for stretches and bends.