The main problem is using the "Keep full IR" setting in combination with a very long sweep. Of your 174.6 second IR 174.0 seconds is noise floor, so not terribly efficient. Use the default "Truncate IR after 1.7 s" setting. It doesn't hurt to use a 4M sweep but it is just a huge waste of time in all but a very, very few specialist cases.
The arithmetic problem was caused by an internal protection to prevent the result getting too long and risking out of memory failures, that protection was kicking in when dividing 4M IRs. I have made a change to accept that case for the next build, but all should work OK using a sensible IR length.
Your target level gets very low at the lowest frequencies due to the high pass shape, that causes heavy boosting of LF noise when dividing. To deal with that use regularisation, even 0.1% will work.