Ok, no one probably can help, so I did my own reasearch across various internet forums. It seems that this bug was quite common and is named "nvidia infinite loop bug" and happens from a certain version of Detonator drivers, probably 23.11 or so. Nvidia seemingly coded their drivers in a way that utilized some OS kernel stuff and that made some more unstable AGP implementations (like in this case) very sensitive.

Most people reported the problem on the Windows XP but it manifested on Windows 98 also. Via made some patch that set register 55 bits 5 to 7 to 0 but in Windows 98 you have to use wpcredit. I have checked by settings and those bits are already set to 0.

Tried various BIOS versions, various Nvidia drivers, multiple Via chipset drivers but nothing helped. Found a tip that in Windows XP you can manually override Via AGP driver with standard PCI-to-PCI driver from Microsoft and even though there is some performance hit (e.g. AGP texturing doesn't work) at least the bug doesn't pop up. Couldn't do it in Win98 as I couldn't manually select such driver. So I given up and changed the card, Geforce 4 MX and Via KT133 are not meant to be put together.

I think I had an mx440 in my jetway 663as-based Athlon thunderbird system back in the day. (Kt133 chipset, like yours) it was a very popular card and I’m sure the combination was/is widely used.

I use 4.43 via drivers and detonator 45.23 for the graphics on my kt133a system now, I have an mx440 but it is faulty. Bad caps I think. My two FX series cards work fine on this board.

Someone else with a kt133a board had issues with fast writes and/or sideband addressing. You may want to see if you can disable one or both of these features in turn to see if your stability improves

I have a ga-7zm board too. A real dog to work with from what I can remember. Relegated to storage a long time ago. Gigabyte were really good with super socket 7 and 440bx but after that they were horrid until sandybridge I reckon