That means digging into the packaging. However, I don't have any power tools small enough to do this with any accuracy

The other damaged pins just fall off one after the other

Decided to take it off. Goes into storage until I can get into the package, or it might become a decap part.

The ones near the TMU and the ones on top of the 2D ASIC seem fine.

I just don't have any 35ns EDO

It is not the RAM, it's this NAND

I take that back. It is the DRAM. But which? I cannot bulk replace because this DRAM is obsolete and eBay has outrageous prices. 35ns is also too fast to just use sacrificial EDO sticks as a source since these are 80 or 70ns. 60ns top. I cannot test them with a small C program writing into the memory and reading back because portions are write only. And there are no THT SOJ40 sockets to be found. I could use the SMD sockets and then either make my own EDO sticks to use memtest86 on or I could use a blue pill stm32 board to build a crude tester

This works too. Testing in circuit would save me from having to desolder and resolder 3x 12 SOJ40 ICs.

So to summarize:
Card 1, the orange one, had bad glitches even in 2D. I fixed some loose pins on the 2D ASIC and now it's dead as a doornail. Just completely hangs on post with not so much as the BIOS complaining about it. Wish I knew where my POST card is.

Card 2 works in Windows, until you start a 3D accelerated application where it shows texture glitches until it hard crashes the system. That can be triggered by blasting the middle DRAM with hot air.

Card 3 has an FRJ texelFX with broken off pins, which could be rescued with fine enough power tools (think dentist drill). Works fine in 2D.