Well the fact it's reporting as 48MB instead of the real 32MB will be the reason it's crashing. What's possibly happening is that the memory is "ghosting" and appearing twice.

This could be for two reasons I can think of:
1) The new memory is being addressed twice. Standard sequential memory tests mightn't pick this up.
2) Somehow capacitance is causing phantom memory.

Mark's very basic memory check should pick up either but it's possible it's not.

I doubt a memory chip since for it to pass the memory test, all four chips would have to behave the same.

Is it possible there is a short on the CPU between the two RAS pins? That would make the new 16MB appear at both locations eventhough it's the same physical memory. It would also pass the memory test I think too. I'm not sure if Mark is writing memory addresses into locations to check for this sort of thing.

After writing all this, I'm tipping a short on the RAS pins at the CPU.