Now the final solution, no one seems to have read the sources 
The mux function table of the 73HC138 was correct - just my assignment of the MUX1-3 pins in the switch statement was plain wrong. I attached the cables on the atmenga2560 wrong first,so this code worked. Now with correctly sorting A0-2, the code has to be corrected as well. Everything works again. 
Attached is the correct and little bit cleaned up code which works as well on ATmega328 as on ATmega2560.
Problem turned out to be real, also on the ATmega2560 chip 4 is detected as defect. Have to rebuild the ram bank properly. Sorry for bothering you with that.
Weird problem: All works great on the ATmega2560.
Since my first Nanos arrived, I wanted to shrink down everything.
First I used pins A0-A2 for interfacing the multiplex 74HC138.
It seems everything is working right, but memory chip 4 is checked as being defect. Looking at the returned buffers, it spits out only random 0 or 255. Unchanged with DIV128 on SPI (~250kHz).
I did electrical checking of all pins and their connection to the bus interfaces, after resoldering every pin just to be sure. I also switched to pin D2-D4 for the multiplex, still the same result, chip 4: Defect.
So I switched to Arduino 1.5.6r2 from 1.0.5 and edited platform.txt to remove the compiler optimizations. Still no changes.
Then I removed the second buffer in my sources and read back into the write buffer, comparing the values directly with the loop counter. I only check one module at a time (need to change the chip value and recompile/reupload the sketch) and not all in a row.
Now it works, every single chip shows up ok. Going back to analog pins A0-A2 shows module 4 as broken again. On D2 to D4 everything as expected. Even when I switch on compiler optimizations again. Still, when running through a loop of checking all 5 chips, module 4 is shown defect.
Can anybody suggest what is going on there, next to my bad coding style? I fail to see where a stack overun / memory overflow of the SRAM could happen in this simple code. 2k SRAM should be enough for these few variables and the 256 bytes buffer.
Edit: Ok, weird. Removing the check-calls from setup() and looping through the checks once in the main loop solves this issue. Find a Nano version of the ram_test attached.
Edit2: weird, after trying A0-A2 again, this only showed chip 5 ok; back to D2-D4, it is chip 4 again which again gets reported defect. I'm lost.
SpiRAMBank-lib.zip (6.3 KB)
SpiRAM_Test.zip (2.43 KB)