Whoever is interested, some more reading on the same issue, encountered by other people on Armel platform:

http://permalink.gmane.org/gmane.lin...arm.omap/52674

same issue, fixed by dissect between 2 kernel trees to find the troubled sourcecode in ubuntu on beagleboard :
https://bugs.launchpad.net/ubuntu/+s...ux/+bug/591941

issue still stays why some kernels work. Basically the 2.6.28 kernels (stock + derived kernel-power line) all work. Are they maybe configured with PREEMPT_NONE=y ?

That could explain the difference in behavior with the 2.6.37 Nemo kernel line.

Minute timing differences in the boot/init process can already influence this race condition, making it hard to pinpoint the issue.

Reading back at the Mer/Nemo install wiki, it almost looks like this issue was known/encountered before but was never really bottomed out, maybe because other more important issue were present.at the time.

Just my 2 cents, to be confirmed.