OK, new theory:

In 3.5.3 watchdogs were miscdevices and accessible trough filesystem. When filesystem got re/u/mounted drivers were getting release file operation called, and kicked the watchdogs, just before init - leaving enough time for userspace to take over.

Now watchdog drivers are in their own kernel class, and only first registered watchdog is linked to /dev/watchdog miscdevice for legacy reasons. So, in our case with 2 watchdogs, only omap_wdt gets kicked and twl4030_wdt doesn't. That doesn't leave enough time for userspace to start DSME that does the kicking in Mer based distros.

Maybe krenel should kick all watchdogs one last time before calling init. I'll test that tomorrow.