I have seen watchdogs implemented in every possible way (main loop, all over, timer interrupt and more). The way to implement a watchdog safely is to kick the puppy in the main loop as many times as needed and nowhere else.
Now the whines:
But that forces me to keep my subroutines/functions small so that the processing time is within the watchdog timeout -- is'nt that great.
But if an interrupt comes in it will extend the time my subroutine runs and possible cause a timeout -- that is part of the reason I tell you to keep interrupt routines short.

In other words, proper use of a watchdog requires proper code layout, you can not save crappy code by adding a watchdog.
The purpose of a watchdog is not to save your crummy code, but to rescue if a particle or whatever makes the uC glitch.

Happy hollidays,

Erik

PS having a watchdog gives one added bonus: you have the greatest "software reset" there is: let the dog time out.