The only thing we actually know with reasonable certainty is that his backup script is broken, and that this is bad. Randomly changing settings so that the brokenness is still there but just less obvious is more dangerous than helpful.
Its been suggested before but pg_dump could be a bit more helpful if it dumped contents to a temporary location first and then, if the dump completes successfully, moves it to the user-specified location. It would be a bit of a subtle error if the dump fails silently and the user doesn't notice that the timestamp of an already existing dump at the named location didn't change - but if the target location is already unoccupied then the absence of a file would be very obvious.
