It's worth noting that the number of actual failures is measurable only after millions had shipped. 10-100 dppm is a reasonable guess (~100 phones out of 2.7M shipped) with relatively low "infant mortality" until they were compressed into a phone and used (and abused as normal).

As a tester, you build a 10k ($10M) unit test run and none of them fail, and then another 10k ($10M) run and again none of them fail... statistically you might not expect reliably catch a single failure until you run 10 times.... that's a $100M test you need to get your boss to sign off on.

Note that when you get a failure it will burn up and you won't be able to identify the cause.

Now, if you knew the failure mode, you could test for it directly and you wouldn't have to waste an entire assembled phone in a destructive evaluation, but that's hind sight. The failure defects they see seem like pretty standard problems, but other than sampling, I'm still not sure how they are going to catch them in manufacturing.

What's amazing is that more batteries don't go up (the original Prius was notorious).