Just one minor nitpick.  We're not testing the hardware RNG directly, we're testing the final output of /dev/random when it's being continuously fed entropy from the hardware RNG.  /dev/random has it's own whitening algorithms and doesn't JUST use the entropy provided from the hwrng, it pulls from a number of other places as well. 

I agree with you that using JUST the output from the hwrng is unsafe.  So that's why we feed the entropy into /dev/random and don't use it directly.  Our test is testing the system as a whole.

Your points are certainly valid though.