In real life, RNG attacks are against the implementation not the noise source, even something as "predictable" as "atmospheric noise" is random enough for all practical applications.
There are measurements you can make that could help you predict Johnson noise. For many quantum schemes, it's much tougher.
(Not that it is easy for an attacker to make those measurements in the first place, but it doesn't seem fair to say it is rigorously just as unpredictable)
Two resistors made of the exactly same material, the same cristaline structure down to every atom and heated in the same way up from 0K will generate different noise.
No, good hardware random number generators are based on nothing, because there is no existing problem with software random number generators that HW rngs can solve.