It's not entirely clear to me that this is what the FDA is demanding. Quantifying the lab error rate is easy, and while it's possible that 23andMe is failing to do so, it's almost like not doing that would take effort.
For example, 23andme ships you a test package, you sample it yourself, send it back and then they process it. While quantifying that last rate is simple (as you say), it's crucial to understand how well the product will work in practice, with an actual customer on the other end.
For something like 23andme, that would likely mean getting a whole bunch of untrained test subjects, having them preform the sample collection as instructed by the instructions shipped with the sample collector, having them shipped over by standard means to the lab, and then processed. The test subjects would likely then have to have their DNA fully sequenced with some gold standard test, and then the results compared. 23andMe would be given leeway, in the sense that inconclusives don't "really" count as a wrong result.
The study that the FDA would have required from 23andMe is to show that the genetic testing provides the correct results at the bare minimal. They would require 23andMe to quantify their actual error rate of their product in use by customers. Depending on 23andMe's claims in their 510(k) submission regarding their intended uses, then the burden of proof may go higher.
The point is that if 23andMe can't actually prove that this test result predicts elevated risk of Alzheimer's (either by their own study, or through other relevant studies) then they sure as hell shouldn't say that their test predicts Alzheimer's risk.
Case in point, clopidogrel response. The link between CYP2C19 and clopidogrel response is well studied. There are already other tests on the market based on this relationship. Given this, 23andMe would not have to prove by themselves that clopidogrel and CYP2C19 are linked, they would merely have to prove their their CYP2C19 test provided correct results.