What does AFMA have to do with it? It has to do with controlling variables in a test environment to ensure the results you are receiving. And despite your protestations, there is an element of accuracy from AFMA related to the precision. We have no way of knowing which of the sample 10 photos is most representative of the most optimum focus his particular camera and lens combination is capable of producing. I can also provide you with test results .pdfs where a non or mis-calibrated lens performed more poorly on the focus consistency tests than it did once properly calibrated (AFMA). That is contrary to the idea that accuracy is unconnected to precision in this application.
I don't know the algorithms involved, so I cannot hypothesize to any degree of accuracy why that might happen. But it does...often enough to convince me there is something to it.
And I have to wonder if you read my post completely. I mentioned my experience in your vaunted "real world" conditions that not only refutes, at least in the 35mm's case, any sort of worrisome AF consistency issues, but that also corroborates what I saw in much more controlled test conditions shooting those pesky test charts.
What you're calling "real world scenarios" is what we call in my line of work "operational testing". Beware that all "real world scenarios" or "operational testing" is not equal for purposes of drawing conclusions from them. Test design and controls have a lot to do with the fidelity of data produced. And I'm saying given the paucity of information regarding test controls on TDP's "real world" shooting, we cannot faithfully put much stock in an idea that this 50mm indeed has any AF issues worth worrying about.
I'm not trying to defend Sigma, per se. I'll dump the Sigmas in a heart beat if the evidence gives me reason to doubt their performance. But I'm similarly not going to simply buy some anecdotal "evidence" as proof without ensuring the evidence is properly controlled to produce the results noted. In this case, the most glaring control missing is sample size. I believe TDP did some fairly exhaustive shooting, probably using fairly common tools, including tripod when necessary. But the results are based on a sample size of 1 lens, which is not adequate to properly draw conclusions from. It's not good enough to dress up a review and AF performance test with accounts of how much and how varied the conditions were that produced the results, even if I grant you 2 test configurations based on his mention of testing both 1DX and 5DMkIII bodies with the lens. Not for me, anyway. I need more data.
AFMA improves accuracy but not precision, and I'm pretty sure that TDP knows to AFMA lenses before evaluating them for their reviews. If you don't trust this review, you might as well not trust ANY of TDP's reviews unless you know that the review process that they used for this lens is different than any other. And the subject distance for that series of infocus/out-of-focus shots is constant. If the degree of OOF were constant, then the dock with the multiple distance adjustments might have been of use, but they weren't. My point is this: TDP finds that the lens that Sigma sent them (not bought retail) can have balky AF. What is more likely -- that TDP's copy has a unique software build that is different from all the others or that all the ones being evaluated (pre-retail) all have the same software. Yes, it was 1 copy but it was one that Sigma chose. More reviews/user experience will be available over time, but that TDP found a difference in S35 and S50 AF performance is significant.
And yes, others in this forum have had issues with S35's AF. Eldar has had to change the AFMA amount over time. Why? Perhaps Sigma had to use a different lens ID for the S50 than the S35 and Canon's AF algorithms treat them differently... or not. Perhaps you assumed that all Sigma's new lenses will AF like the S35s you evaluated, but perhaps that is a bad assumption.
I have worked with software/hardware testing, and you can call it operational testing or whatever you want. We "test" every line of code, but there are still issues that are found once the hardware is in the field. But the problems identified in the field are just as valid as those found during the test phase in the lab. The field problems are often hard to duplicate because the environment/exact state under which the error occurred is unknown. Not everything can be duplicated in a controlled environment.