I agree, and I personally wouldn't call golang memory safe for that reason. Thomas's semi-definition includes the word "vulnerability", which narrows the scope so much that golang fits under the bar, since the common data race that causes memory corruption hasn't been shown to be exploitable without being contrived.

My personal definition of memory safety for a language like golang would specify that you can't cause this sort of memory corruption without an explicit call to unsafe, but there's no real definition to fall back on.

The same thing happens any time a message board confronts a professional term of art. The same thing happened with "zero trust", where you'd have long wooly debates about what was meant by "trust", but really the term just meant not having a conventional perimeter network architecture. Sorry, but the term as used in industry, by the ISRG, and in government guidance refers specifically to vulnerabilities.