Can you point me to an authoritative source that explains what exactly is guaranteed under what assumptions and what is done on a best effort basis only?
The authoritative source is the Go source code. Read it, it's relatively easy to understand, and fairly obvious that the STW mark termination phase goroutine (like any other goroutine) will be preempted after forcePreemptNS (which is 10ms), and if that happens even though it hasn't finished its work, it transitions to a concurrent task, and resumes later.
If you're interested in the finer details, I'm sure the GC talk at golangUK later this month will discuss them all, and the slides and/or videos will eventually land here on Hacker News.
