The language you cite is also in the proposed documentation for data-race safety discussed in [1]. I take that documentation as authoritative and cumulative, so your feedback is relevant to that active discussion and doc PR and would likely be welcome there.
The use-case of running an actor method off the actor seems essential. I wonder if inference should/does permit @concurrent as a more-explicit override for classes and actors. However, it should probably balk at overriding conflicting declarations of protocol requirements or method overrides. Overriding those might break (the Liskov substitution principle for) subtype delegates.
So perhaps instead of
Only non-isolated functions can be @concurrent functions, and marking a function with @concurrent implies nonisolated
Is this closer to the intended reality?:
Marking a function with @concurrent implies
nonisolated. This may override the inference from otherwise unmarked actor and class functions, but not when overriding protocol requirements with actor isolation or when overriding class functions with explicit actor isolation.
[1] https://forums.swift.org/t/rfc-data-race-safety-chapter-in-the-tspl-language-reference/80381