Apart from the exception flow, LGTM - I can't see a problem. Though seeing the complexity here - for code which isn't supposed to be perf-critical - I tend to agree with @NinoFloris's offline comment that we could sacrifice extreme perf here and go with a locking approach in order to have simpler/safer code. If needed I think I'd also be OK with just keeping the thread around (i.e. removing the timeout), which would obviously simplify this a lot.

But am also OK with keeping this lock-free code if we really want to (after the exception flow is fixed).

@NinoFloris can we have your pair of eyes on this too?

Thanks for simplifying - this looks great and far less magical/tricky :)

Looks safe to me - see minor comments. @NinoFloris can you give this a review too please?

Modulo the feedback by @roji LGTM!