The delivery seam a fired timer job goes through - the run-liveness check st-ADR-0054 decision 4 requires, plus the feed-back itself.

Spec 6.2 says a delayed send whose session terminated before the delay elapsed MUST be discarded without delivery. Inside statifier-ex, Statifier.Session satisfies that on terminate by cancelling every live timer ref before the process exits; a durable scheduler survives process death by design, so nothing plays that role for a stored job. What replaces it (st-ADR-0054 decision 4): before feeding the fired event back, the host MUST establish the run is still live and discard otherwise. The guarantee is enforced here, at delivery time - a cancel-on-run-end hook may keep the store tidy but is never load-bearing, because the node death durability exists to survive takes the hook down with it.

Whether a run is live is the host's question, not this package's: a host running Statifier.Session processes answers it from the session registry and Statifier.Session.status/1 (StatifierOban.Timer.Delivery.Session, the default), while a process-less host driving Statifier.Interpreter against persisted positions answers it from whatever it stores as the run's terminated/halted state and feeds the event into its next drive. That is why this is a behaviour: StatifierOban.Config carries the implementation (:delivery), and the scheduled job carries it to the worker.

Implementations must be idempotent under redelivery: jobs in this package are at-least-once, so deliver/2 can run more than once for the same fired timer.

Restoring the caller's trace context

caller_context (st-ADR-0063) is the opaque host slot the sending macrostep stamped onto the effect. It rides the job args untouched (StatifierOban.Timer.JobArgs) and is on the %SendDelayed{} handed to deliver/2 days later, on another node, byte-identical to what was scheduled. An implementation MUST copy it onto the event it feeds back, unchanged. That last hop is where the value earns its keep: upstream puts the fed-back event's caller_context on the macrostep telemetry the firing drives ([:statifier, :session, :macrostep, :start | :stop]), which is what lets opentelemetry_statifier link the firing back to the trace that armed the timer instead of leaving it detached. A delivery that drops the slot loses the link at the last hop, and loses it silently.

fired_event/2 builds that event correctly, and host implementations should use it rather than assembling one by hand - it is the same event StatifierOban.Timer.Delivery.Session feeds back.

What a host may put in the slot is the host's business and none of this package's: nothing here reads it, matches on it, or keys anything on it (ADR-0006 decision 7). Two durability rules do bind the host's choice, and they are StatifierOban.Timer.JobArgs's to state.