The :telemetry surface for this package's durable seams (ADR-0006) - the single definition site for every [:statifier_oban, ...] event name, and the one module that calls :telemetry.execute/3.

docs/telemetry.md is the full contract: what each event answers, what it deliberately leaves to Oban and to statifier-ex, and what opentelemetry_statifier does with it. This moduledoc is the reference table; that note is the reasoning.

events/0 returns every name below, built from the same literal-atom lists the emitters use, so the bridge can attach one handler per event name without hand-copying the list (ADR-0006 decision 4, ots-ADR-0003).

What these events are for

Two other surfaces already report on this package's work. Oban instruments the job ([:oban, :job, :start | :stop | :exception], [:oban, :engine, ...]), and statifier-ex instruments the effect ([:statifier, :session, :effect, :send_delayed | :cancel | :invoke]). What neither can see is the durable step between them, and that is the whole of what this module emits: whether the effect became a row and whether the insert was new (conflict?), which statechart identity an opaque job row belongs to, and the spec-level verdicts that are successes for Oban and non-events for the chart (the 6.2 discard, the 6.3/6.4.3 sweeps, a permanently failed invocation).

Duration, attempts, retries, snoozes, queue latency and exceptions are Oban's and are not re-emitted; in particular there is no lateness measurement here, because :queue_time on [:oban, :job, :stop] is the same subtraction against the same two timestamps.

Structural rules (ADR-0006 decisions 3, 4, 5, 8)

• The prefix is [:statifier_oban, ...], fixed and not configurable. Upstream reserves its own second segment (:session) for the logical SCXML session, and nothing here is scoped to one.
• Measurements are numbers; metadata is everything else, integer indexes included - ordinal, macrostep, microstep and round are metadata here, because an opaque index has no numeric meaning to average.
• Every event is a single point-in-time event. There are no :start/:stop pairs: this package owns no interval Oban does not already own, so every event measures system_time plus whatever numbers are genuinely its own.
• Emission is unconditional. There is no config knob and no sampling knob: :telemetry.execute/3 on an event with no handlers is a lookup and a return.
• Amendment discipline. Adding a measurement or a metadata key to an existing event is an amendment and is fine; renaming or removing one, or renaming an event, is breaking and needs a new ADR.

Scheduling seam

Emitted synchronously on the process that drove the macrostep, before any Oban job exists.

conflict? is what answers "did that replay do anything": true means the dedup key held and no second job was stored.

count on the two cancellation events is the sweep's own return, and 0 is data rather than an error - spec 6.3 cancels every timer under a send_id, and a cancel that matches nothing is a no-op.

reason on the two *_rejected events is the typed error the call returned, unchanged. scope is nil on [:statifier_oban, :invoke, :enqueue_rejected] when the rejection is {:invalid_scope, ctx} itself: there is no validated scope to report, and reporting the raw ctx would put unvalidated host state on the event.

Delivery seam

Emitted on the Oban worker process, inside the job.

reason on the two :discarded events is the delivery seam's StatifierOban.Timer.Delivery.discard_reason/0 - the spec 6.2 verdict as data, which Oban buries inside :result.

[:statifier_oban, :invoke, :failed] mirrors StatifierOban.Invoke.Delivery.deliver_failure/3 exactly, including ADR-0005's three-class reason vocabulary ("run_failed", "run_crashed", "undecodable") and its attempts semantics. Its handler is nil on the "undecodable" class alone: that row never reached handler resolution, because the decode that would have named the module is the thing that failed.

Where the seam delivers nothing, this emits nothing: the environment errors :invalid_handler, :invalid_delivery, :invalid_codec and :codec_failed say the deploy is wrong rather than anything about the invocation, and they ride Oban's exception event.

Fan-out seam

Emitted around ADR-0007's fan-out, and recorded by ADR-0006's 2026-09-06 amendment. None of the three is an answer: the fan-out job completes without delivering, a child start creates a run and delivers nothing, and the settlement side answers the invocation once on behalf of all N.

count is the fan-out's width on all three. It is a measurement on :fan_out and :unstarted_cancelled, where it is the number the call itself produced - the status it already has on the two :cancelled events, 0 included - and metadata on :child_started, where index and count together are the child's position rather than a quantity.

:unstarted_cancelled's count is half of sb-ADR-0009 decision 6's first_error cancel: the indices whose start job never ran. The siblings that already have a child run are the settlement side's to cancel and are not counted here.

handler is on :fan_out alone - ChildStartWorker never resolves the handler name to a module and cancel_unstarted/3 is given no handler at all, and a string where every other event carries a module would be worse than the absence.

Cardinality

Every metadata key above is bounded by the chart rather than by traffic, except job_id, which is a correlation id for a span or a log line and never a metric dimension, and caller_context, which is an opaque host term present for the bridge alone.

Nothing from the chart's datamodel is ever on an event. The effect's data, params and content are host-opaque terms (StatifierOban.OpaqueTerm), possibly encrypted through a host codec, and they are never emitted - not truncated, not hashed, not "just the keys".