v0.4.1

Patch Changes

• 4d3efd3: Fix silent terminal-state loss on the observability final flush — a job_executions
  row could stay running with duration_ms: null after its invocation already read
  completed. The final flush deleted the per-invocation buffer BEFORE awaiting the sink and
  swallowed any error, so a partial batch (the invocations mutation lands, then the
  events/jobs mutation throws a deterministic GraphQL-level error, which is never retried)
  permanently and silently lost the job's terminal state.

  Three defenses:

  • flush() no longer discards state before it's durable — the buffer is deleted only
    AFTER the sink resolves. A failed final flush now RETAINS the buffer, so onFlush (which
    runs in the same handle() finally, right after onInvocationEnd) retries the terminal
    write instead of dropping it.
  • Sink failures are surfaced, not swallowed — new onSinkError?(err, { invocationId, final }) option on observability(...), defaulting to a [eventkit] console.warn.
    Ignored under strict (the error is rethrown as before). Silent telemetry loss was the
    worst failure mode for a telemetry system.
  • graphqlSink graceful-degrades an events/jobs GraphQL error — a deterministic
    GraphqlResponseError on the job_executions/event_executions mutation retries ONCE
    with the droppable payload columns shed (result, job_options, error stacks), mirroring
    the existing source_job_id degrade, so the row lands lossy-but-terminal (status
    preserved) rather than being lost entirely.

  Additive and backward-compatible: onSinkError defaults to the console warning, and the
  sink degrade only triggers on a mutation that would otherwise have thrown.