2.28.0

Behavior change: an un-awaited select() loser is cancelled when the run
closes. To keep a service-mode activity running, schedule it with
cancellation_policy: abandon; its outcome is then never recorded. See the
first entry below. No migration.

• A run no longer leaves work running after its own code closes it (#1117).
  Before, an activity or timer the workflow never awaited, such as a
  select() loser, stayed open when the run completed, failed or continued as
  new. The activity still executed, and its ActivityStarted and outcome were
  appended to the closed run's history after the terminal event. Now the close
  records ActivityCancelled for each open activity and TimerCancelled for
  each pending timer in the same transaction, before WorkflowCompleted,
  WorkflowFailed or WorkflowContinuedAsNew, and cancels their tasks. Such a
  cancellation carries cancellation_reason: run_closed, and replay ignores
  it, because the workflow never observed it: it never changes a selection
  winner or a parallel group's failure. A
  late claim, heartbeat or outcome for such an activity is refused with
  run_closed, including one left behind by a run closed before upgrading.
  Nothing is appended to the closed run. Exception: a service-mode activity
  scheduled with cancellation_policy: abandon keeps running. The terminal
  event lists it under the new abandoned_activities payload key, and its
  heartbeats and outcome are accepted but not appended. Cancel, terminate,
  timeout and cooperative cancellation already settled open work this way.
  Children keep their parent-close policy and continue-as-new hand-over.
  Behavior change: if a workflow needs a loser's effect, await the loser
  before returning.

• Replaying a workflow that repeats the same message-stream send no longer
  lists every earlier matching row again for each send (#1014). The exactly
  matched candidates for a send identity are now loaded once per execution
  and extended with the rows that execution writes, so N identical sends in
  one replay cost one candidate query instead of N, and O(N) rows instead of
  O(N²). Matching is unchanged: exact field comparison, send order and the
  redrive chain.

• An expired open activity row of a closed run is now closed instead of being
  selected and skipped on every timeout sweep (#1117 follow-up). An activity
  that its run's close left running under the abandon policy is marked failed
  when its own schedule_to_close deadline passes without an outcome. Any
  other open row of a closed run, such as one left by a close before
  upgrading, is cancelled. Its task is cancelled and nothing is appended to
  the closed run's history. Before, such rows stayed open and could fill the
  sweep's batch limit.