an embedder shuts its runtime down before calling Hydrant::shutdown, so a shutdown that lands during startup cancels the blocking task that queues gone backfills, and run logged `failed to queue gone backfills err=task N was cancelled` as an error. spindle restarts hydrant every round, so it showed up in spindle's logs. db.run returns a Cancelled error for work its runtime cancelled now, and run lets that one go quietly, since the next start queues them anyway. checking stop instead wouldn't catch it, because the runtime goes down before stop is set.