[alexluong]

When the retry scheduler stops, Monitor returns as soon as its context is cancelled, while tasks it already received keep running on that cancelled context. This happens on every restart with retries in flight:

• A task cancelled before it publishes fails and backs off like any failed retry, so the retry is delayed and the receive counts toward the dead-letter limit.
• A task that published but stops before deleting its message is received again after the visibility timeout, and the retry is delivered twice.

Monitor now waits for running tasks before it returns:

1. Tasks run on a context that shutdown doesn't cancel (context.WithoutCancel, so logger and trace values carry over), so a received task finishes and deletes its message. This matches how internal/consumer shuts down.
2. The wait is bounded by a shutdown grace, 5 s by default (WithShutdownGrace). Tasks still running after it are cancelled and handled like any failed task, as today.

Follow-up from the review on #1162.

🤖 Generated with Claude Code