Actors eliminate data races, not retain cycles. If your actor stores a Task whose body captures self, you just built a memory leak that won’t crash — it will quietly persist.

• Keep cancellation close to creation: only store a Task when you must call cancel() across scopes; otherwise make it local and let scope exit drop it.
• Prefer AsyncStream for repeating push sources; set continuation.onTermination and expose a stop() that finishes the stream and removes NotificationCenter observers.
• Use nonisolated only for methods that truly don’t touch actor state; bounce back into the actor with await self.method() when mutation is required.
• Bridge legacy callbacks with withCheckedThrowingContinuation and avoid capturing the actor; resume exactly once to prevent pinning tasks.
• Validate lifecycle: add deinit that cancels timers/loops (Task.sleep) and nils stored closures; assert deallocation in XCTest.

Choose a stored Task only when you require cross‑scope cancellation; otherwise keep work local to avoid owning lifetimes you can’t reliably terminate.

Trace with OSSignposter and Instruments, and monitor field behavior