`registry::update` took the lock with the blocking `take`, justified by a doc comment claiming its callers were sync. Every one of them is reached from an `async fn`, so a contended registry write parked a tokio worker for up to thirty seconds — very often the thread that would run the peer's token refresh. `take` had no other caller and is gone, leaving `take_async` as the one way in. Change-Id: I3ebd16fb59e839ebb46bf22af1a453fe4198e141