ruststream/runtime/

dispatch.rs

//! The per-subscriber dispatch loop: pulls messages off one subscriber and invokes its handler
//! until shutdown is signalled or the stream ends. [`RustStream`](super::RustStream) owns the
//! task spawning.

use std::fmt;
use std::future::{Future, poll_fn};
use std::hash::{DefaultHasher, Hash, Hasher};
use std::num::NonZeroUsize;
use std::panic::AssertUnwindSafe;
use std::pin::Pin;
use std::sync::Arc;
use std::task::{Context as TaskContext, Poll};
use std::time::Duration;

use bytes::{Bytes, BytesMut};
use futures::{FutureExt, Stream};
#[cfg(test)]
use tokio::runtime::Handle;
use tokio::task::{JoinError, JoinHandle, JoinSet};
use tokio_util::task::TaskTracker;
use tracing::{debug, error, warn};

use crate::{
    AckError, BatchSubscriber, HeaderMap, IncomingMessage, OutgoingMessage, RetryDeclaration, Str,
    Subscriber,
};

use super::batch::{BatchHandler, TakenBatch};
use super::context::{Context, FromDelivery};
use super::failure::{DispatchFailure, FailurePolicy, panic_reason};
use super::handler::{Handler, HandlerResult};
use super::main_runtime::MainRuntime;
#[cfg(all(test, feature = "poll-diagnostics"))]
use super::poll_diagnostics::PollDiagnostics;
#[cfg(feature = "poll-diagnostics")]
use super::poll_diagnostics::PollSampler;
use super::publish::PublishContext;
#[cfg(all(test, feature = "memory"))]
use super::redelivery::ErasedRetryPublisher;
use super::redelivery::{DeferredRetry, ScopeDelivery};
use super::shutdown::Shutdown;
#[cfg(feature = "testing")]
use crate::testing::coordinator::{
    Coordinator, Delivered, HarnessScope, Record, TestHooks, in_harness_scope,
};
pub(crate) use threads::StartThreadError;
use threads::Threads;

/// Header carrying the framework's own retry count.
///
/// The runtime increments it on every copy of a delivery it publishes, and reads it back where the
/// transport counts nothing of its own
/// ([`IncomingMessage::redelivery_count`](crate::IncomingMessage::redelivery_count)): that is what
/// a registration's [`max_attempts`](super::RouterWith::max_attempts) cap counts there, on the
/// runtime's copy path and on a delay a broker crate honours by publishing a copy of its own
/// alike. A handler can read it too, to tell a first delivery from a redelivery.
///
/// # Examples
///
/// ```
/// # #[cfg(all(feature = "macros", feature = "json"))]
/// # mod demo {
/// use ruststream::prelude::*;
/// use ruststream::runtime::RETRY_COUNT_HEADER;
/// use serde::Deserialize;
///
/// #[derive(Deserialize)]
/// struct Order {
///     id: u64,
/// }
///
/// #[subscriber("orders")]
/// async fn reconcile(order: &Order, ctx: &mut Context<'_>) -> HandlerOutcome {
///     // The copies the runtime publishes count themselves; the first delivery carries no header.
///     let attempt: u64 = ctx
///         .headers()
///         .get_str(RETRY_COUNT_HEADER)
///         .and_then(|v| v.parse().ok())
///         .unwrap_or(0);
///     if attempt >= 3 {
///         tracing::warn!(order.id, attempt, "giving up");
///         return HandlerOutcome::drop();
///     }
///     HandlerOutcome::retry()
/// }
/// # }
/// # fn main() {}
/// ```
pub const RETRY_COUNT_HEADER: &str = "x-ruststream-retry-count";

/// Parses the current [`RETRY_COUNT_HEADER`] value, defaulting to zero when absent or malformed.
fn current_retry_count(headers: &HeaderMap) -> u64 {
    headers
        .get_str(RETRY_COUNT_HEADER)
        .and_then(|v| v.parse().ok())
        .unwrap_or(0)
}

/// Concurrency policy for one subscriber's dispatch loop, declared with the `workers(..)` or
/// `threads(..)` macro argument (or [`Workers::sequential`] by default).
///
/// - `workers(n)`: `n` long-lived workers, tasks of the runtime the app runs on, process up to
///   `n` deliveries of the subscriber concurrently; a free worker takes the next one.
///   Back-pressure holds: the stream is not polled while `n` deliveries are in flight. Global
///   processing order is lost by design.
/// - `workers(n, by_key)`: the same `n` workers as sequential lanes; a delivery goes to the lane
///   picked by hashing its [`partition_key`](crate::IncomingMessage::partition_key), so per-key
///   ordering is preserved. Messages without a key rotate over the lanes.
/// - `threads(n)`: `n` dedicated threads of the subscription's own, each with a current-thread
///   runtime, for a handler that computes. The subscription's loop stays on the app's runtime and
///   spreads deliveries over the threads round-robin; decoding, the handler and settling run on
///   the thread a delivery was handed to, and only there. The app's runtime stays free for I/O,
///   timers and the other subscriptions.
/// - `threads(n, by_key)`: the same threads as sequential lanes by partition key.
///
/// The workers or threads and the queues that feed them are made when the subscription starts:
/// handing a delivery to one allocates nothing.
///
/// The default is sequential dispatch (`workers(1)`).
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct Workers {
    count: usize,
    by_key: bool,
    placement: Placement,
}

/// Where a subscription's concurrent deliveries run.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub(crate) enum Placement {
    /// As tasks of the runtime the app runs on.
    Runtime,
    /// On dedicated threads of the subscription's own.
    Threads,
}

impl Workers {
    /// Sequential dispatch: one delivery at a time, in stream order. The default.
    #[must_use]
    pub const fn sequential() -> Self {
        Self {
            count: 1,
            by_key: false,
            placement: Placement::Runtime,
        }
    }

    /// A pool of up to `count` concurrent deliveries. One behaves like sequential dispatch.
    #[must_use]
    pub const fn pool(count: NonZeroUsize) -> Self {
        Self {
            count: count.get(),
            by_key: false,
            placement: Placement::Runtime,
        }
    }

    /// `count` sequential lanes keyed by the message
    /// [`partition_key`](crate::IncomingMessage::partition_key): per-key ordering is preserved.
    /// One lane behaves like sequential dispatch.
    #[must_use]
    pub const fn keyed(count: NonZeroUsize) -> Self {
        Self {
            count: count.get(),
            by_key: true,
            placement: Placement::Runtime,
        }
    }

    /// `count` dedicated threads for a CPU-bound handler: the subscription's deliveries are
    /// handled on threads of its own, each running a current-thread runtime, so the handler's
    /// computation holds up none of the app runtime's threads. The router-chain spelling of
    /// `threads(n)`.
    ///
    /// # Examples
    ///
    /// ```
    /// # #[cfg(all(feature = "macros", feature = "memory", feature = "json"))]
    /// # mod demo {
    /// use ruststream::memory::prelude::*;
    /// use serde::Deserialize;
    ///
    /// #[derive(Deserialize)]
    /// struct Image {
    ///     id: u64,
    /// }
    ///
    /// #[subscriber("images.resize")]
    /// async fn resize(image: &Image) -> HandlerOutcome {
    ///     tracing::info!(image.id, "resized");
    ///     HandlerOutcome::ack()
    /// }
    ///
    /// fn images() -> Router<MemoryBroker, impl RouterDef<MemoryBroker>> {
    ///     // Resizing computes, so it runs on eight threads of its own.
    ///     Router::new()
    ///         .include(resize)
    ///         .workers(Workers::threads(nonzero!(8)))
    /// }
    ///
    /// fn app() -> RustStream {
    ///     RustStream::new(AppInfo::new("images", "0.1.0")).with_broker(MemoryBroker::new(), |b| {
    ///         b.include_router(images());
    ///     })
    /// }
    /// # }
    /// # fn main() {}
    /// ```
    #[must_use]
    pub const fn threads(count: NonZeroUsize) -> Self {
        Self {
            count: count.get(),
            by_key: false,
            placement: Placement::Threads,
        }
    }

    /// `count` dedicated threads as sequential lanes keyed by the message
    /// [`partition_key`](crate::IncomingMessage::partition_key): a key always lands on the same
    /// thread, so per-key ordering is preserved. The router-chain spelling of
    /// `threads(n, by_key)`.
    ///
    /// # Examples
    ///
    /// ```
    /// # #[cfg(all(feature = "macros", feature = "memory", feature = "json"))]
    /// # mod demo {
    /// use ruststream::memory::prelude::*;
    /// use serde::Deserialize;
    ///
    /// #[derive(Deserialize)]
    /// struct Image {
    ///     id: u64,
    /// }
    ///
    /// #[subscriber("images.resize")]
    /// async fn resize(image: &Image) -> HandlerOutcome {
    ///     tracing::info!(image.id, "resized");
    ///     HandlerOutcome::ack()
    /// }
    ///
    /// fn images() -> Router<MemoryBroker, impl RouterDef<MemoryBroker>> {
    ///     // One tenant's images stay in order: a key always lands on the same thread.
    ///     Router::new()
    ///         .include(resize)
    ///         .workers(Workers::threads_keyed(nonzero!(4)))
    /// }
    ///
    /// fn app() -> RustStream {
    ///     RustStream::new(AppInfo::new("images", "0.1.0")).with_broker(MemoryBroker::new(), |b| {
    ///         b.include_router(images());
    ///     })
    /// }
    /// # }
    /// # fn main() {}
    /// ```
    #[must_use]
    pub const fn threads_keyed(count: NonZeroUsize) -> Self {
        Self {
            count: count.get(),
            by_key: true,
            placement: Placement::Threads,
        }
    }

    /// One worker on the app's runtime is indistinguishable from the sequential loop; one
    /// dedicated thread is still a thread of its own.
    pub(crate) const fn is_sequential(&self) -> bool {
        self.count <= 1 && matches!(self.placement, Placement::Runtime)
    }

    /// The same concurrency as tasks of the runtime the loop runs on: what a `threads(n)`
    /// subscription becomes under the test harness, which runs every subscription on the test's
    /// own runtime.
    #[cfg(feature = "testing")]
    pub(crate) const fn on_runtime(self) -> Self {
        Self {
            placement: Placement::Runtime,
            ..self
        }
    }
}

impl Default for Workers {
    fn default() -> Self {
        Self::sequential()
    }
}

/// Per-subscription publish context threaded into every delivery's [`Context`]: the
/// broker-agnostic `retry_after` fallback of this registration and the app-wide tracker for
/// post-settle continuations.
/// An `and_after` continuation is spawned onto `tasks` so a graceful shutdown drains it.
pub(crate) struct Delivery<C = ()> {
    /// The retry path of this subscription: the publisher a copy of a delivery leaves through,
    /// with the destination the mount site or the descriptor named. `None` where the broker moves
    /// the delivery itself, in which case a `NackAfter` on a transport with no native delayed
    /// redelivery has nothing left to do but requeue.
    pub(crate) retry: Option<DeferredRetry<C>>,
    /// What the mount site declared with `max_attempts(..)` and `dead_letter(..)`. Read only
    /// where the runtime is the one moving the delivery.
    pub(crate) declaration: RetryDeclaration,
    /// Per-scope task tracker for post-settle `and_after` continuations. The
    /// dispatcher spawns each element's continuation onto it after settling, so a graceful
    /// shutdown drains them.
    pub(crate) tasks: TaskTracker,
    /// The runtime the app connected its brokers on, captured when the subscription opens. Work
    /// that must outlive the delivery (a continuation, a delayed retry copy) runs here, wherever
    /// the handler itself runs, and a handler reaches it through
    /// [`Context::main_runtime`](super::context::Context::main_runtime).
    pub(crate) runtime: MainRuntime,
    /// On a dedicated thread (`threads(n)`), the thread's own tracker: the thread waits for what
    /// its deliveries left behind before it ends, so a shutdown does not drop it with the
    /// thread's runtime. `None` everywhere else.
    pub(crate) thread_tasks: Option<TaskTracker>,
    /// The harness's recording-and-quiescence hooks for this scope. Empty (uninstalled) outside a
    /// [`TestApp`](crate::testing::TestApp) run, so the per-delivery read is a single atomic load.
    #[cfg(feature = "testing")]
    pub(crate) hooks: Arc<TestHooks>,
    /// This broker's registration index, used to scope recorded deliveries per broker.
    #[cfg(feature = "testing")]
    pub(crate) scope_id: usize,
    /// The subscription of the app this delivery comes through, identifying its records.
    #[cfg(feature = "testing")]
    pub(crate) subscription: usize,
    /// Times every Nth poll of this subscription's handler.
    #[cfg(feature = "poll-diagnostics")]
    pub(crate) poll: PollSampler,
}

/// The handler's future, observed by the poll-time diagnostics where the feature is on and handed
/// through untouched where it is off.
#[cfg(feature = "poll-diagnostics")]
macro_rules! observed {
    ($delivery:expr, $future:expr) => {
        $delivery.poll.observe($future)
    };
}

/// See the feature's counterpart above: without the feature, the expression itself.
#[cfg(not(feature = "poll-diagnostics"))]
macro_rules! observed {
    ($delivery:expr, $future:expr) => {
        $future
    };
}

impl<C> Delivery<C> {
    /// The context one subscription dispatches under: its own retry path over what the whole
    /// scope shares.
    #[cfg_attr(not(feature = "poll-diagnostics"), allow(unused_variables))]
    pub(crate) fn for_subscription(
        scope: &ScopeDelivery,
        subscription: &str,
        retry: Option<DeferredRetry<C>>,
        declaration: RetryDeclaration,
    ) -> Self {
        Self {
            retry,
            declaration,
            tasks: scope.tasks().clone(),
            // A subscription opens inside the app's startup, on the runtime it connected on.
            runtime: MainRuntime::current(),
            thread_tasks: None,
            #[cfg(feature = "testing")]
            hooks: Arc::clone(scope.hooks()),
            #[cfg(feature = "testing")]
            scope_id: scope.scope_id(),
            #[cfg(feature = "testing")]
            subscription: scope.subscription(),
            #[cfg(feature = "poll-diagnostics")]
            poll: scope.poll().register(subscription),
        }
    }

    /// A delivery context outside any scope, for tests that drive the dispatch functions
    /// directly.
    #[cfg(test)]
    pub(crate) fn detached(retry: Option<DeferredRetry<C>>, tasks: TaskTracker) -> Self {
        Self {
            retry,
            declaration: RetryDeclaration::new(),
            tasks,
            runtime: MainRuntime::new(test_runtime()),
            thread_tasks: None,
            #[cfg(feature = "testing")]
            hooks: Arc::new(TestHooks::detached()),
            #[cfg(feature = "testing")]
            scope_id: 0,
            #[cfg(feature = "testing")]
            subscription: 0,
            #[cfg(feature = "poll-diagnostics")]
            poll: PollDiagnostics::new().register("detached"),
        }
    }

    /// A delivery context publishing its retry copies through `publisher` at `destination`. For
    /// tests.
    #[cfg(all(test, feature = "memory"))]
    pub(crate) fn deferring_to(
        publisher: Arc<dyn ErasedRetryPublisher<C>>,
        destination: &str,
        tasks: TaskTracker,
    ) -> Self {
        Self::detached(
            Some(DeferredRetry {
                publisher,
                destination: Some(Arc::from(destination)),
            }),
            tasks,
        )
    }

    /// The same context with a declaration on it, as a mount site's `max_attempts(..)` and
    /// `dead_letter(..)` leave it. For tests.
    #[cfg(all(test, feature = "memory"))]
    pub(crate) fn declaring(mut self, declaration: RetryDeclaration) -> Self {
        self.declaration = declaration;
        self
    }

    /// An empty delivery context: no deferred retry, a fresh continuation tracker. For tests.
    #[cfg(test)]
    pub(crate) fn empty() -> Self {
        Self::with_tasks(TaskTracker::new())
    }

    /// An empty delivery context carrying a caller-owned continuation tracker, so a test can
    /// observe the post-settle continuations spawned through it.
    #[cfg(test)]
    pub(crate) fn with_tasks(tasks: TaskTracker) -> Self {
        Self::detached(None, tasks)
    }
}

impl<C> Delivery<C> {
    /// Spawns work the delivery leaves behind (a continuation, a post-settle hook, the timer of a
    /// deferred copy) on the runtime the delivery runs on, tracked so a graceful shutdown waits
    /// for it.
    pub(crate) fn spawn_after<Work>(&self, work: Work)
    where
        Work: Future<Output = ()> + Send + 'static,
    {
        match &self.thread_tasks {
            None => drop(self.tasks.spawn(work)),
            Some(own) => drop(self.tasks.spawn(own.track_future(work))),
        }
    }

    /// The same context for one dedicated thread, whose leftovers `own` tracks as well.
    fn on_thread(&self, own: TaskTracker) -> Self {
        Self {
            retry: self.retry.as_ref().map(|retry| DeferredRetry {
                publisher: Arc::clone(&retry.publisher),
                destination: retry.destination.clone(),
            }),
            declaration: self.declaration.clone(),
            tasks: self.tasks.clone(),
            runtime: self.runtime.clone(),
            thread_tasks: Some(own),
            #[cfg(feature = "testing")]
            hooks: Arc::clone(&self.hooks),
            #[cfg(feature = "testing")]
            scope_id: self.scope_id,
            #[cfg(feature = "testing")]
            subscription: self.subscription,
            #[cfg(feature = "poll-diagnostics")]
            poll: self.poll.for_thread(),
        }
    }
}

/// The runtime a delivery context built by a unit test answers `main_runtime` with: the test's
/// own where it runs inside one, and a shared idle one where a synchronous test builds a context.
#[cfg(test)]
fn test_runtime() -> Handle {
    static IDLE: std::sync::LazyLock<tokio::runtime::Runtime> = std::sync::LazyLock::new(|| {
        tokio::runtime::Builder::new_current_thread()
            .enable_all()
            .build()
            .expect("a current-thread runtime builds")
    });
    Handle::try_current().unwrap_or_else(|_| IDLE.handle().clone())
}

impl<C> fmt::Debug for Delivery<C> {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        f.debug_struct("Delivery")
            .field(
                "retry_destination",
                &self.retry.as_ref().map(|retry| &retry.destination),
            )
            .field("declaration", &self.declaration)
            .field("pending_continuations", &self.tasks.len())
            .finish_non_exhaustive()
    }
}

/// One delivery, between the stream that yielded it and the settle that consumes it.
///
/// Everything on the path borrows the delivery out of the slot and the settle takes it, so a
/// delivery is copied twice whatever its size: once out of the stream, once into
/// [`ack(self)`](IncomingMessage::ack). The futures in between hold a pointer to the slot rather
/// than the value, which is what keeps a pinned dispatch the same size for a 16-byte delivery and
/// a 16-kilobyte one.
///
/// The inside is an `Option` because Rust cannot move a value out of a place its caller owns. A
/// slot is never empty on the dispatch path: the sequential loop's one slot is filled by the poll
/// that takes a delivery off the stream and emptied by the settle, and a pooled worker's is filled
/// where it is declared; taking the delivery is the last thing that happens to it.
pub(crate) struct Slot<M>(Option<M>);

/// What a slot says when it is read after the settle has taken the delivery.
const SETTLED: &str = "the delivery has already been settled";

impl<M> Slot<M> {
    /// The slot holding `msg`.
    pub(crate) const fn new(msg: M) -> Self {
        Self(Some(msg))
    }

    /// A slot with nothing in it yet: the sequential loop's own, which the stream fills in place.
    const fn empty() -> Self {
        Self(None)
    }

    /// Puts the next delivery in the slot.
    fn fill(&mut self, msg: M) {
        self.0 = Some(msg);
    }

    /// The delivery, for everything that only reads it.
    ///
    /// # Panics
    ///
    /// Panics once the settle has taken the delivery, which is the last thing the dispatch path
    /// does with a slot.
    pub(crate) fn get(&self) -> &M {
        self.0.as_ref().expect(SETTLED)
    }

    /// The delivery, for the settle that consumes it.
    ///
    /// # Panics
    ///
    /// Panics if called twice; see [`Slot::get`].
    pub(crate) fn take(&mut self) -> M {
        self.0.take().expect(SETTLED)
    }
}

/// Spawns a task that drives `subscriber` through `handler` until `shutdown` is triggered or the
/// stream terminates. Each delivery is given a [`Context`] built from `name`, the message headers,
/// shared `state`, and the `delivery` publish context.
pub(crate) fn spawn_dispatch<S, H, C, St>(
    mut subscriber: S,
    handler: Arc<H>,
    shutdown: Shutdown,
    name: Arc<str>,
    state: Arc<St>,
    delivery: Arc<Delivery<C>>,
    failure: DispatchFailure,
) -> JoinHandle<()>
where
    S: Subscriber + Send + 'static,
    H: Handler<S::Message, C, St> + 'static,
    C: crate::BuildContext<S::Message> + Send + Sync + 'static,
    St: Send + Sync + 'static,
{
    tokio::spawn(async move {
        let mut stream = std::pin::pin!(subscriber.stream());
        let mut cancelled = std::pin::pin!(shutdown.cancelled());
        // One encode buffer for the whole loop: a reply that leaves through a publisher which
        // reads the payload is written into this one every time, so the subscription allocates
        // it once and never again.
        let mut encode = BytesMut::new();
        // The loop's slot, one for the whole loop: the stream's poll writes each delivery
        // straight into it, and the settle takes it out again. A delivery returned by value
        // through the turn and bound to a local on its way here would be copied twice more.
        let mut slot = Slot::empty();
        loop {
            match turn_into(&shutdown, stream.as_mut(), cancelled.as_mut(), &mut slot).await {
                Turn::Delivery(Ok(())) => {
                    // Pinned rather than awaited in place: a delivery's future carries the
                    // context, the handler's own state and the settle path, and awaiting the
                    // call expression makes the loop build it on the stack and copy it into
                    // its own state on every delivery.
                    let handling = std::pin::pin!(dispatch(
                        &*handler,
                        &mut slot,
                        &mut encode,
                        &name,
                        &state,
                        &delivery,
                        &failure
                    ));
                    handling.await;
                }
                Turn::Delivery(Err(err)) => {
                    error!(
                        target: "ruststream::dispatch",
                        error = %err,
                        "subscriber stream error",
                    );
                }
                Turn::Ended => {
                    debug!(
                        target: "ruststream::dispatch",
                        subscriber = %name,
                        "subscriber stream ended",
                    );
                    break;
                }
                Turn::Shutdown => break,
            }
        }
    })
}

/// Waits for whichever comes first: the next item off `stream`, or `shutdown`. A delivery is
/// written into the loop's slot where it lands rather than returned by value.
///
/// The stream is polled first and `cancelled` only where it answers `Pending`, so a turn that
/// finds a delivery ready never touches the token's waiter list: what shutdown costs a busy
/// subscription is the one relaxed load of [`Shutdown`]'s flag. `cancelled` is the token's own
/// wait future, built once per subscription by the caller and pinned across the whole loop, so
/// the turn that does park registers with the waiter list once rather than on every iteration,
/// which is what a `select!` over `cancelled()` charges.
///
/// # Cancel safety
///
/// Both halves are cancel-safe: a delivery is written into the slot in the same poll that takes
/// it off the stream, and a registration with the token's waiter list outlives the poll. Once
/// `cancelled` has resolved the flag stays set - a token never un-cancels - so the check at the
/// top answers every later turn and the resolved future is never polled again.
///
/// The other loops write the same wait out rather than share it: one instantiation shared with
/// them is enough for the compiler to stop inlining it here.
fn turn_into<'s, St, Message, Failure, Wait>(
    shutdown: &Shutdown,
    stream: Pin<&'s mut St>,
    cancelled: Pin<&'s mut Wait>,
    slot: &'s mut Slot<Message>,
) -> TurnInto<'s, St, Message, Wait>
where
    St: Stream<Item = Result<Message, Failure>>,
    Wait: Future<Output = ()>,
{
    TurnInto {
        shutdown: shutdown.is_cancelled(),
        stream,
        cancelled,
        slot,
    }
}

/// The future [`turn_into`] returns: written out rather than an `async fn` so its poll carries an
/// inline hint, which a closure inside `poll_fn` cannot.
struct TurnInto<'s, St, Message, Wait> {
    shutdown: bool,
    stream: Pin<&'s mut St>,
    cancelled: Pin<&'s mut Wait>,
    slot: &'s mut Slot<Message>,
}

impl<St, Message, Failure, Wait> Future for TurnInto<'_, St, Message, Wait>
where
    St: Stream<Item = Result<Message, Failure>>,
    Wait: Future<Output = ()>,
{
    type Output = Turn<Result<(), Failure>>;

    #[inline]
    fn poll(self: Pin<&mut Self>, cx: &mut TaskContext<'_>) -> Poll<Self::Output> {
        // Every field is a reference or a flag, so the future is `Unpin`.
        let this = self.get_mut();
        if this.shutdown {
            return Poll::Ready(Turn::Shutdown);
        }
        match this.stream.as_mut().poll_next(cx) {
            Poll::Ready(Some(Ok(msg))) => {
                this.slot.fill(msg);
                Poll::Ready(Turn::Delivery(Ok(())))
            }
            Poll::Ready(Some(Err(err))) => Poll::Ready(Turn::Delivery(Err(err))),
            Poll::Ready(None) => Poll::Ready(Turn::Ended),
            Poll::Pending => this.cancelled.as_mut().poll(cx).map(|()| Turn::Shutdown),
        }
    }
}

/// What one turn of a dispatch loop found on its subscriber.
enum Turn<T> {
    /// The subscriber yielded a delivery, or an error reading one.
    Delivery(T),
    /// The subscriber's stream ended; the loop is done.
    Ended,
    /// Shutdown was signalled; the loop stops without touching the stream again.
    Shutdown,
}

/// Waits for whichever comes first: a finished worker off `tasks`, or `shutdown`.
///
/// [`turn_into`]'s counterpart for a full pool, in the same order and over the same pinned token
/// future: a worker that has already finished is reaped without the token's waiter list being
/// touched. `Ended` answers an empty set, which a full pool cannot be.
///
/// # Cancel safety
///
/// Cancel-safe, like [`turn_into`]: a `JoinSet` keeps its finished workers across polls, and the
/// registration a parked poll made with the waiter list outlives that poll.
async fn reap<Wait, Done>(
    tasks: &mut JoinSet<Done>,
    mut cancelled: Pin<&mut Wait>,
) -> Turn<Result<Done, JoinError>>
where
    Wait: Future<Output = ()>,
    Done: 'static,
{
    poll_fn(|cx| match tasks.poll_join_next(cx) {
        Poll::Ready(Some(joined)) => Poll::Ready(Turn::Delivery(joined)),
        Poll::Ready(None) => Poll::Ready(Turn::Ended),
        Poll::Pending => cancelled.as_mut().poll(cx).map(|()| Turn::Shutdown),
    })
    .await
}

/// Spawns a task that drives `subscriber` through `handler` with `workers.count` long-lived
/// workers: a pool that hands each delivery to a free worker, or with `by_key` lanes that hand it
/// to the worker its key hashes to (see [`pool`]).
///
/// Sequential policies delegate to [`spawn_dispatch`]. On shutdown the stream stops being
/// polled and the workers drain what they hold; if the app's `shutdown_timeout` aborts this task,
/// the workers abort with it.
// The parts are independent and each spawn site passes its own; bundling them into a struct
// would hide that.
#[allow(clippy::too_many_arguments)]
pub(crate) fn spawn_dispatch_workers<S, H, C, St>(
    subscriber: S,
    handler: Arc<H>,
    shutdown: Shutdown,
    name: Arc<str>,
    state: Arc<St>,
    delivery: Arc<Delivery<C>>,
    failure: DispatchFailure,
    workers: Workers,
) -> Result<JoinHandle<()>, StartThreadError>
where
    S: Subscriber + Send + 'static,
    S::Message: Send + Sync + 'static,
    H: Handler<S::Message, C, St> + 'static,
    C: crate::BuildContext<S::Message> + Send + Sync + 'static,
    St: Send + Sync + 'static,
{
    // Before the harness maps the placement away: the declaration is what the warning advises.
    #[cfg(feature = "poll-diagnostics")]
    if matches!(workers.placement, Placement::Threads) {
        delivery.poll.on_dedicated_threads();
    }
    // The harness drives every subscription on the test's own runtime, so a timer the handler
    // arms is one `advance` reaches; it does not reproduce the thread topology.
    #[cfg(feature = "testing")]
    let workers = if delivery.hooks.coordinator().is_some() {
        workers.on_runtime()
    } else {
        workers
    };
    if workers.is_sequential() {
        return Ok(spawn_dispatch(
            subscriber, handler, shutdown, name, state, delivery, failure,
        ));
    }
    if matches!(workers.placement, Placement::Threads) {
        return spawn_dispatch_threads(
            subscriber, handler, shutdown, name, state, delivery, failure, workers,
        );
    }
    Ok(pool::spawn_dispatch_pool(
        subscriber, handler, shutdown, name, state, delivery, failure, workers,
    ))
}

/// Starts the dedicated threads of a `threads(n)` subscription and spawns its loop on the app's
/// runtime (see [`threads`]).
// See `spawn_dispatch_workers`: each part is the registration's own.
#[allow(clippy::too_many_arguments)]
fn spawn_dispatch_threads<S, H, C, St>(
    mut subscriber: S,
    handler: Arc<H>,
    shutdown: Shutdown,
    name: Arc<str>,
    state: Arc<St>,
    delivery: Arc<Delivery<C>>,
    failure: DispatchFailure,
    workers: Workers,
) -> Result<JoinHandle<()>, StartThreadError>
where
    S: Subscriber + Send + 'static,
    S::Message: Send + Sync + 'static,
    H: Handler<S::Message, C, St> + 'static,
    C: crate::BuildContext<S::Message> + Send + Sync + 'static,
    St: Send + Sync + 'static,
{
    let shared = Arc::new(pool::Shared {
        handler,
        name,
        state,
        delivery,
        failure,
    });
    let threads = Threads::start(&shared.name, workers.count, workers.by_key, |own| {
        let shared = Arc::clone(&shared);
        let delivery = shared.delivery.on_thread(own);
        // One encode buffer per thread, for the reason the sequential loop has one.
        let mut encode = BytesMut::new();
        async move |msg: S::Message| {
            let mut slot = Slot::new(msg);
            dispatch(
                &*shared.handler,
                &mut slot,
                &mut encode,
                &shared.name,
                &shared.state,
                &delivery,
                &shared.failure,
            )
            .await;
        }
    })?;
    Ok(tokio::spawn(async move {
        threads
            .feed(
                subscriber.stream(),
                &shared.name,
                &shutdown,
                <S::Message as IncomingMessage>::partition_key,
            )
            .await;
    }))
}

fn lane_of(key: &[u8], lanes: usize) -> usize {
    let mut hasher = DefaultHasher::new();
    key.hash(&mut hasher);
    // The modulo keeps the value below `lanes`, which is a usize.
    #[allow(clippy::cast_possible_truncation)]
    {
        (hasher.finish() % lanes as u64) as usize
    }
}

/// What a finished worker handed back, or nothing and a log line where it failed.
fn log_worker_exit<Done>(joined: Result<Done, JoinError>) -> Option<Done> {
    match joined {
        Ok(done) => Some(done),
        Err(err) => {
            error!(target: "ruststream::dispatch", error = %err, "worker task failed");
            None
        }
    }
}

/// Spawns a task that drives `subscriber` through a batch `handler`, one
/// [`BatchSubscriber::batches`] item per invocation, until `shutdown` is triggered or the stream
/// terminates. The handler owns the batch's deliveries and settles each of them.
///
/// With a non-sequential `workers` policy, up to `workers.count` batches are in flight at once,
/// each in its own task; keyed lanes do not apply at batch granularity (the macro rejects
/// `by_key` on batch forms), so a keyed policy degrades to the plain pool.
#[allow(clippy::too_many_arguments)] // See spawn_dispatch_workers.
pub(crate) fn spawn_batch_dispatch<S, H, C, St>(
    mut subscriber: S,
    handler: Arc<H>,
    shutdown: Shutdown,
    name: Arc<str>,
    state: Arc<St>,
    delivery: Arc<Delivery<C>>,
    failure: DispatchFailure,
    workers: Workers,
    batch_size: NonZeroUsize,
) -> Result<JoinHandle<()>, StartThreadError>
where
    S: BatchSubscriber + Send + 'static,
    S::Message: Send + 'static,
    S::Batch: 'static,
    H: BatchHandler<S::Batch, C, St> + 'static,
    C: Send + Sync + 'static,
    St: Send + Sync + 'static,
{
    // See `spawn_dispatch_workers`: the declaration is what the warning advises, and the harness
    // runs every subscription on the test's runtime.
    #[cfg(feature = "poll-diagnostics")]
    if matches!(workers.placement, Placement::Threads) {
        delivery.poll.on_dedicated_threads();
    }
    #[cfg(feature = "testing")]
    let workers = if delivery.hooks.coordinator().is_some() {
        workers.on_runtime()
    } else {
        workers
    };
    if matches!(workers.placement, Placement::Threads) {
        return spawn_batch_threads(
            subscriber, handler, shutdown, name, state, delivery, failure, workers, batch_size,
        );
    }
    Ok(tokio::spawn(async move {
        // The registration's own batch size, straight to the broker: whatever comes back is the
        // batch the handler sees.
        let mut stream = std::pin::pin!(subscriber.batches(batch_size));
        let mut cancelled = std::pin::pin!(shutdown.cancelled());
        let mut tasks = JoinSet::new();
        // One decode buffer for the whole loop: the sequential path lends the same one to every
        // batch, so the slice a handler reads is allocated once for the subscription. The encode
        // buffer beside it is the same bargain on the way out. A pooled batch runs beside the
        // others and cannot share these, so it takes a pair of its own from `spare`, where the
        // pool keeps what finished workers handed back.
        let mut scratch = <H as BatchHandler<S::Batch, C, St>>::Scratch::default();
        let mut encode = BytesMut::new();
        let mut spare = Vec::new();
        loop {
            if shutdown.is_cancelled() {
                break;
            }
            if tasks.len() >= workers.count {
                // The pool is full: reap a finished worker before polling for more.
                match reap(&mut tasks, cancelled.as_mut()).await {
                    Turn::Delivery(joined) => spare.extend(log_worker_exit(joined)),
                    // A full pool holds a worker, so `Ended` is shutdown's twin here: either
                    // way the drain below joins whatever is still running.
                    Turn::Ended | Turn::Shutdown => break,
                }
                continue;
            }
            // The stream first, the token only where the stream has nothing: see `turn_into`, whose
            // wait this is.
            let pulled = poll_fn(|cx| match stream.as_mut().poll_next(cx) {
                Poll::Ready(Some(item)) => Poll::Ready(Turn::Delivery(item)),
                Poll::Ready(None) => Poll::Ready(Turn::Ended),
                Poll::Pending => cancelled.as_mut().poll(cx).map(|()| Turn::Shutdown),
            })
            .await;
            match pulled {
                Turn::Delivery(Ok(batch)) => {
                    if workers.is_sequential() {
                        // Turbofish: the adapter handlers are generic over the batch context,
                        // so the spawn's own parameter names it.
                        run_batch::<_, _, C, _>(
                            &*handler,
                            batch,
                            &mut scratch,
                            &mut encode,
                            &name,
                            &state,
                            &delivery,
                            &failure,
                        )
                        .await;
                    } else {
                        let handler = Arc::clone(&handler);
                        let name = Arc::clone(&name);
                        let state = Arc::clone(&state);
                        let delivery = Arc::clone(&delivery);
                        let failure = failure.clone();
                        let (mut scratch, mut encode) = spare.pop().unwrap_or_default();
                        tasks.spawn(async move {
                            run_batch::<_, _, C, _>(
                                &*handler,
                                batch,
                                &mut scratch,
                                &mut encode,
                                &name,
                                &state,
                                &delivery,
                                &failure,
                            )
                            .await;
                            (scratch, encode)
                        });
                    }
                }
                Turn::Delivery(Err(err)) => {
                    error!(
                        target: "ruststream::dispatch",
                        error = %err,
                        "subscriber stream error",
                    );
                }
                Turn::Ended => {
                    debug!(
                        target: "ruststream::dispatch",
                        subscriber = %name,
                        "subscriber stream ended",
                    );
                    break;
                }
                Turn::Shutdown => break,
            }
        }
        while let Some(joined) = tasks.join_next().await {
            log_worker_exit(joined);
        }
    }))
}

/// The batch counterpart of [`spawn_dispatch_threads`]: each thread takes whole batches off its
/// ring, with a decode and an encode buffer of its own. Keyed lanes do not apply to batches, so a
/// keyed policy spreads batches as the plain one does.
#[allow(clippy::too_many_arguments)] // See spawn_dispatch_workers.
fn spawn_batch_threads<S, H, C, St>(
    mut subscriber: S,
    handler: Arc<H>,
    shutdown: Shutdown,
    name: Arc<str>,
    state: Arc<St>,
    delivery: Arc<Delivery<C>>,
    failure: DispatchFailure,
    workers: Workers,
    batch_size: NonZeroUsize,
) -> Result<JoinHandle<()>, StartThreadError>
where
    S: BatchSubscriber + Send + 'static,
    S::Message: Send + 'static,
    S::Batch: 'static,
    H: BatchHandler<S::Batch, C, St> + 'static,
    C: Send + Sync + 'static,
    St: Send + Sync + 'static,
{
    let shared = Arc::new(pool::Shared {
        handler,
        name,
        state,
        delivery,
        failure,
    });
    let threads = Threads::start(&shared.name, workers.count, false, |own| {
        let shared = Arc::clone(&shared);
        let delivery = shared.delivery.on_thread(own);
        let mut scratch = <H as BatchHandler<S::Batch, C, St>>::Scratch::default();
        let mut encode = BytesMut::new();
        async move |batch: S::Batch| {
            run_batch::<_, _, C, _>(
                &*shared.handler,
                batch,
                &mut scratch,
                &mut encode,
                &shared.name,
                &shared.state,
                &delivery,
                &shared.failure,
            )
            .await;
        }
    })?;
    Ok(tokio::spawn(async move {
        threads
            .feed(
                subscriber.batches(batch_size),
                &shared.name,
                &shutdown,
                |_| None,
            )
            .await;
    }))
}

async fn dispatch<H, M, C, St>(
    handler: &H,
    slot: &mut Slot<M>,
    encode: &mut BytesMut,
    name: &str,
    state: &St,
    delivery: &Delivery<C>,
    failure: &DispatchFailure,
) where
    H: Handler<M, C, St>,
    C: crate::BuildContext<M> + Send + Sync + 'static,
    M: IncomingMessage,
    // The dispatch future is awaited inside a spawned task, so it must be `Send`: the context
    // borrows `&St` across the handler await, which requires `St: Sync`.
    St: Send + Sync,
{
    // Settling the message is what releases the harness's quiescence wait, and the post-settle
    // continuations are spawned after it; one in-flight token spanning the whole dispatch keeps a
    // `drain` from running before they exist.
    #[cfg(feature = "testing")]
    let watcher = delivery.hooks.coordinator().cloned();
    #[cfg(feature = "testing")]
    if let Some(coordinator) = &watcher {
        coordinator.enqueued();
    }
    // Build the broker's typed per-delivery context from the message, then attach the fail-fast
    // handle. The delivery is borrowed out of the slot for the whole handler phase and moves
    // only at the settle below.
    // The context lives in a block of its own and is dropped where it stands when the block ends:
    // a `drop(ctx)` call would move the whole context out of this future's state into the call's
    // argument first, a `memcpy` of the context per delivery.
    let (settle, continuations) = {
        let msg = slot.get();
        let cx = C::build(msg);
        // The delivery itself, not its header map: a handler that reads no headers never reaches
        // the broker's accessor, which is a parse and a buffer on most transports.
        let source = FromDelivery(msg);
        let mut ctx = Context::new(name, &source, state, cx, delivery)
            .with_failfast(&failure.shutdown)
            .with_encode_buffer(encode)
            .with_decode_policy(failure.policies.decode);
        // Catch a panicking handler so it cannot silently kill the dispatch loop (which would stop
        // the subscriber consuming) or leave the message unsettled. AssertUnwindSafe is required
        // because the future borrows `&mut ctx`; that state is discarded with the failed delivery.
        // Under the harness, the invocation runs in a task-local slot scope so publishes made
        // through injected `Out` publishers are attributed to their slot.
        #[cfg(feature = "testing")]
        let result = in_harness_scope(
            harness_scope(delivery),
            observed!(
                delivery,
                AssertUnwindSafe(handler.handle(msg, &mut ctx)).catch_unwind()
            ),
        )
        .await;
        #[cfg(not(feature = "testing"))]
        let result = observed!(
            delivery,
            AssertUnwindSafe(handler.handle(msg, &mut ctx)).catch_unwind()
        )
        .await;
        #[cfg(feature = "testing")]
        let panicked = result.is_err();
        // Resolve into a `HandlerOutcome` regardless of whether the handler panicked. `None`
        // means a fail-fast panic tore the service down and left the message unsettled (a broker
        // with redelivery hands it back after the restart).
        let settle = match result {
            Ok(s) => Some(s),
            Err(payload) => {
                let reason = panic_reason(payload.as_ref());
                error!(
                    target: "ruststream::dispatch",
                    subscription = %name,
                    panic = %reason,
                    "handler panicked",
                );
                match failure.policies.panic {
                    FailurePolicy::FailFast => {
                        failure
                            .shutdown
                            .signal(name, &format!("handler panicked: {reason}"));
                        None
                    }
                    other => Some(
                        other
                            .settlement()
                            .map_or_else(super::handler::HandlerOutcome::drop, Into::into),
                    ),
                }
            }
        };
        // Drain the matching post-settle hooks BEFORE settling: `ctx` borrows `msg`'s headers, and
        // settling consumes `msg`. The drained futures own their captures. A fail-fast (no
        // settlement) runs no hooks. Most deliveries register none, and those pay the branch alone:
        // the list, its scan and the drop glue of both belong to the deliveries that did register
        // one.
        let continuations = match settle.as_ref() {
            Some(s) if ctx.has_hooks() => Some(ctx.take_hooks_for(s.outcome())),
            _ => None,
        };
        // The harness records what the handler saw and how it settled, BEFORE settling the
        // message: the matching decrement runs in the broker message's `Drop` (during
        // `settle_outcome`, or at the end of this function on the fail-fast path), so the record
        // is in place by the time `drive` wakes. Captured here because `settle_outcome` consumes
        // `msg` and dropping `ctx` clears the decode flag.
        #[cfg(feature = "testing")]
        if let Some(coordinator) = delivery.hooks.coordinator() {
            coordinator.record(Record {
                scope_id: delivery.scope_id,
                subscription: delivery.subscription,
                name: name.to_owned(),
                deliveries: vec![Delivered {
                    raw: Bytes::copy_from_slice(msg.payload()),
                    value: ctx.take_carried(),
                    settle: settle.as_ref().map(super::handler::HandlerOutcome::outcome),
                }],
                panicked,
                decode_failed: ctx.took_decode_failed(),
            });
        }
        (settle, continuations)
    };
    if let Some(mut s) = settle {
        // Named for the same reason as the delivery's own future above: the settle path is
        // built where it is polled instead of being copied into this future's state.
        let settling = settle_outcome(slot, s.outcome(), name, delivery, C::build as fn(&M) -> C);
        settling.await;
        // Spawn the `and_after` continuation (if any) onto the tracked set so a graceful shutdown
        // drains it. At-most-once: the message is already settled, so a lost or panicking
        // continuation never redelivers it.
        if let Some(after) = s.take_after() {
            delivery.spawn_after(after);
        }
    } else {
        // A fail-fast left the delivery unsettled: it is released here, as it always was at the
        // end of its dispatch, rather than when the loop's slot is next filled.
        drop(slot.take());
    }
    // Context-registered hooks run after the message is settled: at-most-once, off the delivery
    // path. They ride the same app-wide tracker as an `and_after` continuation, so one drain
    // covers both - the harness's `drain` and the shutdown's alike.
    if let Some(continuations) = continuations {
        for fut in continuations {
            delivery.spawn_after(fut);
        }
    }
    #[cfg(feature = "testing")]
    if let Some(coordinator) = &watcher {
        coordinator.consumed();
    }
}

/// Runs one batch through its handler under panic protection. The handler owns and settles the
/// batch's deliveries, so a panic there has already consumed them: the panic policy can only tear
/// the service down (`fail_fast`) or be logged and skipped. Per-element settlement is out of scope
/// (see the batch decode path for per-element decode handling). Ungated `after_settle` hooks run
/// once the batch has settled (per-element outcomes make a gated hook ill-defined on a batch).
// The loop hands this everything one batch needs, the buffers it lends included: see
// spawn_dispatch_workers.
#[allow(clippy::too_many_arguments)]
async fn run_batch<H, Batch, C, St>(
    handler: &H,
    batch: Batch,
    scratch: &mut H::Scratch,
    encode: &mut BytesMut,
    name: &str,
    state: &St,
    delivery: &Delivery<C>,
    failure: &DispatchFailure,
) where
    H: BatchHandler<Batch, C, St>,
    C: Send + Sync + 'static,
    St: Send + Sync,
{
    let batch = H::take(batch);
    // A batch spans many deliveries, so its context carries only subscription-scoped data, built
    // from its first delivery, or from the batch itself where the handler lends its values; the
    // shared app state is threaded the same way as on the single-message path. A batch with
    // nothing in it has nothing to settle and nothing to build a context from.
    let Some(cx) = batch.context() else { return };
    let empty = HeaderMap::new();
    let mut ctx = Context::new(name, &empty, state, cx, delivery)
        .with_failfast(&failure.shutdown)
        .with_encode_buffer(encode)
        .with_decode_policy(failure.policies.decode);
    // See `dispatch`: the harness scope attributes `Out` publishes to their slot, and lets the
    // batch settle path record the batch it applied.
    // A batch settles its own deliveries inside the handler (a panic settles them by dropping
    // them), so the last decrement lands before the batch record and the fail-fast signal below.
    // One extra in-flight token spans the whole dispatch, so a harness driving to quiescence
    // cannot return into that window.
    #[cfg(feature = "testing")]
    let watcher = delivery.hooks.coordinator().cloned();
    // A panicking batch settles nothing, so what it carried is captured here (the handler owns
    // the deliveries and a panic consumes them) to record the call the settle path never reached.
    // Only a harness reads that record, so a dispatch no harness watches captures nothing.
    #[cfg(feature = "testing")]
    let mut batch = batch;
    #[cfg(feature = "testing")]
    let unsettled = watcher.as_ref().map(|_| batch.unsettled());
    #[cfg(feature = "testing")]
    if let Some(coordinator) = &watcher {
        coordinator.enqueued();
    }
    #[cfg(feature = "testing")]
    let result = in_harness_scope(
        harness_scope(delivery),
        observed!(
            delivery,
            AssertUnwindSafe(handler.handle_batch(batch, scratch, &mut ctx)).catch_unwind()
        ),
    )
    .await;
    #[cfg(not(feature = "testing"))]
    let result = observed!(
        delivery,
        AssertUnwindSafe(handler.handle_batch(batch, scratch, &mut ctx)).catch_unwind()
    )
    .await;
    match result {
        Ok(()) => {
            // As on the single-message path: a batch that registered no hook pays the branch.
            if ctx.has_hooks() {
                for fut in ctx.take_settle_hooks() {
                    delivery.spawn_after(fut);
                }
            }
        }
        Err(payload) => {
            let reason = panic_reason(payload.as_ref());
            error!(
                target: "ruststream::dispatch",
                subscription = %name,
                panic = %reason,
                "batch handler panicked",
            );
            #[cfg(feature = "testing")]
            if let (Some(coordinator), Some(deliveries)) = (&watcher, unsettled) {
                coordinator.record(Record {
                    scope_id: delivery.scope_id,
                    subscription: delivery.subscription,
                    name: name.to_owned(),
                    deliveries,
                    panicked: true,
                    decode_failed: false,
                });
            }
            if failure.policies.panic == FailurePolicy::FailFast {
                failure
                    .shutdown
                    .signal(name, &format!("batch handler panicked: {reason}"));
            }
        }
    }
    #[cfg(feature = "testing")]
    if let Some(coordinator) = &watcher {
        coordinator.consumed();
    }
}

/// The harness scope a delivery runs under, or `None` when no [`TestApp`](crate::testing::TestApp)
/// is driving this app.
#[cfg(feature = "testing")]
fn harness_scope<C>(delivery: &Delivery<C>) -> Option<HarnessScope> {
    delivery
        .hooks
        .coordinator()
        .cloned()
        .map(|coordinator| HarnessScope::new(coordinator, delivery.scope_id, delivery.subscription))
}

/// Settles one delivery by `outcome`, logging an ack / nack failure without propagating it.
///
/// The single place a settlement reaches the broker, single-message and batch paths alike: a
/// second one would be free to answer a [`NackAfter`](HandlerResult::NackAfter) differently, and
/// the retry path below is exactly the part that is easy to leave out.
pub(crate) async fn settle_outcome<M, C>(
    slot: &mut Slot<M>,
    outcome: HandlerResult,
    name: &str,
    delivery: &Delivery<C>,
    build_cx: fn(&M) -> C,
) where
    M: IncomingMessage,
    C: Send + Sync + 'static,
{
    let ack_result = match outcome {
        HandlerResult::Ack => slot.take().ack().await,
        HandlerResult::Nack { requeue: false } => slot.take().nack(false).await,
        HandlerResult::Nack { requeue: true } => settle_retry(slot, name, delivery, build_cx).await,
        HandlerResult::NackAfter { delay } => {
            settle_nack_after(slot, name, delay, delivery, build_cx).await
        }
    };
    if let Err(err) = ack_result {
        warn!(
            target: "ruststream::dispatch",
            subscription = %name,
            error = %err,
            "ack / nack failed",
        );
    }
}

/// Where the next copy of one delivery goes, read off the registration's declaration and the
/// count the delivery carries.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
enum Redelivery<'a> {
    /// Back to the subscription itself: the attempts are not spent and no destination has taken
    /// the copies over.
    Subscription,
    /// To the declared destination.
    DeadLetter {
        /// Where the copy is published.
        destination: &'a str,
        /// Whether it goes there because the attempts are spent, rather than because the
        /// registration named a destination and no cap.
        at_cap: bool,
    },
    /// Rejected without a copy: the attempts are spent and the registration named nowhere to send
    /// it. A rejection keeps the broker's own dead-letter policy in play where one is configured.
    Reject,
}

/// How many times this message has been delivered, counting this delivery.
///
/// The broker's own count where the transport keeps one, the framework's header otherwise, never
/// both. A transport that counts its deliveries answers for all of them, the ones a delay or a
/// requeue brought round included. The header starts absent and is incremented on every copy
/// published for the delivery, so the first delivery counts as one either way.
fn attempt_of<M: IncomingMessage>(msg: &M) -> u64 {
    msg.redelivery_count()
        .unwrap_or_else(|| current_retry_count(msg.headers()) + 1)
}

/// Reads the registration's declaration against this delivery.
fn redelivery_of<'d, M: IncomingMessage>(
    msg: &M,
    declaration: &'d RetryDeclaration,
) -> Redelivery<'d> {
    match (declaration.max_attempts(), declaration.dead_letter()) {
        (Some(cap), destination) if attempt_of(msg) >= u64::from(cap.get()) => {
            destination.map_or(Redelivery::Reject, |destination| Redelivery::DeadLetter {
                destination,
                at_cap: true,
            })
        }
        (None, Some(destination)) => Redelivery::DeadLetter {
            destination,
            at_cap: false,
        },
        _ => Redelivery::Subscription,
    }
}

/// Settles an immediate retry ([`Nack { requeue: true }`](HandlerResult::Nack)).
///
/// A registration that declared neither a cap nor a destination gets the broker's own requeue,
/// which is what an immediate retry has always been. Under a declaration the outcome obeys it:
/// where the transport counts its own redeliveries the requeue stays the broker's and the count
/// is read from the delivery, and where it counts none the runtime publishes the copy itself so
/// that the framework's header carries the count forward and the cap applies to an immediate
/// retry as it does to a delayed one.
///
/// A registration whose copies the broker moves itself is never capped here, exactly as on the
/// native delayed path: the broker's requeue is the immediate retry, and the declaration reached
/// the subscription descriptor at startup for the broker to apply. Reading the cap here would
/// settle a spent delivery with a plain rejection, and on a queue that deletes a rejected
/// delivery (SQS) that loses the message instead of leaving it to the redrive policy.
///
/// # Errors
///
/// Returns the [`AckError`] from settling the original delivery.
async fn settle_retry<M, C>(
    slot: &mut Slot<M>,
    name: &str,
    delivery: &Delivery<C>,
    build_cx: fn(&M) -> C,
) -> Result<(), AckError>
where
    M: IncomingMessage,
    C: Send + Sync + 'static,
{
    if delivery.declaration.declares_nothing() {
        return slot.take().nack(true).await;
    }
    // The broker moves this subscription's deliveries itself: its requeue is what an immediate
    // retry is, and the declaration is the broker's to apply, so no cap is read here, as on the
    // native delayed path.
    let Some(retry) = delivery.retry.as_ref() else {
        return slot.take().nack(true).await;
    };
    match redelivery_of(slot.get(), &delivery.declaration) {
        // The broker counts its own redeliveries, so its requeue carries the count forward and
        // there is nothing for a copy to add.
        Redelivery::Subscription if slot.get().redelivery_count().is_some() => {
            slot.take().nack(true).await
        }
        Redelivery::Subscription => {
            let destination = retry.destination.clone();
            publish_copy(slot, name, destination, None, retry, delivery, build_cx).await
        }
        Redelivery::DeadLetter {
            destination,
            at_cap,
        } => {
            if at_cap {
                warn_at_cap(name, attempt_of(slot.get()), Some(destination));
            }
            let destination = Arc::from(destination);
            publish_copy(
                slot,
                name,
                Some(destination),
                None,
                retry,
                delivery,
                build_cx,
            )
            .await
        }
        Redelivery::Reject => {
            warn_at_cap(name, attempt_of(slot.get()), None);
            slot.take().nack(false).await
        }
    }
}

/// Settles a [`NackAfter`](HandlerResult::NackAfter) outcome, choosing native delayed redelivery
/// or the copy the runtime publishes.
///
/// When the broker reports native support (`supports_nack_after`), this defers to
/// [`IncomingMessage::nack_after`] and nothing is published - unless the delivery is already at
/// the registration's cap. The cap is read first, from the one count the transport has (see
/// [`attempt_of`]): the broker's own where it keeps one, the framework's header otherwise, never
/// both. A delivery at the cap goes to the declared dead-letter destination or is rejected,
/// exactly as on the copy path.
/// Where the broker moves a spent delivery itself the cap is not read on this path: the
/// declaration is the subscription descriptor's to map onto the broker's own mechanism.
///
/// Without native support this captures the message, drops the original, and schedules a copy of it
/// after the delay, with the [`RETRY_COUNT_HEADER`] incremented. A zero delay publishes the copy at
/// once, on the dispatch path, as an immediate retry does. The copy goes to the address the
/// subscription's descriptor reported at startup, not to the subscription's name: the two differ
/// wherever a subscription is a resource of its own. It leaves through the registration's retry
/// slot, so the mount site's transforms and the publisher's own headers reach it as they reach any
/// slot publish; it carries bytes already, so no codec encodes it and the call adjusts no
/// per-message settings.
///
/// Once the registration's attempts are spent the copy goes to the declared dead-letter
/// destination instead, immediately, and with no destination declared the delivery is rejected.
///
/// A transport with no settlement at all ([`AckError::Unsupported`] from `nack`, as on MQTT at
/// `QoS` 0, `ZeroMQ`, or Redis pub/sub) still gets the copy: there is no original to drop and no
/// redelivery of the broker's own to fall back on, so the copy is the only form the retry can
/// take.
///
/// Where the broker moves the subscription's deliveries itself and cannot hold one back, the
/// delivery is requeued at once, and a warning says the delay is dropped. A delivery the
/// declaration sends elsewhere (its attempts are spent, or a destination takes every copy) gets
/// the same requeue without the warning: the broker moves it on that requeue, so no delay was due.
///
/// # Errors
///
/// Returns the [`AckError`] from settling the original when the transport does settle but this
/// settle failed. The delivery then stays with the broker, which redelivers it on its own timers,
/// and a copy on top of that would duplicate the message; the caller logs the error.
///
/// # Cancel safety
///
/// A deferred copy runs on a detached task that sleeps for `delay`. It is at-most-once over
/// that window: if the process exits (or the runtime is dropped) before the timer fires, the
/// deferred message is lost, since the original has already been dropped. Brokers that need
/// at-least-once delayed redelivery across a crash must provide native support.
async fn settle_nack_after<M, C>(
    slot: &mut Slot<M>,
    name: &str,
    delay: Duration,
    delivery: &Delivery<C>,
    build_cx: fn(&M) -> C,
) -> Result<(), AckError>
where
    M: IncomingMessage,
    C: Send + Sync + 'static,
{
    if slot.get().supports_nack_after() {
        // The cap is read before the delay reaches the broker: a native redelivery would otherwise
        // circle past a cap nothing in this process ever gets to apply. It is read wherever this
        // process publishes the subscription's copies, because a crate that honours the delay by
        // republishing the delivery increments the header the cap counts with where the transport
        // counts nothing; a broker that moves a spent delivery itself applies the declaration
        // itself.
        if let Some(retry) = delivery.retry.as_ref()
            && delivery.declaration.max_attempts().is_some()
        {
            match redelivery_of(slot.get(), &delivery.declaration) {
                Redelivery::DeadLetter { destination, .. } => {
                    warn_at_cap(name, attempt_of(slot.get()), Some(destination));
                    let destination = Arc::from(destination);
                    return publish_copy(
                        slot,
                        name,
                        Some(destination),
                        None,
                        retry,
                        delivery,
                        build_cx,
                    )
                    .await;
                }
                Redelivery::Reject => {
                    warn_at_cap(name, attempt_of(slot.get()), None);
                    return slot.take().nack(false).await;
                }
                Redelivery::Subscription => {}
            }
        }
        let accepted = slot.take().nack_after(delay).await;
        // The broker holds this redelivery on its own timer, out of the harness's sight, so the
        // harness is told when it falls due.
        #[cfg(feature = "testing")]
        if accepted.is_ok()
            && let Some(coordinator) = delivery.hooks.coordinator()
        {
            coordinator.expect_redelivery(delivery.subscription, delay);
        }
        return accepted;
    }

    let Some(retry) = delivery.retry.as_ref() else {
        // A delivery the declaration sends elsewhere is the broker's to move, and this requeue is
        // what moves it, so it was never due a delay. Only one that comes back loses its delay.
        if matches!(
            redelivery_of(slot.get(), &delivery.declaration),
            Redelivery::Subscription
        ) {
            warn!(
                target: "ruststream::dispatch",
                subscription = %name,
                "retry_after on a broker with neither native delayed redelivery nor a copy path \
                 of its own; requeuing immediately (the delay is dropped)",
            );
        }
        return slot.take().nack(true).await;
    };

    match redelivery_of(slot.get(), &delivery.declaration) {
        Redelivery::Subscription => {
            let destination = retry.destination.clone();
            publish_copy(
                slot,
                name,
                destination,
                Some(delay),
                retry,
                delivery,
                build_cx,
            )
            .await
        }
        Redelivery::DeadLetter {
            destination,
            at_cap,
        } => {
            if at_cap {
                warn_at_cap(name, attempt_of(slot.get()), Some(destination));
            }
            // A delivery that has run out of attempts is not retried, so the delay it asked for
            // does not apply to the copy that carries it away.
            let destination = Arc::from(destination);
            publish_copy(
                slot,
                name,
                Some(destination),
                None,
                retry,
                delivery,
                build_cx,
            )
            .await
        }
        Redelivery::Reject => {
            warn_at_cap(name, attempt_of(slot.get()), None);
            slot.take().nack(false).await
        }
    }
}

/// Says once per exhausted delivery where it went, because that is the point at which a message
/// stops coming back and an operator has to know.
fn warn_at_cap(name: &str, attempt: u64, destination: Option<&str>) {
    if let Some(destination) = destination {
        warn!(
            target: "ruststream::dispatch",
            subscription = %name,
            attempt,
            dead_letter = %destination,
            "retry attempts exhausted; publishing the delivery to the dead-letter destination",
        );
    } else {
        warn!(
            target: "ruststream::dispatch",
            subscription = %name,
            attempt,
            "retry attempts exhausted and no dead-letter destination is declared; rejecting the \
             delivery",
        );
    }
}

/// Drops the original delivery and sends a copy of it to `destination`, now or after `delay`; a
/// zero delay is now.
///
/// `destination` is `None` only where the registration's transforms name one per delivery; the
/// copy then starts at the subscription's own name, which is what the transforms read as the
/// delivery's channel, and one of them renames it.
///
/// # Errors
///
/// Returns the [`AckError`] from dropping the original where that failed and the transport does
/// settle: the delivery then stays with the broker, and a copy on top of it would duplicate the
/// message.
async fn publish_copy<M, C>(
    slot: &mut Slot<M>,
    name: &str,
    destination: Option<Arc<str>>,
    delay: Option<Duration>,
    retry: &DeferredRetry<C>,
    delivery: &Delivery<C>,
    build_cx: fn(&M) -> C,
) -> Result<(), AckError>
where
    M: IncomingMessage,
    C: Send + Sync + 'static,
{
    let publisher = Arc::clone(&retry.publisher);
    let subscription: Arc<str> = Arc::from(name);
    let target = destination.unwrap_or_else(|| Arc::clone(&subscription));

    // Settling consumes the message, so capture everything the copy needs first: its bytes, the
    // headers the handler saw, and the broker's own per-delivery context, which is what a
    // transform on this position reads. The bytes are lent on from here rather than handed over:
    // this path runs only after a delivery has failed, so a broker that keeps owned bytes copies
    // once there, where handing them over would cost the runtime a conversion on every retry.
    let msg = slot.get();
    let payload = Bytes::copy_from_slice(msg.payload());
    let delivered = msg.headers().clone();
    let mut headers = delivered.clone();
    let next_count = current_retry_count(&headers) + 1;
    headers.insert(Str::from_static(RETRY_COUNT_HEADER), next_count.to_string());
    let context = build_cx(msg);

    // Drop the original so the broker does not also redeliver it; the copy carries the retry
    // forward. A transport with no settlement at all has nothing to drop and no redelivery of its
    // own, so there the copy is the only way the message survives: aborting on that error would
    // lose it. Any other settle failure leaves the delivery with the broker, which will redeliver
    // it on its own timers, so a copy on top would duplicate it.
    //
    // Under the harness the original's settlement closes its count before the copy's publish
    // opens the copy's; the hold keeps the reaction open across that gap until this function
    // returns, so `drive` cannot settle between the two.
    #[cfg(feature = "testing")]
    let _hold = delivery.hooks.coordinator().map(Coordinator::hold);
    match slot.take().nack(false).await {
        Ok(()) | Err(AckError::Unsupported) => {}
        Err(err) => return Err(err),
    }

    let republish = async move {
        let cx = PublishContext::new(&subscription, &delivered, &context);
        let copy = OutgoingMessage::new(target.as_ref(), payload.as_ref()).with_headers(headers);
        if let Err(err) = publisher.publish_copy(copy, &cx).await {
            warn!(
                target: "ruststream::dispatch",
                subscription = %subscription,
                destination = %target,
                error = %err,
                "publishing the retry copy failed; message lost",
            );
        }
    };

    // The copy leaves after the handler returned, so the harness watching the delivery is
    // handed on to it: the copy is one of the delivery's publishes.
    #[cfg(feature = "testing")]
    let republish = in_harness_scope(harness_scope(delivery), republish);

    // A zero delay is an immediate copy: a timer task for it would cost a task allocation and
    // two scheduler turns to publish the same copy a little later.
    let Some(delay) = delay.filter(|delay| !delay.is_zero()) else {
        // An immediate copy is awaited on the dispatch path: there is no timer to wait for, and a
        // detached task would let the loop pull the next delivery before this one is back.
        republish.await;
        return Ok(());
    };

    // Under the harness the copy is scheduled through the coordinator, the way a broker schedules
    // its native delayed redelivery: `TestApp::advance` then fires it and counts it before it
    // drives the reaction, so a test sees the copy rather than a race.
    #[cfg(feature = "testing")]
    if let Some(coordinator) = delivery.hooks.coordinator() {
        coordinator.schedule_redelivery_future(delay, republish);
        return Ok(());
    }
    // The timer is the delivery's own work, so it runs where the delivery ran; being tracked, a
    // graceful shutdown waits for it and the copy leaves before the broker closes.
    delivery.spawn_after(async move {
        tokio::time::sleep(delay).await;
        republish.await;
    });
    Ok(())
}

#[cfg(all(test, feature = "testing", feature = "memory"))]
mod harness_tests;
mod pool;
#[cfg(all(test, feature = "memory"))]
mod tests;
mod threads;