datum/stream/

mod.rs

//! The linear Source/Flow/Sink DSL and the runtime that materializes it — Datum's
//! primary public surface, mirroring Akka/Pekko Streams Typed.
//!
//! This module root owns the shared vocabulary: the `BoxStream`/`PureTransform`/
//! `RuntimeTransform` type aliases, the `SourceHints`/`FlowHints` optimization-hint
//! system, the fan-in stream combinators (`merge_*`/`zip_*`), the spin-then-park
//! constants, and the re-export block. The public types live in the submodules
//! (`source`, `flow`, `sink`, `runtime`, `rate`, `time`, `restart`, `error`,
//! `completion`). See this directory's `AGENTS.md` for the map and the crate
//! `CLAUDE.md` for the blueprint-vs-materialization rule and execution model.

use std::{
    collections::{BTreeMap, HashMap, VecDeque},
    fmt,
    future::Future,
    hash::Hash,
    marker::PhantomData,
    panic::{AssertUnwindSafe, catch_unwind},
    pin::Pin,
    sync::{
        Arc, Condvar, Mutex, OnceLock,
        atomic::{AtomicBool, AtomicUsize, Ordering},
    },
    task::{Context, Poll},
    thread,
    time::Duration,
};

use futures::{channel::oneshot, executor::block_on};
use thiserror::Error;
use tokio::{
    runtime::{Builder as TokioRuntimeBuilder, Runtime as TokioRuntime},
    task::{JoinError, JoinHandle},
};

pub(crate) type BoxStream<T> = Box<dyn Iterator<Item = StreamResult<T>> + Send>;
pub(crate) type PureTransform<In, Out> = Arc<dyn Fn(BoxStream<In>) -> BoxStream<Out> + Send + Sync>;
pub(crate) type RuntimeTransform<In, Out> =
    Arc<dyn Fn(BoxStream<In>, &Materializer) -> StreamResult<BoxStream<Out>> + Send + Sync>;
type SinkRunner<In, Mat> = dyn Fn(BoxStream<In>, &Materializer) -> StreamResult<Mat> + Send + Sync;
type HintedSinkRunner<In, Mat> =
    dyn Fn(BoxStream<In>, &Materializer, SourceRuntimeHints) -> StreamResult<Mat> + Send + Sync;
type RunnableGraphRunner<Mat> = dyn Fn(&Materializer) -> StreamResult<Mat> + Send + Sync;
const STREAM_READY_SPINS: usize = 256;
/// Spin-loop hints between consecutive drain-thread readiness checks. Back off
/// a small block of hints between polls so the spin window still covers the
/// common fast-completion case without burning a full busy loop before
/// parking.
const STREAM_SPIN_BACKOFF: usize = 8;
const STREAM_MAX_PARK: Duration = Duration::from_millis(1);

/// Private metadata for known finite, synchronous micro-sources.
/// Set only on Datum-owned constructors whose `next()` is guaranteed
/// memory-backed and whose total success item count is known at blueprint time.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
struct InlineMicroSourceHint {
    max_success_items: usize,
}

#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
struct SourceHints {
    inline_head_terminal: bool,
    /// Set only for constructors where Datum owns the full iterator and can
    /// prove it is finite, synchronous, and memory-backed.
    inline_micro: Option<InlineMicroSourceHint>,
    /// Set only when terminal `fold`/`collect`/`ignore` may amortize the
    /// checked-stream cancellation/shutdown wrapper. Unknown, probe, queue, IO,
    /// and timer-backed sources keep the per-element checked path.
    terminal_consumer_batch: bool,
}

#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
pub(crate) struct SourceRuntimeHints {
    pub(crate) inline_micro_max_success_items: Option<usize>,
    pub(crate) terminal_consumer_batch: bool,
}

impl SourceHints {
    const fn with_inline_micro(max_success_items: usize) -> Self {
        Self {
            inline_head_terminal: true,
            inline_micro: Some(InlineMicroSourceHint { max_success_items }),
            terminal_consumer_batch: true,
        }
    }

    const fn with_terminal_consumer_batch() -> Self {
        Self {
            inline_head_terminal: false,
            inline_micro: None,
            terminal_consumer_batch: true,
        }
    }

    fn after_flow(self, flow: FlowHints) -> Self {
        if flow.preserves_inline_head_terminal {
            // inline_micro is intentionally not propagated through any flow:
            // even a trivial map wraps the iterator, making eligibility uncertain.
            Self {
                inline_head_terminal: true,
                inline_micro: None,
                terminal_consumer_batch: self.terminal_consumer_batch
                    && flow.preserves_terminal_consumer_batch,
            }
        } else {
            Self {
                inline_head_terminal: false,
                inline_micro: None,
                terminal_consumer_batch: self.terminal_consumer_batch
                    && flow.preserves_terminal_consumer_batch,
            }
        }
    }

    fn without_inline_micro(self) -> Self {
        Self {
            inline_head_terminal: self.inline_head_terminal,
            inline_micro: None,
            terminal_consumer_batch: self.terminal_consumer_batch,
        }
    }

    fn runtime(self) -> SourceRuntimeHints {
        SourceRuntimeHints {
            inline_micro_max_success_items: self.inline_micro.map(|hint| hint.max_success_items),
            terminal_consumer_batch: self.terminal_consumer_batch,
        }
    }
}

#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
struct FlowHints {
    preserves_inline_head_terminal: bool,
    preserves_terminal_consumer_batch: bool,
    scalar_chunk_prefix: bool,
}

impl FlowHints {
    const PRESERVES_INLINE_HEAD_TERMINAL: Self = Self {
        preserves_inline_head_terminal: true,
        preserves_terminal_consumer_batch: true,
        scalar_chunk_prefix: false,
    };

    const PRESERVES_TERMINAL_CONSUMER_BATCH: Self = Self {
        preserves_inline_head_terminal: false,
        preserves_terminal_consumer_batch: true,
        scalar_chunk_prefix: false,
    };

    fn then(self, next: Self) -> Self {
        Self {
            preserves_inline_head_terminal: self.preserves_inline_head_terminal
                && next.preserves_inline_head_terminal,
            preserves_terminal_consumer_batch: self.preserves_terminal_consumer_batch
                && next.preserves_terminal_consumer_batch,
            scalar_chunk_prefix: self.scalar_chunk_prefix && next.scalar_chunk_prefix,
        }
    }

    fn without_scalar_chunk_prefix(mut self) -> Self {
        self.scalar_chunk_prefix = false;
        self
    }
}

struct PartitionSlot<Key, Out> {
    key: Option<Key>,
    active: usize,
    queued: VecDeque<(usize, Out)>,
    in_ready_queue: bool,
}

struct AbortOnDropHandle<T> {
    handle: JoinHandle<T>,
}

impl<T> AbortOnDropHandle<T> {
    fn new(handle: JoinHandle<T>) -> Self {
        Self { handle }
    }
}

impl<T> Drop for AbortOnDropHandle<T> {
    fn drop(&mut self) {
        self.handle.abort();
    }
}

impl<T> Future for AbortOnDropHandle<T> {
    type Output = Result<T, JoinError>;

    fn poll(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
        Pin::new(&mut self.handle).poll(cx)
    }
}

impl<T> Unpin for AbortOnDropHandle<T> {}

pub(crate) fn stream_tokio_runtime() -> &'static TokioRuntime {
    static RUNTIME: OnceLock<TokioRuntime> = OnceLock::new();
    RUNTIME.get_or_init(|| {
        TokioRuntimeBuilder::new_multi_thread()
            .enable_all()
            .thread_name("datum-stream-tokio")
            .build()
            .expect("stream tokio runtime")
    })
}

fn spawn_tokio_task<Fut, T>(future: Fut) -> AbortOnDropHandle<T>
where
    Fut: Future<Output = T> + Send + 'static,
    T: Send + 'static,
{
    AbortOnDropHandle::new(stream_tokio_runtime().spawn(future))
}

pub(crate) fn current_stream_cancelled() -> Option<Arc<AtomicBool>> {
    runtime::current_stream_cancelled()
}

pub(super) fn catch_unwind_failed<T, F>(context: &'static str, f: F) -> StreamResult<T>
where
    F: FnOnce() -> T,
{
    catch_unwind(AssertUnwindSafe(f))
        .map_err(|_| StreamError::Failed(format!("{context} panicked")))
}

impl<Key, Out> PartitionSlot<Key, Out> {
    fn new(key: Key) -> Self {
        Self {
            key: Some(key),
            active: 0,
            queued: VecDeque::new(),
            in_ready_queue: false,
        }
    }
}

#[inline(always)]
fn partition_slot_for<Key, Out>(
    key: Key,
    slots_by_key: &mut HashMap<Key, usize>,
    slots: &mut Vec<PartitionSlot<Key, Out>>,
    free_slots: &mut Vec<usize>,
) -> usize
where
    Key: Clone + Eq + Hash,
{
    if let Some(slot) = slots_by_key.get(&key) {
        return *slot;
    }

    let slot = if let Some(slot) = free_slots.pop() {
        let state = &mut slots[slot];
        state.key = Some(key.clone());
        state.active = 0;
        state.queued.clear();
        state.in_ready_queue = false;
        slot
    } else {
        slots.push(PartitionSlot::new(key.clone()));
        slots.len() - 1
    };
    slots_by_key.insert(key, slot);
    slot
}

#[inline(always)]
fn retire_partition_slot<Key, Out>(
    slot: usize,
    slots_by_key: &mut HashMap<Key, usize>,
    slots: &mut [PartitionSlot<Key, Out>],
    free_slots: &mut Vec<usize>,
) where
    Key: Eq + Hash,
{
    let state = &mut slots[slot];
    if let Some(key) = state.key.take() {
        slots_by_key.remove(&key);
    }
    state.active = 0;
    state.queued.clear();
    state.in_ready_queue = false;
    free_slots.push(slot);
}

#[inline(always)]
fn ready_partition_slot<Key, Out>(
    slots: &mut [PartitionSlot<Key, Out>],
    ready_slots: &mut VecDeque<usize>,
    slot: usize,
    per_partition: usize,
) {
    if let Some(state) = slots.get_mut(slot)
        && state.key.is_some()
        && !state.in_ready_queue
        && state.active < per_partition
        && !state.queued.is_empty()
    {
        state.in_ready_queue = true;
        ready_slots.push_back(slot);
    }
}

#[inline(always)]
fn pop_ready_partition_slot<Key, Out>(
    slots: &mut [PartitionSlot<Key, Out>],
    ready_slots: &mut VecDeque<usize>,
    per_partition: usize,
) -> Option<(usize, usize, Out)> {
    while let Some(slot) = ready_slots.pop_front() {
        let mut requeue = false;
        let item = if let Some(state) = slots.get_mut(slot) {
            state.in_ready_queue = false;
            if state.key.is_some() && state.active < per_partition {
                let item = state.queued.pop_front().map(|(index, item)| {
                    state.active += 1;
                    (index, slot, item)
                });
                if !state.queued.is_empty() && state.active < per_partition {
                    state.in_ready_queue = true;
                    requeue = true;
                }
                item
            } else {
                None
            }
        } else {
            None
        };

        if requeue {
            ready_slots.push_back(slot);
        }
        if item.is_some() {
            return item;
        }
    }
    None
}

pub(crate) trait SourceFactory<Out, Mat>: Send + Sync {
    fn create(self: Arc<Self>, materializer: &Materializer) -> StreamResult<(BoxStream<Out>, Mat)>;

    fn append_scalar_i64(
        self: Arc<Self>,
        _steps: &[scalar::ScalarChunkStep<i64>],
    ) -> Option<Arc<dyn SourceFactory<Out, Mat>>> {
        None
    }
}

struct FnSourceFactory<F>(F);

impl<Out, Mat, F> SourceFactory<Out, Mat> for FnSourceFactory<F>
where
    F: Fn(&Materializer) -> StreamResult<(BoxStream<Out>, Mat)> + Send + Sync,
{
    fn create(self: Arc<Self>, materializer: &Materializer) -> StreamResult<(BoxStream<Out>, Mat)> {
        (self.0)(materializer)
    }
}

struct MapSourceFactory<In, Out, Mat, F> {
    source: Arc<dyn SourceFactory<In, Mat>>,
    stage: F,
    _marker: PhantomData<fn(In) -> Out>,
}

impl<In, Out, Mat, F> SourceFactory<Out, Mat> for MapSourceFactory<In, Out, Mat, F>
where
    In: Send + 'static,
    Out: Send + 'static,
    Mat: Send + 'static,
    F: Fn(In) -> Out + Send + Sync + 'static,
{
    fn create(self: Arc<Self>, materializer: &Materializer) -> StreamResult<(BoxStream<Out>, Mat)> {
        let (stream, mat) = Arc::clone(&self.source).create(materializer)?;
        Ok((
            Box::new(MapSourceStream {
                input: stream,
                factory: self,
            }),
            mat,
        ))
    }
}

struct MapSourceStream<In, Out, Mat, F> {
    input: BoxStream<In>,
    factory: Arc<MapSourceFactory<In, Out, Mat, F>>,
}

impl<In, Out, Mat, F> Iterator for MapSourceStream<In, Out, Mat, F>
where
    F: Fn(In) -> Out,
{
    type Item = StreamResult<Out>;

    fn next(&mut self) -> Option<Self::Item> {
        self.input
            .next()
            .map(|item| item.map(|item| (self.factory.stage)(item)))
    }
}

fn merge_streams<Out>(streams: Vec<BoxStream<Out>>, eager_complete: bool) -> BoxStream<Out>
where
    Out: Send + 'static,
{
    let mut streams: Vec<Option<BoxStream<Out>>> = streams.into_iter().map(Some).collect();
    let mut current = 0usize;
    Box::new(std::iter::from_fn(move || {
        loop {
            let index = next_active_optional_stream(&streams, current, |_| true)?;
            current = (index + 1) % streams.len().max(1);
            let Some(stream) = streams[index].as_mut() else {
                continue;
            };
            match stream.next() {
                Some(item) => return Some(item),
                None => {
                    streams[index] = None;
                    if eager_complete {
                        return None;
                    }
                }
            }
        }
    }))
}

fn merge_prioritized_streams<Out>(
    streams: Vec<BoxStream<Out>>,
    priorities: Vec<usize>,
    eager_complete: bool,
) -> BoxStream<Out>
where
    Out: Send + 'static,
{
    let mut streams: Vec<Option<BoxStream<Out>>> = streams.into_iter().map(Some).collect();
    let schedule: Vec<usize> = priorities
        .into_iter()
        .enumerate()
        .flat_map(|(index, weight)| std::iter::repeat_n(index, weight))
        .collect();
    let mut schedule_index = 0usize;
    Box::new(std::iter::from_fn(move || {
        loop {
            if streams.iter().all(Option::is_none) {
                return None;
            }
            let index = next_weighted_stream(&streams, &schedule, &mut schedule_index)?;
            let Some(stream) = streams[index].as_mut() else {
                continue;
            };
            match stream.next() {
                Some(item) => return Some(item),
                None => {
                    streams[index] = None;
                    if eager_complete {
                        return None;
                    }
                }
            }
        }
    }))
}

fn merge_sorted_stream<Out>(mut left: BoxStream<Out>, mut right: BoxStream<Out>) -> BoxStream<Out>
where
    Out: Ord + Send + 'static,
{
    let mut left_next: Option<Out> = None;
    let mut right_next: Option<Out> = None;
    let mut left_done = false;
    let mut right_done = false;
    Box::new(std::iter::from_fn(move || {
        loop {
            if left_next.is_none() && !left_done {
                match left.next() {
                    Some(Ok(item)) => left_next = Some(item),
                    Some(Err(error)) => return Some(Err(error)),
                    None => left_done = true,
                }
            }
            if right_next.is_none() && !right_done {
                match right.next() {
                    Some(Ok(item)) => right_next = Some(item),
                    Some(Err(error)) => return Some(Err(error)),
                    None => right_done = true,
                }
            }

            let next = match (&left_next, &right_next) {
                (Some(left_item), Some(right_item)) => {
                    if left_item <= right_item {
                        left_next.take()
                    } else {
                        right_next.take()
                    }
                }
                (Some(_), None) if right_done => left_next.take(),
                (None, Some(_)) if left_done => right_next.take(),
                (None, None) if left_done && right_done => return None,
                _ => continue,
            };
            if let Some(item) = next {
                return Some(Ok(item));
            }
        }
    }))
}

fn merge_latest_streams<Out>(
    streams: Vec<BoxStream<Out>>,
    eager_complete: bool,
) -> BoxStream<Vec<Out>>
where
    Out: Clone + Send + 'static,
{
    let mut streams: Vec<Option<BoxStream<Out>>> = streams.into_iter().map(Some).collect();
    let mut latest = vec![None; streams.len()];
    let mut seen = 0usize;
    let mut current = 0usize;
    let mut pending = VecDeque::<Vec<Out>>::new();
    Box::new(std::iter::from_fn(move || {
        loop {
            if let Some(output) = pending.pop_front() {
                return Some(Ok(output));
            }
            if streams.iter().all(Option::is_none) {
                return None;
            }
            let index = next_active_optional_stream(&streams, current, |_| true)?;
            current = (index + 1) % streams.len().max(1);
            let Some(stream) = streams[index].as_mut() else {
                continue;
            };
            match stream.next() {
                Some(Ok(item)) => {
                    if latest[index].is_none() {
                        seen += 1;
                    }
                    latest[index] = Some(item);
                    if seen == latest.len() {
                        pending.push_back(
                            latest
                                .iter()
                                .map(|item| item.clone().expect("merge-latest initialized"))
                                .collect(),
                        );
                    }
                }
                Some(Err(error)) => return Some(Err(error)),
                None => {
                    streams[index] = None;
                    if eager_complete {
                        return None;
                    }
                }
            }
        }
    }))
}

fn zip_streams<Left, Right>(
    mut left: BoxStream<Left>,
    mut right: BoxStream<Right>,
) -> BoxStream<(Left, Right)>
where
    Left: Send + 'static,
    Right: Send + 'static,
{
    let mut left_next: Option<Left> = None;
    let mut right_next: Option<Right> = None;
    let mut left_done = false;
    let mut right_done = false;
    Box::new(std::iter::from_fn(move || {
        loop {
            if left_next.is_none() && !left_done {
                match left.next() {
                    Some(Ok(item)) => left_next = Some(item),
                    Some(Err(error)) => return Some(Err(error)),
                    None => left_done = true,
                }
            }
            if right_next.is_none() && !right_done {
                match right.next() {
                    Some(Ok(item)) => right_next = Some(item),
                    Some(Err(error)) => return Some(Err(error)),
                    None => right_done = true,
                }
            }
            match (left_next.take(), right_next.take()) {
                (Some(left_item), Some(right_item)) => return Some(Ok((left_item, right_item))),
                (left_item, right_item) => {
                    left_next = left_item;
                    right_next = right_item;
                    if (left_done && left_next.is_none()) || (right_done && right_next.is_none()) {
                        return None;
                    }
                }
            }
        }
    }))
}

fn zip_latest_with_stream<Left, Right, Out, F>(
    mut left: BoxStream<Left>,
    mut right: BoxStream<Right>,
    eager_complete: bool,
    combine: Arc<F>,
) -> BoxStream<Out>
where
    Left: Clone + Send + 'static,
    Right: Clone + Send + 'static,
    Out: Send + 'static,
    F: Fn(Left, Right) -> Out + Send + Sync + 'static,
{
    let mut left_latest: Option<Left> = None;
    let mut right_latest: Option<Right> = None;
    let mut left_done = false;
    let mut right_done = false;
    let mut turn_left = true;
    let mut pending = VecDeque::<Out>::new();

    Box::new(std::iter::from_fn(move || {
        loop {
            if let Some(output) = pending.pop_front() {
                return Some(Ok(output));
            }
            if eager_complete && (left_done || right_done) {
                return None;
            }
            if left_done && right_done {
                return None;
            }
            // A side that completed without ever emitting can never be paired:
            // no further output is possible, so draining the other (possibly
            // infinite) side would loop forever producing nothing.
            if (left_done && left_latest.is_none()) || (right_done && right_latest.is_none()) {
                return None;
            }

            let pull_left = if left_done {
                false
            } else if right_done {
                true
            } else {
                let value = turn_left;
                turn_left = !turn_left;
                value
            };

            if pull_left {
                match left.next() {
                    Some(Ok(item)) => {
                        left_latest = Some(item);
                        if let (Some(left_item), Some(right_item)) = (&left_latest, &right_latest) {
                            pending.push_back(combine(left_item.clone(), right_item.clone()));
                        }
                    }
                    Some(Err(error)) => return Some(Err(error)),
                    None => {
                        left_done = true;
                        if eager_complete {
                            return None;
                        }
                    }
                }
            } else {
                match right.next() {
                    Some(Ok(item)) => {
                        right_latest = Some(item);
                        if let (Some(left_item), Some(right_item)) = (&left_latest, &right_latest) {
                            pending.push_back(combine(left_item.clone(), right_item.clone()));
                        }
                    }
                    Some(Err(error)) => return Some(Err(error)),
                    None => {
                        right_done = true;
                        if eager_complete {
                            return None;
                        }
                    }
                }
            }
        }
    }))
}

fn zip_all_stream<Left, Right>(
    mut left: BoxStream<Left>,
    mut right: BoxStream<Right>,
    left_fill: Left,
    right_fill: Right,
) -> BoxStream<(Left, Right)>
where
    Left: Clone + Send + 'static,
    Right: Clone + Send + 'static,
{
    let mut left_done = false;
    let mut right_done = false;
    Box::new(std::iter::from_fn(move || {
        if left_done && right_done {
            return None;
        }

        let left_item = if left_done {
            None
        } else {
            match left.next() {
                Some(Ok(item)) => Some(item),
                Some(Err(error)) => return Some(Err(error)),
                None => {
                    left_done = true;
                    None
                }
            }
        };
        let right_item = if right_done {
            None
        } else {
            match right.next() {
                Some(Ok(item)) => Some(item),
                Some(Err(error)) => return Some(Err(error)),
                None => {
                    right_done = true;
                    None
                }
            }
        };

        match (left_item, right_item) {
            (None, None) if left_done && right_done => None,
            (Some(left_value), Some(right_value)) => Some(Ok((left_value, right_value))),
            (Some(left_value), None) => Some(Ok((left_value, right_fill.clone()))),
            (None, Some(right_value)) => Some(Ok((left_fill.clone(), right_value))),
            (None, None) => None,
        }
    }))
}

fn zip_n_streams<Out, Next, F>(streams: Vec<BoxStream<Out>>, zipper: Arc<F>) -> BoxStream<Next>
where
    Out: Send + 'static,
    Next: Send + 'static,
    F: Fn(Vec<Out>) -> Next + Send + Sync + 'static,
{
    let count = streams.len();
    if count == 0 {
        return Box::new(std::iter::empty());
    }
    let mut streams: Vec<Option<BoxStream<Out>>> = streams.into_iter().map(Some).collect();
    let mut slots: Vec<Option<Out>> = (0..count).map(|_| None).collect();
    let mut current = 0usize;
    Box::new(std::iter::from_fn(move || {
        loop {
            if slots.iter().all(Option::is_some) {
                let values = slots
                    .iter_mut()
                    .map(|slot| slot.take().expect("zip-n slot filled"))
                    .collect();
                return Some(Ok(zipper(values)));
            }

            let index = next_active_optional_stream(&streams, current, |idx| slots[idx].is_none())?;
            current = (index + 1) % count.max(1);
            let Some(stream) = streams[index].as_mut() else {
                continue;
            };
            match stream.next() {
                Some(Ok(item)) => slots[index] = Some(item),
                Some(Err(error)) => return Some(Err(error)),
                None => {
                    streams[index] = None;
                    slots[index].as_ref()?;
                }
            }
        }
    }))
}

fn next_active_optional_stream<T, F>(
    streams: &[Option<BoxStream<T>>],
    current: usize,
    predicate: F,
) -> Option<usize>
where
    T: Send + 'static,
    F: Fn(usize) -> bool,
{
    if streams.is_empty() {
        return None;
    }
    for offset in 0..streams.len() {
        let index = (current + offset) % streams.len();
        if streams[index].is_some() && predicate(index) {
            return Some(index);
        }
    }
    None
}

fn next_weighted_stream<T>(
    streams: &[Option<BoxStream<T>>],
    schedule: &[usize],
    schedule_index: &mut usize,
) -> Option<usize>
where
    T: Send + 'static,
{
    if streams.is_empty() || schedule.is_empty() {
        return None;
    }
    for _ in 0..schedule.len() {
        let index = schedule[*schedule_index % schedule.len()];
        *schedule_index = (*schedule_index + 1) % schedule.len();
        if streams.get(index).is_some_and(Option::is_some) {
            return Some(index);
        }
    }
    None
}

pub(crate) mod async_boundary;
mod completion;
mod error;
mod flow;
mod rate;
mod restart;
mod runtime;
pub(crate) mod scalar;
mod sink;
mod source;
mod time;
mod timer;

/// Opaque hook on `Source<T>` for split-segment fast-path sources.
/// Non-`None` only on `Source`s returned by the split fast path.
/// Cleared by all composition operators (via/map/to_mat etc.).
pub(crate) trait SplitSegmentHookDyn: Send + Sync + 'static {
    fn as_any_arc(self: Arc<Self>) -> Arc<dyn std::any::Any + Send + Sync>;
}

/// Opaque hook on `Source<T>` for a private source-to-terminal fast path.
///
/// Implementations synchronously fill bounded batches on the sink worker
/// spawned by the terminal sink descriptor. Sink descriptors process the batch
/// after the hook returns, so user closures do not run under source locks.
pub(crate) trait TerminalSourceHookDyn<In>: Send + Sync + 'static {
    fn drain_terminal_batch(
        &self,
        materializer: &Materializer,
        cancelled: &Arc<AtomicBool>,
        batch: &mut Vec<In>,
    ) -> StreamResult<TerminalSourceStatus>;

    fn supports_direct_terminal(&self) -> bool {
        false
    }

    fn try_register_direct_terminal(
        &self,
        _consumer: Box<dyn TerminalSinkConsumerDyn<In>>,
        _cancelled: Arc<AtomicBool>,
    ) -> Option<StreamResult<()>> {
        None
    }

    fn cancel_terminal(&self) {}
}

#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub(crate) enum TerminalSourceStatus {
    Active,
    Completed,
}

pub(crate) trait TerminalSinkConsumerDyn<In>: Send + 'static {
    fn on_item(&mut self, item: In) -> StreamResult<()>;
    fn finish(self: Box<Self>, result: StreamResult<()>);
}

/// Fast-path descriptor for fold/collect/ignore sinks in split-segment context.
/// Non-`None` only for recognised sinks (`Sink::fold`, `fold_result`, `collect`, `ignore`).
pub(crate) trait FoldFastPathDyn<In: Send + 'static>: Send + Sync + 'static {
    /// Try to register directly with a split segment hook.
    /// Returns `None` if the hook type is not recognised.
    /// Returns `Some(Ok(mat))` on success where `mat` is `Box<StreamCompletion<Acc>>`.
    fn try_register(
        &self,
        hook: Arc<dyn SplitSegmentHookDyn>,
    ) -> Option<StreamResult<Box<dyn std::any::Any + Send>>>;

    fn supports_terminal_drain(&self) -> bool {
        false
    }

    fn try_register_direct_terminal(
        &self,
        _hook: Arc<dyn TerminalSourceHookDyn<In>>,
        _materializer: &Materializer,
    ) -> Option<StreamResult<Box<dyn std::any::Any + Send>>> {
        None
    }

    fn try_register_terminal_drain(
        &self,
        _hook: Arc<dyn TerminalSourceHookDyn<In>>,
        _materializer: &Materializer,
    ) -> Option<StreamResult<Box<dyn std::any::Any + Send>>> {
        None
    }
}

use self::runtime::{runtime_checked_stream, set_current_stream_cancelled};

pub(crate) use self::completion::StreamCancellation;

pub use self::{
    completion::{Cancellable, StreamCompletion},
    error::{StreamError, StreamResult, Supervision, SupervisionDecider, SupervisionDirective},
    flow::{BidiFlow, Flow},
    rate::{AggregateTimer, OverflowStrategy},
    restart::{RestartFlow, RestartSettings, RestartSink, RestartSource, RetryFlow},
    runtime::{Materializer, Runtime},
    scalar::{ScalarArithmeticOp, ScalarCompareOp},
    sink::{RunnableGraph, Sink, SinkCombineStrategy},
    source::{
        Demand, IntoSource, Keep, MaybeHandle, NotUsed, PushOutlet, Source, SourceCombineStrategy,
    },
    time::{DelayOverflowStrategy, ThrottleMode},
};

#[cfg(test)]
mod tests {
    use super::*;
    use crate::Attributes;
    use crate::testkit::TestSink;
    use std::fs;
    use std::sync::{
        Arc as StdArc,
        atomic::{
            AtomicBool as StdAtomicBool, AtomicUsize as StdAtomicUsize, Ordering as StdOrdering,
        },
        mpsc,
    };
    use std::time::Duration as StdDuration;
    use std::time::Instant;

    fn wait<T>(completion: StreamCompletion<T>) -> T {
        completion.wait().unwrap()
    }

    fn wait_until(timeout: Duration, mut condition: impl FnMut() -> bool) -> bool {
        let deadline = Instant::now() + timeout;
        while Instant::now() < deadline {
            if condition() {
                return true;
            }
            thread::sleep(Duration::from_millis(2));
        }
        condition()
    }

    fn linux_thread_count(thread_name: &str) -> usize {
        fs::read_dir("/proc/self/task")
            .expect("task directory readable")
            .filter_map(Result::ok)
            .filter_map(|entry| fs::read_to_string(entry.path().join("comm")).ok())
            .filter(|name| name.trim() == thread_name)
            .count()
    }

    #[test]
    fn source_run_terminal_shortcuts_match_explicit_sinks() {
        let explicit_fold: StreamResult<StreamCompletion<u64>> =
            Source::from_iter(1_u64..=4).run_with(Sink::fold(0_u64, |acc, item| acc + item));
        let sugared_fold: StreamResult<StreamCompletion<u64>> =
            Source::from_iter(1_u64..=4).run_fold(0_u64, |acc, item| acc + item);
        assert_eq!(wait(explicit_fold.unwrap()), 10);
        assert_eq!(wait(sugared_fold.unwrap()), 10);

        let explicit_reduce: StreamResult<StreamCompletion<u64>> =
            Source::from_iter(1_u64..=4).run_with(Sink::reduce(|left: u64, right| left + right));
        let sugared_reduce: StreamResult<StreamCompletion<u64>> =
            Source::from_iter(1_u64..=4).run_reduce(|left, right| left + right);
        assert_eq!(wait(explicit_reduce.unwrap()), 10);
        assert_eq!(wait(sugared_reduce.unwrap()), 10);

        let explicit_empty_reduce = Source::<u64>::empty()
            .run_with(Sink::reduce(|left: u64, right| left + right))
            .unwrap()
            .wait();
        let sugared_empty_reduce = Source::<u64>::empty()
            .run_reduce(|left, right| left + right)
            .unwrap()
            .wait();
        assert_eq!(sugared_empty_reduce, explicit_empty_reduce);
        assert_eq!(sugared_empty_reduce, Err(StreamError::EmptyStream));

        let explicit_sum = StdArc::new(StdAtomicUsize::new(0));
        let explicit_sum_sink = StdArc::clone(&explicit_sum);
        let explicit_foreach: StreamResult<StreamCompletion<NotUsed>> =
            Source::from_iter(1_usize..=4).run_with(Sink::foreach(move |item| {
                explicit_sum_sink.fetch_add(item, StdOrdering::SeqCst);
            }));
        assert_eq!(wait(explicit_foreach.unwrap()), NotUsed);

        let sugared_sum = StdArc::new(StdAtomicUsize::new(0));
        let sugared_sum_sink = StdArc::clone(&sugared_sum);
        let sugared_foreach: StreamResult<StreamCompletion<NotUsed>> =
            Source::from_iter(1_usize..=4).run_foreach(move |item| {
                sugared_sum_sink.fetch_add(item, StdOrdering::SeqCst);
            });
        assert_eq!(wait(sugared_foreach.unwrap()), NotUsed);

        let alias_sum = StdArc::new(StdAtomicUsize::new(0));
        let alias_sum_sink = StdArc::clone(&alias_sum);
        let alias_for_each: StreamResult<StreamCompletion<NotUsed>> =
            Source::from_iter(1_usize..=4).run_for_each(move |item| {
                alias_sum_sink.fetch_add(item, StdOrdering::SeqCst);
            });
        assert_eq!(wait(alias_for_each.unwrap()), NotUsed);

        assert_eq!(explicit_sum.load(StdOrdering::SeqCst), 10);
        assert_eq!(sugared_sum.load(StdOrdering::SeqCst), 10);
        assert_eq!(alias_sum.load(StdOrdering::SeqCst), 10);
    }

    #[test]
    fn source_constructor_sugar_matches_from_iter() {
        let expected = Source::from_iter(1_u64..=4).run_collect().unwrap();

        assert_eq!(
            Source::from(vec![1_u64, 2, 3, 4]).run_collect().unwrap(),
            expected
        );
        assert_eq!(
            Source::from([1_u64, 2, 3, 4]).run_collect().unwrap(),
            expected
        );
        assert_eq!((1_u64..=4).into_source().run_collect().unwrap(), expected);
    }

    #[test]
    fn source_constructor_sugar_keeps_existing_inference_paths() {
        let from_vec_into: Source<u64> = vec![1, 2, 3].into();
        let from_array_into: Source<u64> = [1, 2, 3].into();
        let from_iter = Source::from_iter(1_u64..=3);
        let from_iterable = Source::from_iterable(1_u64..=3);
        let from_iterator: Source<u64> = (1_u64..=3).collect();
        let from_range_into_source: Source<u64> = (1_u64..=3).into_source();

        let expected = vec![1, 2, 3];
        assert_eq!(from_vec_into.run_collect().unwrap(), expected);
        assert_eq!(from_array_into.run_collect().unwrap(), expected);
        assert_eq!(from_iter.run_collect().unwrap(), expected);
        assert_eq!(from_iterable.run_collect().unwrap(), expected);
        assert_eq!(from_iterator.run_collect().unwrap(), expected);
        assert_eq!(from_range_into_source.run_collect().unwrap(), expected);
    }

    #[test]
    fn source_async_boundary_preserves_results() {
        let expected = Source::from_iter(0_u64..128)
            .map(|item| item.wrapping_add(1))
            .filter(|item| item % 3 != 0)
            .map(|item| item * 2)
            .run_collect()
            .unwrap();

        let actual = Source::from_iter(0_u64..128)
            .map(|item| item.wrapping_add(1))
            .async_boundary()
            .filter(|item| item % 3 != 0)
            .map(|item| item * 2)
            .run_collect()
            .unwrap();

        assert_eq!(actual, expected);
    }

    #[test]
    fn flow_async_boundary_preserves_results() {
        let expected = Source::from_iter(0_u64..128)
            .map(|item| item + 1)
            .map(|item| item * 3)
            .run_collect()
            .unwrap();

        let flow = Flow::identity()
            .map(|item: u64| item + 1)
            .r#async()
            .map(|item| item * 3);
        let actual = Source::from_iter(0_u64..128)
            .via(flow)
            .run_collect()
            .unwrap();

        assert_eq!(actual, expected);
    }

    #[test]
    fn linear_async_boundary_matches_graph_async_boundary_shape() {
        use crate::{
            AsyncBoundary, AsyncBoundaryExecutionConfig, FusedExecutionConfig, GraphDsl,
            GraphFlowShape, MapStage,
        };

        let graph = GraphDsl::try_create(|builder| {
            let first = builder.add(MapStage::new(|item: u64| item + 1));
            let boundary = builder.add(AsyncBoundary::<u64>::new());
            let second = builder.add(MapStage::new(|item: u64| item * 2));

            builder.connect(first.outlet(), boundary.inlet())?;
            builder.connect(boundary.outlet(), second.inlet())?;

            Ok(GraphFlowShape::new(first.inlet(), second.outlet()))
        })
        .unwrap();

        let linear = Source::from_iter(1_u64..=4)
            .map(|item| item + 1)
            .async_boundary_with_buffer(4)
            .map(|item| item * 2)
            .run_collect()
            .unwrap();
        let graph_output = graph.run_with_input(1_u64..=4).unwrap();
        let report = graph
            .run_async_boundary_count_with_input_report(
                1_u64..=4,
                AsyncBoundaryExecutionConfig {
                    fused: FusedExecutionConfig { event_limit: 1024 },
                    buffer_size: 4,
                },
            )
            .unwrap();

        assert_eq!(linear, graph_output);
        assert_eq!(report.result, linear.len());
        assert_eq!(report.async_boundary_crossings, linear.len());
    }

    #[test]
    fn async_boundary_regions_run_concurrently() {
        let (upstream_tx, upstream_rx) = mpsc::channel::<u64>();
        let (downstream_blocked_tx, downstream_blocked_rx) = mpsc::channel::<()>();
        let (release_tx, release_rx) = mpsc::channel::<()>();
        let release_rx = StdArc::new(Mutex::new(release_rx));

        let completion = Source::from_iter(0_u64..3)
            .map(move |item| {
                upstream_tx.send(item).expect("upstream probe receives");
                item
            })
            .async_boundary_with_buffer(1)
            .map({
                let release_rx = StdArc::clone(&release_rx);
                move |item| {
                    if item == 0 {
                        downstream_blocked_tx
                            .send(())
                            .expect("downstream probe receives");
                        release_rx
                            .lock()
                            .expect("release receiver lock")
                            .recv_timeout(StdDuration::from_secs(2))
                            .expect("downstream release arrives");
                    }
                    item
                }
            })
            .run_with(Sink::collect())
            .unwrap();

        assert_eq!(
            downstream_blocked_rx.recv_timeout(StdDuration::from_secs(2)),
            Ok(())
        );
        assert_eq!(upstream_rx.recv_timeout(StdDuration::from_secs(2)), Ok(0));
        assert_eq!(upstream_rx.recv_timeout(StdDuration::from_secs(2)), Ok(1));

        release_tx.send(()).expect("release downstream");
        assert_eq!(completion.wait().unwrap(), vec![0, 1, 2]);
    }

    #[test]
    fn async_boundary_backpressures_slow_downstream() {
        let (produced_tx, produced_rx) = mpsc::channel::<u64>();
        let (release_tx, release_rx) = mpsc::channel::<()>();
        let release_rx = StdArc::new(Mutex::new(release_rx));

        let completion = Source::from_iter(0_u64..8)
            .map(move |item| {
                produced_tx.send(item).expect("producer probe receives");
                item
            })
            .async_boundary_with_buffer(1)
            .map({
                let release_rx = StdArc::clone(&release_rx);
                move |item| {
                    if item == 0 {
                        release_rx
                            .lock()
                            .expect("release receiver lock")
                            .recv_timeout(StdDuration::from_secs(2))
                            .expect("downstream release arrives");
                    }
                    item
                }
            })
            .run_with(Sink::collect())
            .unwrap();

        assert_eq!(produced_rx.recv_timeout(StdDuration::from_secs(2)), Ok(0));
        assert_eq!(produced_rx.recv_timeout(StdDuration::from_secs(2)), Ok(1));
        if let Ok(item) = produced_rx.recv_timeout(StdDuration::from_millis(100)) {
            assert_eq!(item, 2);
        }
        match produced_rx.recv_timeout(StdDuration::from_millis(100)) {
            Err(mpsc::RecvTimeoutError::Timeout) => {}
            other => panic!("async boundary handoff was not bounded: {other:?}"),
        }

        release_tx.send(()).expect("release downstream");
        assert_eq!(completion.wait().unwrap(), (0_u64..8).collect::<Vec<_>>());
    }

    #[test]
    fn source_blueprints_are_reusable() {
        let source = Source::from_iter(0..5).map(|item| item + 1);

        assert_eq!(source.clone().run_collect().unwrap(), vec![1, 2, 3, 4, 5]);
        assert_eq!(source.run_collect().unwrap(), vec![1, 2, 3, 4, 5]);
    }

    #[test]
    fn source_map_preserves_materialized_value() {
        let graph = Source::single(1)
            .map_materialized_value(|_| "source")
            .map(|item| item + 1)
            .to_mat(Sink::head(), Keep::both);

        let materialized = graph.run().unwrap();
        assert_eq!(materialized.0, "source");
        assert_eq!(wait(materialized.1), 2);
    }

    #[test]
    fn source_and_flow_compose() {
        let flow = Flow::identity()
            .map(|item: i32| item * 2)
            .filter(|item| item % 3 == 0);

        let result = Source::from_iter(0..8).via(flow).run_collect().unwrap();

        assert_eq!(result, vec![0, 6, 12]);
    }

    #[test]
    fn sink_setup_sees_materializer_defaults_and_local_attributes() {
        let observed = StdArc::new(Mutex::new(None));
        let observed_in_setup = StdArc::clone(&observed);
        let sink = Sink::<i32, StreamCompletion<NotUsed>>::setup(move |_materializer, attrs| {
            *observed_in_setup.lock().unwrap() = Some((
                attrs.name().map(str::to_owned),
                attrs.input_buffer_hint(),
                attrs.dispatcher_hint().map(str::to_owned),
            ));
            Sink::ignore()
        })
        .add_attributes(Attributes::named("sink-inner"))
        .add_attributes(Attributes::input_buffer(4, 4))
        .add_attributes(Attributes::dispatcher("bench-dispatcher"));

        let materializer = Materializer::new().with_attributes(Attributes::named("mat-outer"));
        wait(
            Source::from_iter([1, 2, 3])
                .run_with_materializer(sink, &materializer)
                .unwrap(),
        );

        assert_eq!(
            *observed.lock().unwrap(),
            Some((
                Some("sink-inner".to_owned()),
                Some((4, 4)),
                Some("bench-dispatcher".to_owned())
            ))
        );
    }

    #[test]
    fn sink_pre_materialize_feeds_existing_materialization() {
        let materializer = Materializer::new();
        let (completion, pre) = Sink::<i32, StreamCompletion<Vec<i32>>>::collect()
            .pre_materialize(&materializer)
            .unwrap();

        Source::from_iter([1, 2, 3])
            .run_with_materializer(pre, &materializer)
            .unwrap();

        assert_eq!(wait(completion), vec![1, 2, 3]);
    }

    #[test]
    fn flow_from_sink_and_source_connects_both_sides() {
        assert_eq!(
            Source::from_iter([1, 2, 3])
                .via(Flow::from_sink_and_source(
                    Sink::foreach(|_item: i32| {}),
                    Source::from_iter([10, 20, 30]),
                ))
                .run_collect()
                .unwrap(),
            vec![10, 20, 30]
        );
    }

    #[test]
    fn from_sink_and_source_keeps_sink_running_after_source_side_completes() {
        let completed = StdArc::new(StdAtomicBool::new(false));
        let on_complete = StdArc::clone(&completed);
        let flow = Flow::from_sink_and_source(
            Sink::on_complete(move || {
                on_complete.store(true, StdOrdering::SeqCst);
            }),
            Source::single(10),
        );

        let result = Source::from_iter([1, 2, 3])
            .via(flow)
            .run_collect()
            .unwrap();

        assert_eq!(result, vec![10]);
        assert!(wait_until(StdDuration::from_secs(1), || {
            completed.load(StdOrdering::SeqCst)
        }));
    }

    #[test]
    fn from_sink_and_source_coupled_cancels_source_when_sink_finishes_first() {
        let cancellable = StdArc::new(Mutex::new(None));
        let observed = StdArc::clone(&cancellable);
        let flow = Flow::from_sink_and_source_coupled(
            Sink::ignore(),
            Source::tick(
                StdDuration::from_millis(50),
                StdDuration::from_millis(50),
                10,
            )
            .map_materialized_value(move |handle| {
                *observed.lock().unwrap() = Some(handle.clone());
                handle
            }),
        );

        let completion = Source::from_iter(std::iter::empty::<i32>())
            .via(flow)
            .run_with(Sink::ignore())
            .unwrap();
        assert!(wait_until(StdDuration::from_secs(1), || {
            cancellable
                .lock()
                .unwrap()
                .as_ref()
                .is_some_and(Cancellable::is_cancelled)
        }));
        assert_eq!(wait(completion), NotUsed);
    }

    #[test]
    fn bidi_flow_join_and_atop_compose() {
        let codec = BidiFlow::from_flows(
            Flow::identity().map(|item: i32| item + 1),
            Flow::identity().map(|item: i32| item * 2),
        )
        .named("codec");
        let framing = BidiFlow::from_flows(
            Flow::identity().map(|item: i32| item * 3),
            Flow::identity().map(|item: i32| item - 4),
        );

        let joined = codec
            .clone()
            .join(Flow::identity().map(|item: i32| item - 5));
        let stacked = codec.atop(framing).join(Flow::identity());

        assert_eq!(
            Source::single(10).via(joined).run_collect().unwrap(),
            vec![12]
        );
        assert_eq!(
            Source::single(10).via(stacked).run_collect().unwrap(),
            vec![58]
        );
    }

    #[test]
    fn flow_buffer_then_map_runs_end_to_end() {
        let flow = Flow::identity()
            .buffer(8, OverflowStrategy::Backpressure)
            .map(|item: i32| item + 1);

        let result = Source::from_iter(0..4).via(flow).run_collect().unwrap();

        assert_eq!(result, vec![1, 2, 3, 4]);
    }

    #[test]
    fn public_flow_combinators_preserve_runtime_transform_after_buffer() {
        fn buffered_flow() -> Flow<i32, i32> {
            Flow::identity().buffer(8, OverflowStrategy::Backpressure)
        }

        assert_eq!(
            Source::from_iter(0..4)
                .via(buffered_flow().filter(|item| *item % 2 == 0))
                .run_collect()
                .unwrap(),
            vec![0, 2]
        );
        assert_eq!(
            Source::from_iter(0..4)
                .via(buffered_flow().filter_not(|item| *item % 2 == 0))
                .run_collect()
                .unwrap(),
            vec![1, 3]
        );
        assert_eq!(
            Source::from_iter(0..4)
                .via(buffered_flow().filter_map(|item| (item % 2 == 0).then_some(item + 10)))
                .run_collect()
                .unwrap(),
            vec![10, 12]
        );
        assert_eq!(
            Source::from_iter(0..3)
                .via(buffered_flow().map_concat(|item| [item, item + 10]))
                .run_collect()
                .unwrap(),
            vec![0, 10, 1, 11, 2, 12]
        );
        assert_eq!(
            Source::from_iter(0..3)
                .via(buffered_flow().stateful_map(5, |state, item| {
                    *state += item;
                    *state
                }))
                .run_collect()
                .unwrap(),
            vec![5, 6, 8]
        );
        assert_eq!(
            Source::from_iter(0..3)
                .via(buffered_flow().stateful_map_concat(0, |state, item| {
                    *state += item;
                    [*state, item]
                }))
                .run_collect()
                .unwrap(),
            vec![0, 0, 1, 1, 3, 2]
        );
        assert_eq!(
            Source::from_iter(0..4)
                .via(buffered_flow().map_async(2, |item| async move { Ok(item + 1) }))
                .run_collect()
                .unwrap(),
            vec![1, 2, 3, 4]
        );
        assert_eq!(
            Source::from_iter(0..4)
                .via(buffered_flow().map_async_unordered(2, |item| async move { Ok(item + 1) }))
                .run_collect()
                .unwrap(),
            vec![1, 2, 3, 4]
        );
        assert_eq!(
            Source::from_iter(0..4)
                .via(buffered_flow().map_async_partitioned(
                    2,
                    1,
                    |item| item % 2,
                    |item| async move { Ok(item + 1) },
                ))
                .run_collect()
                .unwrap(),
            vec![1, 2, 3, 4]
        );
        assert_eq!(
            Source::from_iter(0..5)
                .via(buffered_flow().take(3))
                .run_collect()
                .unwrap(),
            vec![0, 1, 2]
        );
        assert_eq!(
            Source::from_iter(0..5)
                .via(buffered_flow().drop(2))
                .run_collect()
                .unwrap(),
            vec![2, 3, 4]
        );
        assert_eq!(
            Source::from_iter(0..5)
                .via(buffered_flow().take_while(|item| *item < 3))
                .run_collect()
                .unwrap(),
            vec![0, 1, 2]
        );
        assert_eq!(
            Source::from_iter(0..5)
                .via(buffered_flow().drop_while(|item| *item < 3))
                .run_collect()
                .unwrap(),
            vec![3, 4]
        );
        assert_eq!(
            Source::from_iter(0..3)
                .via(buffered_flow().limit(5))
                .run_collect()
                .unwrap(),
            vec![0, 1, 2]
        );
        assert_eq!(
            Source::from_iter(0..5)
                .via(buffered_flow().grouped(2))
                .run_collect()
                .unwrap(),
            vec![vec![0, 1], vec![2, 3], vec![4]]
        );
        assert_eq!(
            Source::from_iter(1..=3)
                .via(buffered_flow().scan(0, |acc, item| acc + item))
                .run_collect()
                .unwrap(),
            vec![0, 1, 3, 6]
        );
        assert_eq!(
            Source::from_iter(1..=4)
                .via(buffered_flow().sliding(2, 1))
                .run_collect()
                .unwrap(),
            vec![vec![1, 2], vec![2, 3], vec![3, 4]]
        );
        assert_eq!(
            Source::from_iter(1..=4)
                .via(buffered_flow().fold(0, |acc, item| acc + item))
                .run_collect()
                .unwrap(),
            vec![10]
        );
        assert_eq!(
            Source::from_iter(1..=4)
                .via(buffered_flow().reduce(|acc, item| acc + item))
                .run_collect()
                .unwrap(),
            vec![10]
        );
        assert_eq!(
            Source::from_factory(|| {
                Box::new(vec![Ok(1), Err(StreamError::Failed("boom".into())), Ok(2)].into_iter())
            })
            .via(buffered_flow().map_error(|_| StreamError::Failed("mapped".into())))
            .run_collect(),
            Err(StreamError::Failed("mapped".into()))
        );
        assert_eq!(
            Source::<i32>::failed(StreamError::Failed("boom".into()))
                .via(buffered_flow().recover(|_| Some(42)))
                .run_collect()
                .unwrap(),
            vec![42]
        );
        assert_eq!(
            Source::<i32>::failed(StreamError::Failed("boom".into()))
                .via(buffered_flow().recover_with(|_| Some(Source::from_iter([7, 8]))))
                .run_collect()
                .unwrap(),
            vec![7, 8]
        );
        assert_eq!(
            Source::<i32>::failed(StreamError::Failed("boom".into()))
                .via(buffered_flow().recover_with_retries(1, |_| Some(Source::from_iter([9]))))
                .run_collect()
                .unwrap(),
            vec![9]
        );
        assert_eq!(
            Source::from_factory(|| {
                Box::new(vec![Ok(1), Err(StreamError::Failed("ignored".into())), Ok(2)].into_iter())
            })
            .via(buffered_flow().on_error_complete())
            .run_collect()
            .unwrap(),
            vec![1]
        );

        let materialized = Source::from_iter([1, 2, 3])
            .run_with(
                buffered_flow()
                    .via(Flow::identity().map(|item| item + 1))
                    .map_materialized_value(|_| "buffered-flow")
                    .to_mat(Sink::fold(0, |acc, item| acc + item), Keep::both),
            )
            .unwrap();
        assert_eq!(materialized.0, "buffered-flow");
        assert_eq!(wait(materialized.1), 9);

        let kept = Source::from_iter([1, 2, 3])
            .run_with(
                buffered_flow()
                    .via_mat_with(Flow::identity().map(|item| item + 1), |_, _| "combined")
                    .to(Sink::fold(0, |acc, item| acc + item)),
            )
            .unwrap();
        assert_eq!(kept, "combined");
    }

    #[test]
    fn runtime_rate_flows_compose_in_flow_form() {
        let conflate = Flow::identity()
            .conflate(|left: i32, right| left + right)
            .map(|item| item + 1);
        assert_eq!(
            Source::single(4).via(conflate).run_collect().unwrap(),
            vec![5]
        );

        let batch = Flow::identity()
            .batch(4, |item: i32| item, |left, right| left + right)
            .map(|item| item + 1);
        assert_eq!(Source::single(4).via(batch).run_collect().unwrap(), vec![5]);

        let expand = Flow::identity()
            .expand(std::iter::once::<i32>)
            .map(|item| item + 1);
        assert_eq!(
            Source::from_iter(0..4).via(expand).run_collect().unwrap(),
            vec![1, 2, 3, 4]
        );

        let aggregate = Flow::identity()
            .aggregate_with_boundary(
                Vec::<i32>::new,
                |mut items, item| {
                    items.push(item);
                    let ready = !items.is_empty();
                    (items, ready)
                },
                |items| items.into_iter().sum::<i32>(),
                None,
            )
            .map(|item| item + 1);
        assert_eq!(
            Source::from_iter(0..4)
                .via(aggregate)
                .run_collect()
                .unwrap(),
            vec![1, 2, 3, 4]
        );

        let detached = Flow::identity().detach().map(|item: i32| item + 1);
        assert_eq!(
            Source::from_iter(0..4).via(detached).run_collect().unwrap(),
            vec![1, 2, 3, 4]
        );
    }

    #[test]
    fn high_use_source_flow_operators_work() {
        let result = Source::from_iter(0..8)
            .drop(1)
            .take(5)
            .filter_not(|item| item % 2 == 0)
            .map_concat(|item| [item, item + 10])
            .grouped(3)
            .run_collect()
            .unwrap();

        assert_eq!(result, vec![vec![1, 11, 3], vec![13, 5, 15]]);
    }

    #[test]
    fn prefix_and_tail_emits_prefix_and_live_tail() {
        let mut outer = Source::from_iter(0..5)
            .prefix_and_tail(2)
            .run_collect()
            .unwrap();
        assert_eq!(outer.len(), 1);
        let (prefix, tail) = outer.pop().unwrap();
        assert_eq!(prefix, vec![0, 1]);
        assert_eq!(tail.clone().run_collect().unwrap(), vec![2, 3, 4]);
        assert_eq!(
            tail.run_collect(),
            Err(StreamError::Failed(
                "substream source cannot be materialized more than once".into()
            ))
        );
    }

    #[test]
    fn prefix_and_tail_fails_before_prefix_is_ready() {
        let result = Source::from_factory(|| {
            Box::new(vec![Ok(1), Err(StreamError::Failed("boom".into())), Ok(2)].into_iter())
        })
        .prefix_and_tail(2)
        .run_collect();
        assert!(matches!(result, Err(StreamError::Failed(message)) if message == "boom"));
    }

    #[test]
    fn prefix_and_tail_tail_propagates_late_upstream_failure() {
        let mut outer = Source::from_factory(|| {
            Box::new(vec![Ok(1), Ok(2), Err(StreamError::Failed("boom".into())), Ok(3)].into_iter())
        })
        .prefix_and_tail(2)
        .run_collect()
        .unwrap();
        let (prefix, tail) = outer.pop().unwrap();
        assert_eq!(prefix, vec![1, 2]);
        assert_eq!(tail.run_collect(), Err(StreamError::Failed("boom".into())));
    }

    #[test]
    fn prefix_and_tail_accepts_non_clone_elements() {
        #[derive(Debug, PartialEq, Eq)]
        struct NonClone(u8);

        let mut outer = Source::from_factory(|| {
            Box::new(vec![Ok(NonClone(1)), Ok(NonClone(2)), Ok(NonClone(3))].into_iter())
        })
        .prefix_and_tail(2)
        .run_collect()
        .unwrap();
        let (prefix, tail) = outer.pop().unwrap();
        assert_eq!(prefix, vec![NonClone(1), NonClone(2)]);
        assert_eq!(tail.run_collect().unwrap(), vec![NonClone(3)]);
    }

    #[test]
    fn flat_map_prefix_materializes_on_short_upstream_completion() {
        let values = Source::from_iter([1, 2])
            .flat_map_prefix(3, |prefix| {
                let sum = prefix.into_iter().sum::<i32>();
                Flow::identity().prepend(Source::single(sum))
            })
            .run_collect()
            .unwrap();
        assert_eq!(values, vec![3]);
    }

    #[test]
    fn flat_map_prefix_does_not_materialize_on_early_upstream_failure() {
        let invoked = StdArc::new(StdAtomicBool::new(false));
        let invoked_for_stage = StdArc::clone(&invoked);
        let result = Source::from_factory(|| {
            Box::new(vec![Ok(1), Err(StreamError::Failed("boom".into()))].into_iter())
        })
        .flat_map_prefix(3, move |_prefix| {
            invoked_for_stage.store(true, StdOrdering::SeqCst);
            Flow::identity()
        })
        .run_collect();
        assert_eq!(result, Err(StreamError::Failed("boom".into())));
        assert!(!invoked.load(StdOrdering::SeqCst));
    }

    #[test]
    fn flat_map_concat_flattens_nested_sources_sequentially() {
        let values = Source::from_iter([1, 2, 3])
            .flat_map_concat(|item| Source::from_iter(0..item))
            .run_collect()
            .unwrap();
        assert_eq!(values, vec![0, 0, 1, 0, 1, 2]);
    }

    #[test]
    fn flat_map_merge_respects_breadth_bound() {
        let active = StdArc::new(StdAtomicUsize::new(0));
        let max_active = StdArc::new(StdAtomicUsize::new(0));
        let active_for_stage = StdArc::clone(&active);
        let max_for_stage = StdArc::clone(&max_active);

        let mut values = Source::from_iter(0..6)
            .flat_map_merge(2, move |item| {
                let active = StdArc::clone(&active_for_stage);
                let max_active = StdArc::clone(&max_for_stage);
                Source::future(move || {
                    let active = StdArc::clone(&active);
                    let max_active = StdArc::clone(&max_active);
                    async move {
                        let now = active.fetch_add(1, StdOrdering::SeqCst) + 1;
                        loop {
                            let seen = max_active.load(StdOrdering::SeqCst);
                            if now <= seen {
                                break;
                            }
                            if max_active
                                .compare_exchange(
                                    seen,
                                    now,
                                    StdOrdering::SeqCst,
                                    StdOrdering::SeqCst,
                                )
                                .is_ok()
                            {
                                break;
                            }
                        }
                        thread::sleep(StdDuration::from_millis(20));
                        active.fetch_sub(1, StdOrdering::SeqCst);
                        Ok(item)
                    }
                })
            })
            .run_collect()
            .unwrap();
        values.sort_unstable();
        assert_eq!(values, vec![0, 1, 2, 3, 4, 5]);
        assert!(max_active.load(StdOrdering::SeqCst) <= 2);
    }

    #[test]
    fn flat_map_merge_propagates_inner_failures() {
        let result = Source::from_iter([0, 1, 2])
            .flat_map_merge(2, |item| {
                if item == 1 {
                    Source::failed(StreamError::Failed("boom".into()))
                } else {
                    Source::single(item)
                }
            })
            .run_collect();
        assert_eq!(result, Err(StreamError::Failed("boom".into())));
    }

    #[test]
    fn flat_map_merge_emits_ready_inner_output_while_upstream_is_blocked() {
        let (release_tx, release_rx) = mpsc::channel();
        let release_rx = StdArc::new(std::sync::Mutex::new(Some(release_rx)));
        let queue = Source::from_factory(move || {
            let release_rx = StdArc::clone(&release_rx);
            let mut step = 0_u8;
            Box::new(std::iter::from_fn(move || {
                let item = match step {
                    0 => Some(Ok(0)),
                    1 => {
                        release_rx
                            .lock()
                            .unwrap()
                            .as_ref()
                            .expect("release receiver available")
                            .recv_timeout(StdDuration::from_secs(1))
                            .expect("timed out waiting to release second upstream element");
                        Some(Ok(1))
                    }
                    _ => None,
                };
                step += 1;
                item
            }))
        })
        .flat_map_merge(2, |item| Source::single(item + 10))
        .run_with(Sink::queue())
        .unwrap();

        assert_eq!(queue.pull().unwrap(), Some(10));
        release_tx.send(()).unwrap();
        assert_eq!(queue.pull().unwrap(), Some(11));
        assert!(queue.pull().unwrap().is_none());
    }

    #[test]
    fn group_by_routes_keys_and_drops_closed_keys() {
        let outer = Source::from_iter([0, 1, 2, 3, 4])
            .group_by(4, |item| item % 2, false)
            .run_with(Sink::queue())
            .unwrap();

        let even = outer.pull().unwrap().unwrap();
        let even_completion = even.run_with(Sink::ignore()).unwrap();
        let odd = outer.pull().unwrap().unwrap();
        drop(even_completion);

        assert_eq!(odd.run_collect().unwrap(), vec![1, 3]);
        assert!(outer.pull().unwrap().is_none());
    }

    #[test]
    fn group_by_fails_when_distinct_key_limit_is_exceeded() {
        let outer = Source::from_iter([0, 1, 2])
            .group_by(2, |item| *item, false)
            .run_with(Sink::queue())
            .unwrap();

        let _ = outer.pull().unwrap().unwrap();
        let _ = outer.pull().unwrap().unwrap();
        assert!(matches!(
            outer.pull(),
            Err(StreamError::Failed(message)) if message == "group_by reached max_substreams (2)"
        ));
    }

    #[test]
    fn group_by_can_recreate_closed_substreams_when_enabled() {
        let (release_tx, release_rx) = mpsc::channel();
        let release_rx = StdArc::new(std::sync::Mutex::new(Some(release_rx)));
        let outer = Source::from_factory(move || {
            let release_rx = StdArc::clone(&release_rx);
            let mut step = 0_u8;
            Box::new(std::iter::from_fn(move || {
                let item = match step {
                    0 => Some(Ok(0)),
                    1 => Some(Ok(1)),
                    2 => {
                        release_rx
                            .lock()
                            .unwrap()
                            .as_ref()
                            .expect("release receiver available")
                            .recv_timeout(StdDuration::from_secs(1))
                            .expect("timed out waiting to release recreated key");
                        Some(Ok(0))
                    }
                    _ => None,
                };
                step += 1;
                item
            }))
        })
        .group_by(4, |item| item % 2, true)
        .run_with(Sink::queue())
        .unwrap();

        let even = outer.pull().unwrap().unwrap();
        assert_eq!(wait(even.run_with(Sink::head()).unwrap()), 0);
        release_tx.send(()).unwrap();

        let odd = outer.pull().unwrap().unwrap();
        assert_eq!(odd.run_collect().unwrap(), vec![1]);

        let recreated_even = outer.pull().unwrap().unwrap();
        assert_eq!(recreated_even.run_collect().unwrap(), vec![0]);
        assert!(outer.pull().unwrap().is_none());
    }

    #[test]
    fn group_by_panicking_key_fn_abruptly_terminates_live_substreams() {
        let outer = Source::from_iter([0, 1])
            .group_by(
                4,
                |item| {
                    assert_ne!(*item, 1, "boom");
                    item % 2
                },
                false,
            )
            .run_with(Sink::queue())
            .unwrap();

        let substream = outer.pull().unwrap().unwrap();
        let (result_tx, result_rx) = mpsc::channel();
        thread::spawn(move || {
            let _ = result_tx.send(substream.run_collect());
        });

        assert_eq!(
            result_rx.recv_timeout(StdDuration::from_secs(1)).unwrap(),
            Err(StreamError::AbruptTermination)
        );
        assert!(matches!(outer.pull(), Err(StreamError::AbruptTermination)));
    }

    #[test]
    fn split_when_starts_new_substream_on_boundary_element() {
        let outer = Source::from_iter([1, 2, 0, 3, 0, 4, 5])
            .split_when(|item| *item == 0)
            .run_with(Sink::queue())
            .unwrap();

        let first = outer.pull().unwrap().unwrap();
        assert_eq!(first.run_collect().unwrap(), vec![1, 2]);
        let second = outer.pull().unwrap().unwrap();
        assert_eq!(second.run_collect().unwrap(), vec![0, 3]);
        let third = outer.pull().unwrap().unwrap();
        assert_eq!(third.run_collect().unwrap(), vec![0, 4, 5]);
        assert!(outer.pull().unwrap().is_none());
    }

    #[test]
    fn split_after_ends_current_substream_on_boundary_element() {
        let outer = Source::from_iter([1, 2, 0, 3, 0, 4, 5])
            .split_after(|item| *item == 0)
            .run_with(Sink::queue())
            .unwrap();

        let first = outer.pull().unwrap().unwrap();
        assert_eq!(first.run_collect().unwrap(), vec![1, 2, 0]);
        let second = outer.pull().unwrap().unwrap();
        assert_eq!(second.run_collect().unwrap(), vec![3, 0]);
        let third = outer.pull().unwrap().unwrap();
        assert_eq!(third.run_collect().unwrap(), vec![4, 5]);
        assert!(outer.pull().unwrap().is_none());
    }

    #[test]
    fn split_when_panicking_predicate_abruptly_terminates_live_substreams() {
        let outer = Source::from_iter([1, 2])
            .split_when(|item| {
                assert_ne!(*item, 2, "boom");
                false
            })
            .run_with(Sink::queue())
            .unwrap();

        let substream = outer.pull().unwrap().unwrap();
        let (result_tx, result_rx) = mpsc::channel();
        thread::spawn(move || {
            let _ = result_tx.send(substream.run_collect());
        });

        assert_eq!(
            result_rx.recv_timeout(StdDuration::from_secs(1)).unwrap(),
            Err(StreamError::AbruptTermination)
        );
        assert!(matches!(outer.pull(), Err(StreamError::AbruptTermination)));
    }

    #[test]
    fn split_when_pre_buffer_segments_match_expected_count() {
        let outer = Source::from_iter(0..100)
            .split_when(|item| *item != 0 && *item % 10 == 0)
            .run_with(Sink::queue())
            .unwrap();
        let mut segment_count = 0;
        while let Some(substream) = outer.pull().unwrap() {
            let items: Vec<i32> = substream.run_collect().unwrap();
            assert!(!items.is_empty(), "segment should not be empty");
            segment_count += 1;
        }
        assert_eq!(segment_count, 10, "100 elements in segments of 10");
    }

    #[test]
    fn split_after_pre_buffer_segments_match_expected_count() {
        let outer = Source::from_iter(0..100)
            .split_after(|item| (*item + 1) % 10 == 0)
            .run_with(Sink::queue())
            .unwrap();
        let mut segment_count = 0;
        let mut total = 0_i32;
        while let Some(substream) = outer.pull().unwrap() {
            let items: Vec<i32> = substream.run_collect().unwrap();
            assert!(!items.is_empty(), "segment should not be empty");
            total += items.len() as i32;
            segment_count += 1;
        }
        assert_eq!(segment_count, 10);
        assert_eq!(total, 100);
    }

    #[test]
    fn group_by_single_key_fused_matches_general_path() {
        let outer = Source::from_iter(0..1000i64)
            .group_by(1, |_| 0u8, false)
            .run_with(Sink::queue())
            .unwrap();
        let substream = outer.pull().unwrap().unwrap();
        let items: Vec<i64> = substream.run_collect().unwrap();
        assert_eq!(items.len(), 1000);
        assert_eq!(items[0], 0);
        assert_eq!(items[999], 999);
        assert!(outer.pull().unwrap().is_none());
    }

    #[test]
    fn group_by_single_key_fused_handles_key_change_with_substream_limit() {
        let outer = Source::from_iter([0, 1, 0])
            .group_by(2, |item| *item, false)
            .run_with(Sink::queue())
            .unwrap();
        let mut sources = vec![];
        while let Some(source) = outer.pull().unwrap() {
            sources.push(source);
        }
        assert_eq!(sources.len(), 2);
        assert_eq!(sources[0].clone().run_collect().unwrap(), vec![0, 0]);
        assert_eq!(sources[1].clone().run_collect().unwrap(), vec![1]);
    }

    #[test]
    fn flat_map_merge_lock_lighter_matches_expected_count() {
        let items = Source::from_iter(0..20)
            .flat_map_merge(2, |item| Source::single(item + 100))
            .run_with(Sink::queue())
            .unwrap();
        let mut count = 0;
        while items.pull().unwrap().is_some() {
            count += 1;
        }
        assert_eq!(count, 20);
    }

    // Verify that group_by emits the substream BEFORE upstream completes —
    // i.e., the worker does not buffer all elements before delivery.
    //
    // The test uses a rendezvous channel to synchronize: the group_by worker
    // will block waiting for more items (the channel is empty after item 0),
    // so if the outer substream is NOT emitted before item 1 arrives, the
    // `outer.pull()` call in a separate thread cannot return during that window.
    // With the old "single_key_buf" batching, pull() would block until a
    // key-change or EOF — which is a liveness bug for infinite or slow sources.
    #[test]
    fn group_by_single_key_emits_substream_before_upstream_completes() {
        // sync_channel(0): rendezvous — sender blocks until receiver is ready,
        // guaranteeing the worker processes exactly one item then waits.
        let (tx, rx) = mpsc::sync_channel::<i32>(0);
        let rx = StdArc::new(std::sync::Mutex::new(rx));

        let outer = Source::from_factory({
            let rx = StdArc::clone(&rx);
            move || {
                let rx = StdArc::clone(&rx);
                Box::new(std::iter::from_fn(move || {
                    rx.lock().unwrap().recv().ok().map(Ok)
                })) as BoxStream<i32>
            }
        })
        .group_by(1, |_| 0u8, false)
        .run_with(Sink::queue())
        .unwrap();

        // Kick off the producer: sends item 0, then waits for the outer
        // substream to be pulled (via a second channel) before sending more.
        let (sub_tx, sub_rx) = mpsc::channel::<Source<i32>>();
        let outer_thread = thread::spawn(move || {
            let substream = outer.pull().unwrap().expect("expected a substream");
            sub_tx.send(substream).unwrap();
        });

        // Send the first item — the worker processes it and (with the fix)
        // immediately emits the substream to outer.
        tx.send(0).unwrap();

        // The outer pull must complete within the timeout.
        let substream = sub_rx
            .recv_timeout(StdDuration::from_secs(5))
            .expect("timed out — group_by buffered first element before emitting substream");

        // Now deliver the remaining items and close.
        for i in 1..100_i32 {
            tx.send(i).unwrap();
        }
        drop(tx);

        let items: Vec<i32> = substream.run_collect().unwrap();
        assert_eq!(items.len(), 100);
        outer_thread.join().unwrap();
    }

    #[test]
    fn group_by_concurrent_live_substreams_do_not_hold_ready_item_stress() {
        const STREAMS: usize = 32;
        const ROUNDS: usize = 8;
        const ITEMS: i64 = 8;

        for _ in 0..ROUNDS {
            let barrier = StdArc::new(std::sync::Barrier::new(STREAMS));
            let mut handles = Vec::with_capacity(STREAMS);

            for _ in 0..STREAMS {
                let barrier = StdArc::clone(&barrier);
                handles.push(thread::spawn(move || {
                    let (tx, rx) = mpsc::sync_channel::<i64>(0);
                    let rx = StdArc::new(std::sync::Mutex::new(rx));

                    let outer = Source::from_factory({
                        let rx = StdArc::clone(&rx);
                        move || {
                            let rx = StdArc::clone(&rx);
                            Box::new(std::iter::from_fn(move || {
                                rx.lock().unwrap().recv().ok().map(Ok)
                            })) as BoxStream<i64>
                        }
                    })
                    .group_by(1, |_| 0_u8, false)
                    .run_with(Sink::queue())
                    .unwrap();

                    barrier.wait();

                    tx.send(0).unwrap();
                    let substream = outer.pull().unwrap().expect("expected group_by substream");
                    let subqueue = substream.run_with(Sink::queue()).unwrap();
                    assert_eq!(subqueue.pull().unwrap(), Some(0));

                    for item in 1..ITEMS {
                        tx.send(item).unwrap();
                        assert_eq!(subqueue.pull().unwrap(), Some(item));
                    }
                    drop(tx);

                    assert!(subqueue.pull().unwrap().is_none());
                    assert!(outer.pull().unwrap().is_none());
                }));
            }

            for handle in handles {
                handle.join().expect("group_by stress worker panicked");
            }
        }
    }

    // Verify that split_when emits each substream as a live stream, NOT after
    // accumulating all segment elements into memory.  The channel is sized to
    // be larger than LIVE_SUBSTREAM_CAPACITY (256) so that, if the old Vec
    // pre-buffering path were active, the first substream would never be emitted
    // before the boundary.  With live substreams the substream is visible
    // immediately upon the first item.
    #[test]
    fn split_when_emits_substream_before_segment_ends() {
        // 512 > LIVE_SUBSTREAM_CAPACITY (256): the live substream will block
        // on backpressure mid-segment, but that's fine — the outer substream
        // IS emitted immediately and the consumer can drain it concurrently.
        const SEGMENT_LEN: usize = 300;

        let (tx, rx) = mpsc::sync_channel::<i32>(SEGMENT_LEN * 2 + 4);
        for i in 0..SEGMENT_LEN as i32 {
            tx.send(i).unwrap();
        }
        tx.send(-1).unwrap(); // boundary
        tx.send(99).unwrap(); // second segment
        drop(tx);

        let outer = Source::from_iter(rx)
            .split_when(|item| *item == -1)
            .run_with(Sink::queue())
            .unwrap();

        let (result_tx, result_rx) = mpsc::channel();
        thread::spawn(move || {
            let first = outer.pull().unwrap().expect("expected first substream");
            let items: Vec<i32> = first.run_collect().unwrap();
            let second = outer.pull().unwrap().expect("expected second substream");
            let items2: Vec<i32> = second.run_collect().unwrap();
            let done = outer.pull().unwrap().is_none();
            let _ = result_tx.send((items, items2, done));
        });

        let (items, items2, done) = result_rx
            .recv_timeout(StdDuration::from_secs(5))
            .expect("timed out — split_when is buffering the whole segment");
        assert_eq!(items.len(), SEGMENT_LEN);
        assert_eq!(items2, vec![-1, 99]);
        assert!(done);
    }

    #[test]
    fn split_after_emits_substream_before_segment_ends() {
        const SEGMENT_LEN: usize = 300;

        let (tx, rx) = mpsc::sync_channel::<i32>(SEGMENT_LEN * 2 + 4);
        for i in 0..SEGMENT_LEN as i32 {
            tx.send(i).unwrap();
        }
        tx.send(-1).unwrap(); // boundary (included in first segment for split_after)
        tx.send(99).unwrap();
        drop(tx);

        let outer = Source::from_iter(rx)
            .split_after(|item| *item == -1)
            .run_with(Sink::queue())
            .unwrap();

        let (result_tx, result_rx) = mpsc::channel();
        thread::spawn(move || {
            let first = outer.pull().unwrap().expect("expected first substream");
            let items: Vec<i32> = first.run_collect().unwrap();
            let second = outer.pull().unwrap().expect("expected second substream");
            let items2: Vec<i32> = second.run_collect().unwrap();
            let done = outer.pull().unwrap().is_none();
            let _ = result_tx.send((items, items2, done));
        });

        let (items, items2, done) = result_rx
            .recv_timeout(StdDuration::from_secs(5))
            .expect("timed out — split_after is buffering the whole segment");
        // split_after includes the boundary element in the first segment.
        assert_eq!(items.len(), SEGMENT_LEN + 1);
        assert_eq!(items2, vec![99]);
        assert!(done);
    }

    // Stress test: run flat_map_merge 20 times with many concurrent inner
    // streams to confirm the fixed coordinator tail loop never hangs.
    // Intended to be run with --test-threads=32 to expose the race.
    #[test]
    fn flat_map_merge_coordinator_no_lost_wakeup_stress() {
        for _ in 0..20 {
            let result = Source::from_iter(0..50_i32)
                .flat_map_merge(8, |item| Source::from_iter(item..item + 3))
                .run_with(Sink::fold(0i64, |acc, item| acc + item as i64))
                .unwrap()
                .wait();
            // Each i in 0..50 contributes i + (i+1) + (i+2) = 3i + 3.
            // Sum = 3 * (0+1+..+49) + 3*50 = 3*1225 + 150 = 3675 + 150 = 3825.
            assert_eq!(result, Ok(3825), "flat_map_merge produced wrong sum");
        }
    }

    // Stress test for the single-mutex flat_map_merge refactor.
    // 20 runs with high concurrency to verify no deadlock, lost wakeup,
    // or window-accounting corruption under the merged-lock design.
    // Intended to be run with --test-threads=32.
    #[test]
    fn flat_map_merge_single_mutex_race_stress() {
        for _ in 0..20 {
            let result = Source::from_iter(0..100_i64)
                .flat_map_merge(16, |item| Source::from_iter([item, item + 1000]))
                .run_with(Sink::fold(0i64, |acc, v| acc + v))
                .unwrap()
                .wait();
            // sum of item + (item+1000) for item in 0..100
            // = sum(0..100) + sum(0..100) + 100*1000
            // = 4950 + 4950 + 100000 = 109900
            assert_eq!(result, Ok(109_900), "flat_map_merge single-mutex stress");
        }
    }

    // Bounded-memory test for split_when batching: verify that items are
    // delivered to a slow consumer and not held indefinitely in the writer
    // buffer.  Uses a rendezvous-style channel where the producer sends MANY
    // items into the outer queue (larger than LIVE_SUBSTREAM_BATCH = 64), then
    // blocks.  The consumer must be able to drain the first segment fully — i.e.
    // the writer must have flushed even when the batch was smaller than 64.
    #[test]
    fn split_when_bounded_memory_rendezvous() {
        // Segment of 100 items (> LIVE_SUBSTREAM_BATCH=64): the writer will
        // flush at 64, then again at the boundary (remaining 36).
        const SEGMENT: usize = 100;
        let (tx, rx) = mpsc::sync_channel::<i32>(SEGMENT * 4);
        for i in 0..SEGMENT as i32 {
            tx.send(i).unwrap();
        }
        tx.send(-1).unwrap(); // boundary (split_when: starts new segment)
        // Second segment
        for i in 0..10_i32 {
            tx.send(i).unwrap();
        }
        drop(tx);

        let outer = Source::from_iter(rx)
            .split_when(|item| *item == -1)
            .run_with(Sink::queue())
            .unwrap();

        let (result_tx, result_rx) = mpsc::channel();
        thread::spawn(move || {
            let first = outer.pull().unwrap().expect("first segment");
            let seg1: Vec<i32> = first.run_collect().unwrap();
            let second = outer.pull().unwrap().expect("second segment");
            let seg2: Vec<i32> = second.run_collect().unwrap();
            let done = outer.pull().unwrap().is_none();
            result_tx.send((seg1, seg2, done)).unwrap();
        });

        let (seg1, seg2, done) = result_rx
            .recv_timeout(StdDuration::from_secs(5))
            .expect("timed out — split_when writer held items past LIVE_SUBSTREAM_BATCH");
        assert_eq!(seg1.len(), SEGMENT, "first segment length");
        assert_eq!(seg2[0], -1, "boundary element starts second segment");
        assert_eq!(seg2.len(), 11, "second segment: boundary + 10 items");
        assert!(done);
    }

    // Bounded-memory test for group_by batching: verifies that a slow consumer
    // waiting on a single-key substream eventually receives all items (i.e. the
    // write batch is flushed, not held indefinitely).
    #[test]
    fn group_by_single_key_bounded_memory_rendezvous() {
        // 200 items, all same key, batched in groups of 64 by the writer.
        // The consumer blocks on run_collect(); all items must arrive.
        const N: usize = 200;
        let outer = Source::from_iter(0..N as i64)
            .group_by(1, |_| 0u8, false)
            .run_with(Sink::queue())
            .unwrap();

        let (result_tx, result_rx) = mpsc::channel();
        thread::spawn(move || {
            let substream = outer.pull().unwrap().expect("substream");
            let items: Vec<i64> = substream.run_collect().unwrap();
            let done = outer.pull().unwrap().is_none();
            result_tx.send((items, done)).unwrap();
        });

        let (items, done) = result_rx
            .recv_timeout(StdDuration::from_secs(5))
            .expect("timed out — group_by write batch held items beyond LIVE_SUBSTREAM_BATCH");
        assert_eq!(items.len(), N, "all items delivered");
        assert_eq!(items[0], 0);
        assert_eq!(items[N - 1], (N - 1) as i64);
        assert!(done);
    }

    #[test]
    fn scan_emits_seed_and_accumulated_values() {
        let result = Source::from_iter(1..=3)
            .scan(0, |acc, item| acc + item)
            .run_collect()
            .unwrap();

        assert_eq!(result, vec![0, 1, 3, 6]);
    }

    #[test]
    fn limit_fails_after_max_elements() {
        let result = Source::from_iter(0..3).limit(2).run_collect();

        assert_eq!(result, Err(StreamError::LimitExceeded { max: 2 }));
    }

    #[test]
    fn limit_weighted_fails_with_limit_error_like_akka() {
        let result = Source::from_iter(["this", "is", "some", "string"])
            .via(Flow::identity().limit_weighted(15, |item: &&str| item.len()))
            .run_collect();

        assert_eq!(result, Err(StreamError::LimitExceeded { max: 15 }));
    }

    #[test]
    fn grouped_weighted_allows_oversized_first_element_like_akka() {
        let result = Source::from_iter([10_usize, 1, 2])
            .via(Flow::identity().grouped_weighted(5, |item: &usize| *item))
            .run_collect()
            .unwrap();

        assert_eq!(result, vec![vec![10], vec![1, 2]]);
    }

    #[test]
    fn grouped_weighted_keeps_oversized_later_element_in_current_group_like_akka() {
        let result = Source::from_iter([1_usize, 10, 2])
            .via(Flow::identity().grouped_weighted(5, |item: &usize| *item))
            .run_collect()
            .unwrap();

        assert_eq!(result, vec![vec![1, 10], vec![2]]);
    }

    #[test]
    fn sink_terminals_materialize_results() {
        let sum = Source::from_iter(1..=4)
            .run_with(Sink::fold(0, |acc, item| acc + item))
            .unwrap();

        assert_eq!(wait(sum), 10);
        assert_eq!(
            wait(Source::from_iter(1..=4).run_with(Sink::head()).unwrap()),
            1
        );
        assert_eq!(
            wait(Source::from_iter(1..=4).run_with(Sink::last()).unwrap()),
            4
        );
    }

    #[test]
    fn all_terminal_sink_variants_complete() {
        assert_eq!(
            wait(
                Source::from_iter([1, 2, 3])
                    .run_with(Sink::collect())
                    .unwrap()
            ),
            vec![1, 2, 3]
        );
        assert_eq!(
            wait(
                Source::<i32>::empty()
                    .run_with(Sink::head_option())
                    .unwrap()
            ),
            None
        );
        assert_eq!(
            wait(
                Source::from_iter([1, 2, 3])
                    .run_with(Sink::last_option())
                    .unwrap()
            ),
            Some(3)
        );
        assert_eq!(
            wait(
                Source::from_iter([1, 2, 3])
                    .run_with(Sink::reduce(|acc, item| acc + item))
                    .unwrap()
            ),
            6
        );

        let seen = StdArc::new(StdAtomicUsize::new(0));
        let seen_by_sink = StdArc::clone(&seen);
        assert_eq!(
            wait(
                Source::from_iter([1_usize, 2, 3])
                    .run_with(Sink::foreach(move |item| {
                        seen_by_sink.fetch_add(item, StdOrdering::SeqCst);
                    }))
                    .unwrap()
            ),
            NotUsed
        );
        assert_eq!(seen.load(StdOrdering::SeqCst), 6);
    }

    #[test]
    fn take_last_zero_returns_empty_vector() {
        let result = Source::from_iter([1, 2, 3])
            .run_with(Sink::take_last(0))
            .unwrap();

        assert_eq!(wait(result), Vec::<i32>::new());
    }

    #[test]
    fn bounded_head_terminals_complete_inline() {
        let materializer = Materializer::new();

        let mut head = Source::from_iter(0_u64..1_000)
            .run_with_materializer(Sink::head(), &materializer)
            .unwrap();
        assert_eq!(materializer.active_streams(), 0);
        assert_eq!(head.try_wait(), Some(Ok(0)));

        let mut filtered_head = Source::from_iter(0_u64..1_000)
            .filter(|item| *item >= 10)
            .run_with_materializer(Sink::head(), &materializer)
            .unwrap();
        assert_eq!(materializer.active_streams(), 0);
        assert_eq!(filtered_head.try_wait(), Some(Ok(10)));

        let mut head_option = Source::<u64>::empty()
            .run_with_materializer(Sink::head_option(), &materializer)
            .unwrap();
        assert_eq!(materializer.active_streams(), 0);
        assert_eq!(head_option.try_wait(), Some(Ok(None)));
    }

    #[test]
    fn bounded_head_fast_path_preserves_terminal_errors() {
        let materializer = Materializer::new();

        let mut empty = Source::<u64>::empty()
            .run_with_materializer(Sink::head(), &materializer)
            .unwrap();
        assert_eq!(empty.try_wait(), Some(Err(StreamError::EmptyStream)));

        let mut failed = Source::<u64>::failed(StreamError::Failed("boom".into()))
            .run_with_materializer(Sink::head(), &materializer)
            .unwrap();
        assert_eq!(
            failed.try_wait(),
            Some(Err(StreamError::Failed("boom".into())))
        );
        assert_eq!(materializer.active_streams(), 0);
    }

    #[test]
    fn runnable_graph_composes_source_and_sink() {
        let graph = Source::from_iter(1..=4)
            .map(|item| item * 2)
            .to_mat(Sink::fold(0, |acc, item| acc + item), Keep::right);

        assert_eq!(wait(graph.run().unwrap()), 20);

        let graph = Source::single(1)
            .map_materialized_value(|_| 20)
            .to(Sink::ignore())
            .map_materialized_value(|value| value + 1);
        assert_eq!(graph.run().unwrap(), 21);

        let ignored = Source::single(1).to(Sink::ignore()).run().unwrap();
        assert_eq!(ignored, NotUsed);
    }

    #[test]
    fn materialized_values_follow_keep_defaults() {
        let source = Source::single(1).map_materialized_value(|_| "source");
        let flow = Flow::identity().map_materialized_value(|_| "flow");

        let source_mat = source.clone().via(flow.clone()).to(Sink::ignore()).run();
        assert_eq!(source_mat.unwrap(), "source");

        let combined = source
            .via_mat(flow, Keep::both)
            .to_mat(Sink::ignore(), Keep::both)
            .run()
            .unwrap();
        assert_eq!(combined.0, ("source", "flow"));
        assert_eq!(wait(combined.1), NotUsed);

        let sink_mat = Source::single(41)
            .map_materialized_value(|_| "ignored source")
            .run_with(Sink::fold(1, |acc, item| acc + item))
            .unwrap();
        assert_eq!(wait(sink_mat), 42);
    }

    #[test]
    fn flow_to_sink_preserves_flow_materialized_value_by_default() {
        let sink = Flow::identity()
            .map(|item: i32| item + 1)
            .map_materialized_value(|_| "flow")
            .to(Sink::fold(0, |acc, item| acc + item));

        let materialized = Source::from_iter([1, 2, 3]).run_with(sink).unwrap();

        assert_eq!(materialized, "flow");
        let explicit = Flow::identity()
            .map(|item: i32| item + 1)
            .map_materialized_value(|_| "flow")
            .to_mat(Sink::fold(0, |acc, item| acc + item), Keep::both)
            .run_with(Source::from_iter([1, 2, 3]))
            .unwrap();
        assert_eq!(explicit, NotUsed);

        let explicit = Source::from_iter([1, 2, 3])
            .run_with(
                Flow::identity()
                    .map(|item: i32| item + 1)
                    .map_materialized_value(|_| "flow")
                    .to_mat(Sink::fold(0, |acc, item| acc + item), Keep::both),
            )
            .unwrap();
        assert_eq!(explicit.0, "flow");
        assert_eq!(wait(explicit.1), 9);
    }

    #[test]
    fn materializer_shutdown_fails_materialization() {
        let materializer = Materializer::new();
        let named = materializer.with_name_prefix("test-stream");
        materializer.shutdown();

        let graph = Source::single(1).to(Sink::ignore());

        assert_eq!(named.name_prefix(), "test-stream");
        assert_eq!(
            graph.run_with_materializer(&named),
            Err(StreamError::AbruptTermination)
        );
    }

    #[test]
    fn materializer_shutdown_fails_running_stream_completion() {
        let materializer = Materializer::new();
        let completion = Source::repeat(1)
            .run_with_materializer(Sink::ignore(), &materializer)
            .unwrap();

        assert_eq!(materializer.active_streams(), 1);
        materializer.shutdown();
        assert_eq!(completion.wait(), Err(StreamError::AbruptTermination));
        assert_eq!(materializer.active_streams(), 0);
    }

    #[test]
    fn dropped_stream_completion_cancels_running_stream() {
        let materializer = Materializer::new();
        let completion = Source::repeat(1)
            .run_with_materializer(Sink::ignore(), &materializer)
            .unwrap();

        assert_eq!(materializer.active_streams(), 1);
        drop(completion);
        for _ in 0..50 {
            if materializer.active_streams() == 0 {
                break;
            }
            thread::sleep(Duration::from_millis(5));
        }
        assert_eq!(materializer.active_streams(), 0);
    }

    #[test]
    fn runtime_timers_fire_cancel_and_stop_on_shutdown() {
        let materializer = Materializer::new();
        let (once_tx, once_rx) = mpsc::channel();
        let once = materializer.schedule_once(Duration::from_millis(5), move || {
            once_tx.send(()).unwrap();
        });
        once_rx.recv_timeout(Duration::from_millis(250)).unwrap();
        assert!(!once.is_cancelled());

        let (cancelled_tx, cancelled_rx) = mpsc::channel();
        let cancelled = materializer.schedule_once(Duration::from_millis(25), move || {
            cancelled_tx.send(()).unwrap();
        });
        assert!(cancelled.cancel());
        assert!(!cancelled.cancel());
        assert!(cancelled.is_cancelled());
        assert!(
            cancelled_rx
                .recv_timeout(Duration::from_millis(75))
                .is_err()
        );

        let fixed_delay_count = StdArc::new(StdAtomicUsize::new(0));
        let fixed_delay_task_count = StdArc::clone(&fixed_delay_count);
        let fixed_delay = materializer.schedule_with_fixed_delay(
            Duration::from_millis(1),
            Duration::from_millis(5),
            move || {
                fixed_delay_task_count.fetch_add(1, StdOrdering::SeqCst);
            },
        );
        thread::sleep(Duration::from_millis(25));
        assert!(fixed_delay_count.load(StdOrdering::SeqCst) > 0);
        fixed_delay.cancel();

        let fixed_rate_count = StdArc::new(StdAtomicUsize::new(0));
        let fixed_rate_task_count = StdArc::clone(&fixed_rate_count);
        let fixed_rate = materializer.schedule_at_fixed_rate(
            Duration::from_millis(1),
            Duration::from_millis(5),
            move || {
                fixed_rate_task_count.fetch_add(1, StdOrdering::SeqCst);
            },
        );
        thread::sleep(Duration::from_millis(25));
        assert!(fixed_rate_count.load(StdOrdering::SeqCst) > 0);
        fixed_rate.cancel();

        let shutdown_materializer = Materializer::new();
        let (shutdown_tx, shutdown_rx) = mpsc::channel();
        shutdown_materializer.schedule_once(Duration::from_millis(25), move || {
            shutdown_tx.send(()).unwrap();
        });
        shutdown_materializer.shutdown();
        assert!(shutdown_rx.recv_timeout(Duration::from_millis(75)).is_err());
    }

    #[test]
    fn runtime_timer_driver_preserves_fixed_rate_cadence_under_slow_tasks() {
        use std::sync::{Condvar, Mutex};

        #[derive(Debug)]
        enum TimerEvent {
            Started(usize, Instant),
            Completed(usize, Instant),
        }

        let recv_event = |rx: &mpsc::Receiver<TimerEvent>, label: &str| {
            rx.recv_timeout(Duration::from_secs(20))
                .unwrap_or_else(|err| panic!("{label}: expected timer event within 20 s: {err}"))
        };
        let release = |gate: &StdArc<(Mutex<bool>, Condvar)>| {
            let (released, condvar) = &**gate;
            let mut released = released.lock().unwrap_or_else(|poison| poison.into_inner());
            *released = true;
            condvar.notify_all();
        };

        let interval = Duration::from_secs(2);
        let overrun = interval + Duration::from_millis(250);

        let rate_materializer = Materializer::new();
        let (rate_tx, rate_rx) = mpsc::channel();
        let rate_runs = StdArc::new(StdAtomicUsize::new(0));
        let rate_task_runs = StdArc::clone(&rate_runs);
        let rate_gate = StdArc::new((Mutex::new(false), Condvar::new()));
        let rate_task_gate = StdArc::clone(&rate_gate);
        let fixed_rate =
            rate_materializer.schedule_at_fixed_rate(Duration::ZERO, interval, move || {
                let run = rate_task_runs.fetch_add(1, StdOrdering::SeqCst) + 1;
                rate_tx
                    .send(TimerEvent::Started(run, Instant::now()))
                    .unwrap();
                if run == 1 {
                    let (released, condvar) = &*rate_task_gate;
                    let mut released = released.lock().unwrap_or_else(|poison| poison.into_inner());
                    while !*released {
                        released = condvar
                            .wait(released)
                            .unwrap_or_else(|poison| poison.into_inner());
                    }
                    rate_tx
                        .send(TimerEvent::Completed(run, Instant::now()))
                        .unwrap();
                }
            });
        let rate_first_started = match recv_event(&rate_rx, "fixed-rate first task") {
            TimerEvent::Started(1, at) => at,
            other => panic!("fixed-rate first task: unexpected event {other:?}"),
        };
        assert!(wait_until(Duration::from_secs(20), || {
            rate_first_started.elapsed() >= overrun
        }));
        release(&rate_gate);
        let rate_first_completed = match recv_event(&rate_rx, "fixed-rate first completion") {
            TimerEvent::Completed(1, at) => at,
            other => panic!("fixed-rate first completion: unexpected event {other:?}"),
        };
        let rate_second_started = match recv_event(&rate_rx, "fixed-rate second task") {
            TimerEvent::Started(2, at) => at,
            other => panic!("fixed-rate second task: unexpected event {other:?}"),
        };
        fixed_rate.cancel();
        rate_materializer.shutdown();

        let delay_materializer = Materializer::new();
        let (delay_tx, delay_rx) = mpsc::channel();
        let delay_runs = StdArc::new(StdAtomicUsize::new(0));
        let delay_task_runs = StdArc::clone(&delay_runs);
        let delay_gate = StdArc::new((Mutex::new(false), Condvar::new()));
        let delay_task_gate = StdArc::clone(&delay_gate);
        let fixed_delay =
            delay_materializer.schedule_with_fixed_delay(Duration::ZERO, interval, move || {
                let run = delay_task_runs.fetch_add(1, StdOrdering::SeqCst) + 1;
                delay_tx
                    .send(TimerEvent::Started(run, Instant::now()))
                    .unwrap();
                if run == 1 {
                    let (released, condvar) = &*delay_task_gate;
                    let mut released = released.lock().unwrap_or_else(|poison| poison.into_inner());
                    while !*released {
                        released = condvar
                            .wait(released)
                            .unwrap_or_else(|poison| poison.into_inner());
                    }
                    delay_tx
                        .send(TimerEvent::Completed(run, Instant::now()))
                        .unwrap();
                }
            });
        let delay_first_started = match recv_event(&delay_rx, "fixed-delay first task") {
            TimerEvent::Started(1, at) => at,
            other => panic!("fixed-delay first task: unexpected event {other:?}"),
        };
        assert!(wait_until(Duration::from_secs(20), || {
            delay_first_started.elapsed() >= overrun
        }));
        release(&delay_gate);
        let delay_first_completed = match recv_event(&delay_rx, "fixed-delay first completion") {
            TimerEvent::Completed(1, at) => at,
            other => panic!("fixed-delay first completion: unexpected event {other:?}"),
        };
        let delay_second_started = match recv_event(&delay_rx, "fixed-delay second task") {
            TimerEvent::Started(2, at) => at,
            other => panic!("fixed-delay second task: unexpected event {other:?}"),
        };
        fixed_delay.cancel();
        delay_materializer.shutdown();

        let rate_task_time = rate_first_completed.duration_since(rate_first_started);
        let rate_catch_up = rate_second_started.duration_since(rate_first_completed);
        let delay_task_time = delay_first_completed.duration_since(delay_first_started);
        let delay_gap = delay_second_started.duration_since(delay_first_completed);
        assert!(
            rate_task_time >= interval,
            "fixed-rate first task should overrun its interval; ran for {rate_task_time:?}"
        );
        assert!(
            rate_catch_up < interval,
            "fixed-rate second task should catch up after an overrun; waited {rate_catch_up:?}"
        );
        assert!(
            delay_task_time >= interval,
            "fixed-delay first task should overrun its interval; ran for {delay_task_time:?}"
        );
        assert!(
            delay_gap >= interval,
            "fixed-delay second task fired before one full delay elapsed after completion: {delay_gap:?}",
        );
    }

    #[test]
    fn runtime_repeating_timer_cancellation_stops_future_fires() {
        let materializer = Materializer::new();
        let (tx, rx) = mpsc::channel();
        let timer = materializer.schedule_at_fixed_rate(
            Duration::from_millis(1),
            Duration::from_millis(30),
            move || {
                tx.send(()).unwrap();
            },
        );

        rx.recv_timeout(Duration::from_millis(250)).unwrap();
        assert!(timer.cancel());
        assert!(rx.recv_timeout(Duration::from_millis(90)).is_err());
        materializer.shutdown();
    }

    #[test]
    fn runtime_panicking_once_timer_does_not_kill_driver_or_later_timers() {
        let materializer = Materializer::new();
        materializer.schedule_once(Duration::from_millis(1), || {
            panic!("timer boom");
        });

        let (tx, rx) = mpsc::channel();
        materializer.schedule_once(Duration::from_millis(20), move || {
            tx.send(()).unwrap();
        });

        rx.recv_timeout(Duration::from_millis(250)).unwrap();
        materializer.shutdown();
    }

    #[test]
    fn runtime_panicking_fixed_rate_timer_stops_itself_and_leaves_driver_alive() {
        let materializer = Materializer::new();
        let panic_count = StdArc::new(StdAtomicUsize::new(0));
        let panic_count_task = StdArc::clone(&panic_count);
        materializer.schedule_at_fixed_rate(Duration::ZERO, Duration::from_millis(20), move || {
            panic_count_task.fetch_add(1, StdOrdering::SeqCst);
            panic!("fixed-rate boom");
        });

        assert!(wait_until(Duration::from_millis(150), || {
            panic_count.load(StdOrdering::SeqCst) == 1
        }));

        let (tx, rx) = mpsc::channel();
        materializer.schedule_once(Duration::from_millis(30), move || {
            tx.send(()).unwrap();
        });
        rx.recv_timeout(Duration::from_millis(250)).unwrap();

        thread::sleep(Duration::from_millis(90));
        assert_eq!(panic_count.load(StdOrdering::SeqCst), 1);
        materializer.shutdown();
    }

    #[test]
    fn runtime_slow_timer_task_does_not_delay_unrelated_timers() {
        let materializer = Materializer::new();
        let (started_tx, started_rx) = mpsc::channel();
        let release_gate = StdArc::new((Mutex::new(false), Condvar::new()));
        let release_task = StdArc::clone(&release_gate);
        let runs = StdArc::new(StdAtomicUsize::new(0));
        let task_runs = StdArc::clone(&runs);
        let slow_timer = materializer.schedule_at_fixed_rate(
            Duration::ZERO,
            Duration::from_millis(250),
            move || {
                if task_runs.fetch_add(1, StdOrdering::SeqCst) == 0 {
                    started_tx.send(()).unwrap();
                    let (released, condvar) = &*release_task;
                    let mut released = released.lock().unwrap_or_else(|poison| poison.into_inner());
                    while !*released {
                        released = condvar
                            .wait(released)
                            .unwrap_or_else(|poison| poison.into_inner());
                    }
                }
            },
        );

        started_rx
            .recv_timeout(Duration::from_secs(10))
            .expect("slow timer task should start");

        let (tx, rx) = mpsc::channel();
        materializer.schedule_once(Duration::from_millis(10), move || {
            tx.send(()).unwrap();
        });
        let fired = rx.recv_timeout(Duration::from_secs(10));

        let (released, condvar) = &*release_gate;
        let mut released = released.lock().unwrap_or_else(|poison| poison.into_inner());
        *released = true;
        condvar.notify_all();
        drop(released);

        slow_timer.cancel();
        materializer.shutdown();
        fired.expect("unrelated timer should fire while slow timer task is still blocked");
    }

    #[test]
    fn runtime_shutdown_stops_timer_driver_thread() {
        let materializer = Materializer::new();
        assert!(wait_until(Duration::from_secs(1), || materializer
            .timer_driver_is_live()));

        materializer.shutdown();
        assert!(wait_until(Duration::from_secs(2), || !materializer
            .timer_driver_is_live()));
    }

    #[test]
    fn runtime_timer_driver_orders_many_timers_by_deadline() {
        let materializer = Materializer::new();
        let (tx, rx) = mpsc::channel();
        let schedule = [(450_u64, 4_u8), (50, 1), (350, 3), (150, 2), (550, 5)];

        for (delay_ms, value) in schedule {
            let tx = tx.clone();
            materializer.schedule_once(Duration::from_millis(delay_ms), move || {
                tx.send(value).unwrap();
            });
        }
        drop(tx);

        let mut received = Vec::new();
        for _ in 0..schedule.len() {
            received.push(rx.recv_timeout(Duration::from_secs(10)).unwrap());
        }
        materializer.shutdown();

        assert_eq!(received, vec![1, 2, 3, 4, 5]);
    }

    #[test]
    fn runtime_timer_driver_uses_one_thread_per_runtime_regardless_of_timer_count() {
        let materializer = Materializer::new();
        let thread_name = materializer.timer_thread_name().to_owned();
        // Linux exposes thread names through /proc/self/task/*/comm with the
        // kernel's 15-byte task-name limit. Once a full test process has
        // created enough runtimes, `datum-timer-1000` is reported as
        // `datum-timer-100`, so compare against the Linux-visible name and use
        // the observed baseline for collision-safe assertions.
        let linux_thread_name = thread_name.chars().take(15).collect::<String>();
        assert!(wait_until(Duration::from_secs(5), || {
            materializer.timer_driver_is_live() && linux_thread_count(&linux_thread_name) >= 1
        }));
        let live_timer_threads = linux_thread_count(&linux_thread_name);

        for _ in 0..128 {
            materializer.schedule_once(Duration::from_secs(60), || {});
        }

        assert!(
            wait_until(Duration::from_secs(5), || {
                materializer.timer_driver_is_live()
                    && linux_thread_count(&linux_thread_name) == live_timer_threads
            }),
            "scheduling timers should not create extra timer threads for a runtime",
        );
        materializer.shutdown();
        assert!(wait_until(Duration::from_secs(5), || {
            !materializer.timer_driver_is_live()
                && linux_thread_count(&linux_thread_name) < live_timer_threads
        }));
    }

    #[test]
    fn cancelled_and_never_sinks_have_distinct_materialization_results() {
        assert_eq!(
            Source::repeat(1)
                .run_with(Sink::cancelled())
                .expect("cancelled sink materializes"),
            NotUsed
        );
        assert_eq!(
            Source::single(1)
                .run_with(Sink::never())
                .expect("never sink materializes")
                .try_wait(),
            None
        );
    }

    #[test]
    fn never_sink_finishes_on_materializer_shutdown() {
        let materializer = Materializer::new();
        let completion = Source::single(1)
            .run_with_materializer(Sink::never(), &materializer)
            .unwrap();

        materializer.shutdown();
        assert_eq!(completion.wait(), Err(StreamError::AbruptTermination));
    }

    #[test]
    fn dropping_source_never_completion_releases_parked_worker() {
        let materializer = Materializer::new();
        let completion = Source::<i32>::never()
            .run_with_materializer(Sink::ignore(), &materializer)
            .unwrap();

        assert!(wait_until(StdDuration::from_secs(1), || {
            materializer.active_streams() == 1
        }));
        assert_eq!(materializer.active_streams(), 1);

        drop(completion);

        assert!(wait_until(StdDuration::from_secs(15), || {
            materializer.active_streams() == 0
        }));
        assert_eq!(materializer.active_streams(), 0);
    }

    #[test]
    fn future_and_maybe_sources_emit_values() {
        let future_value = Source::future(|| async { Ok(7) }).run_collect().unwrap();
        assert_eq!(future_value, vec![7]);

        let future_source = Source::future_source(|| async { Ok(Source::from_iter([1, 2, 3])) })
            .run_collect()
            .unwrap();
        assert_eq!(future_source, vec![1, 2, 3]);

        let (handle, source) = Source::maybe();
        assert_eq!(
            source.clone().run_collect(),
            Err(StreamError::MaybeIncomplete)
        );
        handle.complete(9).unwrap();
        assert_eq!(source.run_collect().unwrap(), vec![9]);
    }

    #[test]
    fn wp6b_source_generators_emit_and_fail_like_stream_errors() {
        assert_eq!(
            Source::cycle(|| [1, 2, 3].into_iter())
                .take(8)
                .run_collect()
                .unwrap(),
            vec![1, 2, 3, 1, 2, 3, 1, 2]
        );
        assert_eq!(
            Source::<i32>::cycle(std::iter::empty::<i32>).run_collect(),
            Err(StreamError::Failed("empty iterator".into()))
        );
        assert_eq!(
            Source::unfold(0, |state| (state < 4).then_some((state + 1, state)))
                .run_collect()
                .unwrap(),
            vec![0, 1, 2, 3]
        );
        assert_eq!(
            Source::unfold_async(0, |state| async move {
                Ok((state < 4).then_some((state + 1, state * 2)))
            })
            .run_collect()
            .unwrap(),
            vec![0, 2, 4, 6]
        );
        assert!(matches!(
            Source::<i32>::lazy_single(|| panic!("boom")).run_collect(),
            Err(StreamError::Failed(message)) if message == "lazy_single factory panicked"
        ));
    }

    #[test]
    fn wp6b_lazy_sources_defer_until_first_pull_and_complete_deferred_mat() {
        let created = StdArc::new(StdAtomicUsize::new(0));
        let created_for_source = StdArc::clone(&created);
        let source = Source::<i32>::lazy_source(move || {
            created_for_source.fetch_add(1, StdOrdering::SeqCst);
            Source::from_iter([7, 8]).map_materialized_value(|_| 99)
        });
        let materializer = Materializer::new();
        let (mut stream, mut mat) = StdArc::clone(&source.factory)
            .create(&materializer)
            .unwrap();

        assert_eq!(created.load(StdOrdering::SeqCst), 0);
        assert!(mat.try_wait().is_none());
        assert_eq!(stream.next().unwrap().unwrap(), 7);
        assert_eq!(mat.wait().unwrap(), 99);
        assert_eq!(created.load(StdOrdering::SeqCst), 1);
        assert_eq!(stream.next().unwrap().unwrap(), 8);

        let never_created = StdArc::new(StdAtomicUsize::new(0));
        let never_created_for_source = StdArc::clone(&never_created);
        let mat = Source::<i32>::lazy_future_source(move || {
            never_created_for_source.fetch_add(1, StdOrdering::SeqCst);
            async { Ok(Source::single(1)) }
        })
        .to(Sink::cancelled())
        .run()
        .unwrap();
        assert!(matches!(mat.wait(), Err(StreamError::Failed(_))));
        assert_eq!(never_created.load(StdOrdering::SeqCst), 0);

        let lazy_future = StdArc::new(StdAtomicUsize::new(0));
        let lazy_future_for_source = StdArc::clone(&lazy_future);
        let source = Source::lazy_future(move || {
            lazy_future_for_source.fetch_add(1, StdOrdering::SeqCst);
            async { Ok(42) }
        });
        let (mut stream, _) = StdArc::clone(&source.factory)
            .create(&Materializer::new())
            .unwrap();
        assert_eq!(lazy_future.load(StdOrdering::SeqCst), 0);
        assert_eq!(stream.next().unwrap().unwrap(), 42);
        assert_eq!(lazy_future.load(StdOrdering::SeqCst), 1);
    }

    #[test]
    fn wp6b_unfold_resource_closes_on_completion_failure_and_cancellation() {
        let closed = StdArc::new(StdAtomicUsize::new(0));
        let closed_on_complete = StdArc::clone(&closed);
        let values = Source::unfold_resource(
            || Ok(std::collections::VecDeque::from([1, 2, 3])),
            |items| Ok(items.pop_front()),
            move |_items| {
                closed_on_complete.fetch_add(1, StdOrdering::SeqCst);
                Ok(())
            },
        )
        .run_collect()
        .unwrap();
        assert_eq!(values, vec![1, 2, 3]);
        assert_eq!(closed.load(StdOrdering::SeqCst), 1);

        let closed_on_failure = StdArc::new(StdAtomicUsize::new(0));
        let closed_on_failure_for_close = StdArc::clone(&closed_on_failure);
        let failed = Source::<i32>::unfold_resource(
            || Ok(()),
            |_| Err(StreamError::Failed("read".into())),
            move |_| {
                closed_on_failure_for_close.fetch_add(1, StdOrdering::SeqCst);
                Err(StreamError::Failed("close".into()))
            },
        )
        .run_collect();
        assert_eq!(failed, Err(StreamError::Failed("read".into())));
        assert_eq!(closed_on_failure.load(StdOrdering::SeqCst), 1);

        let closed_on_cancel = StdArc::new(StdAtomicUsize::new(0));
        let closed_on_cancel_for_close = StdArc::clone(&closed_on_cancel);
        let first = Source::unfold_resource(
            || Ok(0_usize),
            |next| {
                let item = *next;
                *next += 1;
                Ok(Some(item))
            },
            move |_| {
                closed_on_cancel_for_close.fetch_add(1, StdOrdering::SeqCst);
                Ok(())
            },
        )
        .run_with(Sink::head())
        .unwrap();
        assert_eq!(first.wait().unwrap(), 0);
        assert!(wait_until(Duration::from_millis(250), || {
            closed_on_cancel.load(StdOrdering::SeqCst) == 1
        }));
    }

    #[test]
    fn wp6b_async_resource_and_async_accumulators_are_sequential() {
        let closed = StdArc::new(StdAtomicUsize::new(0));
        let closed_for_close = StdArc::clone(&closed);
        let values = Source::unfold_resource_async(
            || async { Ok(std::collections::VecDeque::from([1, 2, 3])) },
            |items| {
                let item = items.pop_front();
                async move { Ok(item) }
            },
            move |_items| {
                let closed = StdArc::clone(&closed_for_close);
                async move {
                    closed.fetch_add(1, StdOrdering::SeqCst);
                    Ok(())
                }
            },
        )
        .run_collect()
        .unwrap();
        assert_eq!(values, vec![1, 2, 3]);
        assert_eq!(closed.load(StdOrdering::SeqCst), 1);

        let closed_on_failure = StdArc::new(StdAtomicUsize::new(0));
        let closed_on_failure_for_close = StdArc::clone(&closed_on_failure);
        let failed = Source::<i32>::unfold_resource_async(
            || async { Ok(()) },
            |_resource| async { Err(StreamError::Failed("read".into())) },
            move |_resource| {
                let closed_on_failure = StdArc::clone(&closed_on_failure_for_close);
                async move {
                    closed_on_failure.fetch_add(1, StdOrdering::SeqCst);
                    Err(StreamError::Failed("close".into()))
                }
            },
        )
        .run_collect();
        assert_eq!(failed, Err(StreamError::Failed("read".into())));
        assert_eq!(closed_on_failure.load(StdOrdering::SeqCst), 1);

        let active = StdArc::new(StdAtomicUsize::new(0));
        let max_active = StdArc::new(StdAtomicUsize::new(0));
        let active_for_stage = StdArc::clone(&active);
        let max_for_stage = StdArc::clone(&max_active);
        let scanned = Source::from_iter(1..=4)
            .scan_async(0, move |acc, item| {
                let active = StdArc::clone(&active_for_stage);
                let max_active = StdArc::clone(&max_for_stage);
                async move {
                    let now = active.fetch_add(1, StdOrdering::SeqCst) + 1;
                    max_active.fetch_max(now, StdOrdering::SeqCst);
                    tokio::time::sleep(Duration::from_millis(1)).await;
                    active.fetch_sub(1, StdOrdering::SeqCst);
                    Ok(acc + item)
                }
            })
            .run_collect()
            .unwrap();
        assert_eq!(scanned, vec![0, 1, 3, 6, 10]);
        assert_eq!(max_active.load(StdOrdering::SeqCst), 1);

        let folded = Source::from_iter(1..=4)
            .fold_async(0, |acc, item| async move { Ok(acc + item) })
            .run_collect()
            .unwrap();
        assert_eq!(folded, vec![10]);
    }

    #[test]
    fn wp6b_fold_async_materialization_does_not_drain_upstream() {
        let release = StdArc::new((std::sync::Mutex::new(false), std::sync::Condvar::new()));
        let started = StdArc::new(StdAtomicBool::new(false));
        let source = {
            let release = StdArc::clone(&release);
            let started = StdArc::clone(&started);
            Source::from_factory(move || {
                let release = StdArc::clone(&release);
                let started = StdArc::clone(&started);
                let mut emitted = false;
                Box::new(std::iter::from_fn(move || {
                    if emitted {
                        return None;
                    }
                    emitted = true;
                    started.store(true, StdOrdering::SeqCst);
                    let (released, available) = &*release;
                    let mut released = released.lock().unwrap();
                    while !*released {
                        released = available.wait(released).unwrap();
                    }
                    Some(Ok(1))
                }))
            })
        };

        let (materialized_tx, materialized_rx) = mpsc::channel();
        let join = thread::spawn(move || {
            let queue = source
                .fold_async(0, |acc, item| async move { Ok(acc + item) })
                .run_with(Sink::queue())
                .unwrap();
            materialized_tx.send(queue).unwrap();
        });

        let queue = match materialized_rx.recv_timeout(StdDuration::from_secs(1)) {
            Ok(queue) => queue,
            Err(error) => {
                let (released, available) = &*release;
                *released.lock().unwrap() = true;
                available.notify_all();
                let _ = join.join();
                panic!("fold_async materialization did not return before first pull: {error}");
            }
        };
        let (released, _) = &*release;
        assert!(
            !*released.lock().unwrap(),
            "test source was released before materialization returned"
        );

        let (released, available) = &*release;
        *released.lock().unwrap() = true;
        available.notify_all();
        assert_eq!(queue.pull().unwrap(), Some(1));
        assert_eq!(queue.pull().unwrap(), None);
        join.join().unwrap();
        assert!(started.load(StdOrdering::SeqCst));
    }

    #[test]
    fn wp6b_lazy_sink_and_flow_wait_for_first_element() {
        let lazy_sink_created = StdArc::new(StdAtomicUsize::new(0));
        let lazy_sink_created_for_factory = StdArc::clone(&lazy_sink_created);
        let empty_sink = Source::<i32>::empty()
            .run_with(Sink::lazy_sink(move || {
                lazy_sink_created_for_factory.fetch_add(1, StdOrdering::SeqCst);
                Sink::ignore()
            }))
            .unwrap();
        assert!(matches!(empty_sink.wait(), Err(StreamError::Failed(_))));
        assert_eq!(lazy_sink_created.load(StdOrdering::SeqCst), 0);

        let foreach_sum = StdArc::new(StdAtomicUsize::new(0));
        let foreach_sum_for_sink = StdArc::clone(&foreach_sum);
        Source::from_iter([1_usize, 2, 3])
            .run_with(Sink::foreach_async(2, move |item| {
                let foreach_sum = StdArc::clone(&foreach_sum_for_sink);
                async move {
                    foreach_sum.fetch_add(item, StdOrdering::SeqCst);
                    Ok(())
                }
            }))
            .unwrap()
            .wait()
            .unwrap();
        assert_eq!(foreach_sum.load(StdOrdering::SeqCst), 6);

        let lazy_flow_created = StdArc::new(StdAtomicUsize::new(0));
        let lazy_flow_created_for_factory = StdArc::clone(&lazy_flow_created);
        let lazy_flow = Flow::<i32, i32>::lazy_flow(move || {
            lazy_flow_created_for_factory.fetch_add(1, StdOrdering::SeqCst);
            Flow::identity()
                .map(|item: i32| item + 10)
                .map_materialized_value(|_| 123)
        });
        let mat = (lazy_flow.materialize)().unwrap();
        let mut stream = match lazy_flow.transform {
            flow::FlowTransform::Runtime(transform) => {
                transform(Box::new([Ok(1), Ok(2)].into_iter()), &Materializer::new()).unwrap()
            }
            flow::FlowTransform::Pure(_) => panic!("lazy flow must be runtime-backed"),
        };
        assert_eq!(lazy_flow_created.load(StdOrdering::SeqCst), 0);
        assert_eq!(stream.next().unwrap().unwrap(), 11);
        assert_eq!(mat.wait().unwrap(), 123);
        assert_eq!(lazy_flow_created.load(StdOrdering::SeqCst), 1);
        assert_eq!(stream.next().unwrap().unwrap(), 12);

        let future_flow = Source::from_iter([1, 2])
            .via_mat(
                Flow::future_flow(|| async {
                    Ok(Flow::identity()
                        .map(|item: i32| item * 2)
                        .map_materialized_value(|_| 77))
                }),
                Keep::right,
            )
            .to_mat(Sink::collect(), Keep::both)
            .run()
            .unwrap();
        assert_eq!(future_flow.0.wait().unwrap(), 77);
        assert_eq!(future_flow.1.wait().unwrap(), vec![2, 4]);
    }

    #[test]
    fn wp6b_lazy_flow_double_use_in_one_chain_pairs_instances_in_order() {
        // Order-sensitive pin for the per-thread FIFO slot pairing: one
        // cloned lazy-flow blueprint used twice in the SAME chain on the
        // same thread. element = first_stage_id * 100 + second_stage_id, so
        // a cross-wired (swapped) mat pairing would yield id2 * 100 + id1.
        for round in 0..50 {
            let counter = StdArc::new(StdAtomicUsize::new(1));
            let factory_counter = StdArc::clone(&counter);
            let lazy: Flow<usize, usize, _> = Flow::lazy_flow(move || {
                let id = factory_counter.fetch_add(1, StdOrdering::SeqCst);
                Flow::identity()
                    .map(move |x: usize| x * 100 + id)
                    .map_materialized_value(move |_| id)
            });
            let lazy_again = lazy.clone();

            let ((first_mat, second_mat), out) = Source::from_iter([0usize])
                .via_mat(lazy, Keep::right)
                .via_mat(lazy_again, Keep::both)
                .to_mat(Sink::collect(), Keep::both)
                .run()
                .unwrap();

            let first_id = first_mat.wait().unwrap();
            let second_id = second_mat.wait().unwrap();
            let element = out.wait().unwrap()[0];
            assert_eq!(
                element,
                first_id * 100 + second_id,
                "round {round}: mats ({first_id},{second_id}) cross-wired with transform order"
            );
            assert_ne!(
                first_id, second_id,
                "round {round}: same factory instance paired twice"
            );
        }
    }

    #[test]
    fn wp6b_lazy_flow_clones_materialize_concurrently_without_cross_wiring() {
        for _ in 0..20 {
            let next_id = StdArc::new(StdAtomicUsize::new(0));
            let next_id_for_factory = StdArc::clone(&next_id);
            let flow = Flow::<i32, i32>::lazy_flow(move || {
                let id = next_id_for_factory.fetch_add(1, StdOrdering::SeqCst) + 1;
                Flow::identity()
                    .map(move |item: i32| item + (id as i32 * 100))
                    .map_materialized_value(move |_| id)
            });
            let barrier = StdArc::new(std::sync::Barrier::new(3));

            let spawn_materialization = |input: i32| {
                let flow = flow.clone();
                let barrier = StdArc::clone(&barrier);
                thread::spawn(move || {
                    barrier.wait();
                    let (mat, values) = Source::single(input)
                        .via_mat(flow, Keep::right)
                        .to_mat(Sink::collect(), Keep::both)
                        .run()
                        .unwrap();
                    (input, mat.wait().unwrap(), values.wait().unwrap())
                })
            };

            let first = spawn_materialization(1);
            let second = spawn_materialization(2);
            barrier.wait();

            for result in [first.join().unwrap(), second.join().unwrap()] {
                let (input, mat_id, values) = result;
                assert_eq!(values, vec![input + (mat_id as i32 * 100)]);
            }
            assert_eq!(next_id.load(StdOrdering::SeqCst), 2);
        }
    }

    #[test]
    fn wp6b_map_with_resource_emits_close_item_before_terminal_error() {
        let queue = Source::from_factory(|| {
            Box::new(vec![Ok(1), Err(StreamError::Failed("upstream".into()))].into_iter())
        })
        .map_with_resource(
            || Ok(()),
            |_resource, item| Ok(item + 10),
            |_resource| Ok(Some(99)),
        )
        .run_with(Sink::queue())
        .unwrap();

        assert_eq!(queue.pull().unwrap(), Some(11));
        assert_eq!(queue.pull().unwrap(), Some(99));
        assert_eq!(queue.pull(), Err(StreamError::Failed("upstream".into())));

        let failed: StreamResult<Vec<i32>> = Source::single(1)
            .map_with_resource(
                || Ok(()),
                |_resource, _item| -> StreamResult<i32> { Err(StreamError::Failed("map".into())) },
                |_resource| -> StreamResult<Option<i32>> {
                    Err(StreamError::Failed("close".into()))
                },
            )
            .run_collect();
        assert_eq!(failed, Err(StreamError::Failed("map".into())));
    }

    #[test]
    fn stateful_and_terminal_source_operators_work() {
        let stateful = Source::from_iter([1, 2, 3])
            .stateful_map(0, |sum, item| {
                *sum += item;
                *sum
            })
            .run_collect()
            .unwrap();
        assert_eq!(stateful, vec![1, 3, 6]);

        let concat = Source::from_iter([1, 2, 3])
            .stateful_map_concat(0, |sum, item| {
                *sum += item;
                [item, *sum]
            })
            .run_collect()
            .unwrap();
        assert_eq!(concat, vec![1, 1, 2, 3, 3, 6]);

        assert_eq!(
            Source::from_iter([1, 2, 3])
                .fold(10, |acc, item| acc + item)
                .run_collect()
                .unwrap(),
            vec![16]
        );
        assert_eq!(
            Source::from_iter([1, 2, 3])
                .reduce(|acc, item| acc + item)
                .run_collect()
                .unwrap(),
            vec![6]
        );
    }

    #[test]
    fn concat_and_sliding_emit_before_unbounded_upstream_finishes() {
        let concat = Source::single(())
            .map_concat(|_| 0_u64..)
            .take(1)
            .run_collect()
            .unwrap();
        assert_eq!(concat, vec![0]);

        let sliding = Source::repeat(1_u64)
            .sliding(2, 1)
            .take(1)
            .run_collect()
            .unwrap();
        assert_eq!(sliding, vec![vec![1, 1]]);
    }

    #[test]
    fn fan_in_source_operators_follow_ordering_rules() {
        assert_eq!(
            Source::from_iter([1, 2])
                .concat(Source::from_iter([3, 4]))
                .run_collect()
                .unwrap(),
            vec![1, 2, 3, 4]
        );
        assert_eq!(
            Source::from_iter([3, 4])
                .prepend(Source::from_iter([1, 2]))
                .run_collect()
                .unwrap(),
            vec![1, 2, 3, 4]
        );
        assert_eq!(
            Source::empty()
                .or_else(Source::from_iter([10, 20]))
                .run_collect()
                .unwrap(),
            vec![10, 20]
        );
        assert_eq!(
            Source::from_iter([1, 2])
                .or_else(Source::from_iter([10, 20]))
                .run_collect()
                .unwrap(),
            vec![1, 2]
        );
        assert_eq!(
            Source::from_iter([1, 2, 3])
                .interleave(Source::from_iter([10, 11, 12]), 2)
                .run_collect()
                .unwrap(),
            vec![1, 2, 10, 11, 3, 12]
        );
    }

    #[test]
    fn fan_in_flow_operators_compose_with_primary_stream() {
        let concat = Source::from_iter([1, 2])
            .via(Flow::identity().concat(Source::from_iter([3, 4])))
            .run_collect()
            .unwrap();
        assert_eq!(concat, vec![1, 2, 3, 4]);

        let prepend = Source::from_iter([3, 4])
            .via(Flow::identity().prepend(Source::from_iter([1, 2])))
            .run_collect()
            .unwrap();
        assert_eq!(prepend, vec![1, 2, 3, 4]);

        let interleave = Source::from_iter([1, 2, 3])
            .via(Flow::identity().interleave(Source::from_iter([10, 11, 12]), 1))
            .run_collect()
            .unwrap();
        assert_eq!(interleave, vec![1, 10, 2, 11, 3, 12]);

        let merge_sorted = Source::from_iter([1, 4])
            .via(Flow::identity().merge_sorted(Source::from_iter([2, 3, 5])))
            .run_collect()
            .unwrap();
        assert_eq!(merge_sorted, vec![1, 2, 3, 4, 5]);

        let zip_latest = Source::from_iter([1, 2])
            .via(Flow::identity().zip_latest(Source::single(10)))
            .run_collect()
            .unwrap();
        assert_eq!(zip_latest, vec![(1, 10), (2, 10)]);

        let zip_latest_with = Source::from_iter([1, 2])
            .via(
                Flow::identity()
                    .zip_latest_with(Source::single(10), false, |left, right| left + right),
            )
            .run_collect()
            .unwrap();
        assert_eq!(zip_latest_with, vec![11, 12]);
    }

    #[test]
    fn fan_in_operators_propagate_errors_and_eager_close() {
        assert!(matches!(
            Source::failed(StreamError::Failed("boom".into()))
                .or_else(Source::from_iter([1, 2]))
                .run_collect(),
            Err(StreamError::Failed(_))
        ));
        assert!(matches!(
            Source::from_iter([1, 2])
                .prepend(Source::failed(StreamError::Failed("boom".into())))
                .run_collect(),
            Err(StreamError::Failed(_))
        ));
        assert_eq!(
            Source::from_iter([1, 2])
                .interleave_all([Source::empty()], 1, true)
                .run_collect()
                .unwrap(),
            vec![1]
        );
    }

    #[test]
    fn interleave_lazy_pulls_only_inputs_needed_for_first_segment() {
        use std::sync::{Arc, atomic::AtomicUsize, atomic::Ordering};

        let pulls: Arc<[AtomicUsize; 3]> = Arc::new([
            AtomicUsize::new(0),
            AtomicUsize::new(0),
            AtomicUsize::new(0),
        ]);

        let make_source = |idx: usize| {
            let pulls = Arc::clone(&pulls);
            Source::from_materialized_factory(move |_| {
                let pulls = Arc::clone(&pulls);
                let mut emitted = false;
                Ok((
                    Box::new(std::iter::from_fn(move || {
                        pulls[idx].fetch_add(1, Ordering::SeqCst);
                        if !emitted && idx == 0 {
                            emitted = true;
                            Some(Ok(42))
                        } else {
                            None
                        }
                    })) as BoxStream<i32>,
                    NotUsed,
                ))
            })
        };

        let result = make_source(0)
            .interleave_all([make_source(1), make_source(2)], 1, false)
            .run_with(Sink::head());

        assert_eq!(wait(result.unwrap()), 42);
        assert_eq!(pulls[0].load(Ordering::SeqCst), 1);
        assert_eq!(
            pulls[1].load(Ordering::SeqCst),
            0,
            "second input should not be pulled when downstream cancels after first element"
        );
        assert_eq!(
            pulls[2].load(Ordering::SeqCst),
            0,
            "third input should not be pulled before its turn"
        );
    }

    #[test]
    fn interleave_non_eager_drains_remaining_when_one_input_completes() {
        assert_eq!(
            Source::from_iter([1, 2, 3, 4])
                .interleave_all(
                    [Source::from_iter([10]), Source::from_iter([20, 21, 22])],
                    1,
                    false
                )
                .run_collect()
                .unwrap(),
            vec![1, 10, 20, 2, 21, 3, 22, 4]
        );
    }

    #[test]
    fn remaining_merge_and_zip_family_matches_expected_ordering() {
        assert_eq!(
            Source::from_iter([1, 4])
                .merge_sorted(Source::from_iter([2, 3, 5]))
                .run_collect()
                .unwrap(),
            vec![1, 2, 3, 4, 5]
        );

        assert_eq!(
            Source::from_iter([1, 2])
                .merge_latest(Source::single(10), false)
                .run_collect()
                .unwrap(),
            vec![vec![1, 10], vec![2, 10]]
        );

        assert_eq!(
            Source::from_iter([1, 2, 3])
                .merge_all([Source::from_iter([10, 11])], false)
                .run_collect()
                .unwrap(),
            vec![1, 10, 2, 11, 3]
        );

        assert_eq!(
            Source::from_iter([1, 2, 3])
                .zip_with(Source::from_iter([10, 11, 12]), |left, right| left + right)
                .run_collect()
                .unwrap(),
            vec![11, 13, 15]
        );

        assert_eq!(
            Source::from_iter([1, 2])
                .zip_latest(Source::single(10))
                .run_collect()
                .unwrap(),
            vec![(1, 10), (2, 10)]
        );

        assert_eq!(
            Source::from_iter([1, 2, 3])
                .zip_latest_with(Source::from_iter([10]), false, |left, right| left + right)
                .run_collect()
                .unwrap(),
            vec![11, 12, 13]
        );

        assert_eq!(
            Source::from_iter([1, 2])
                .zip_all(Source::from_iter([10, 11, 12]), -1, -2)
                .run_collect()
                .unwrap(),
            vec![(1, 10), (2, 11), (-1, 12)]
        );

        assert_eq!(
            Source::from_iter([5, 6, 7])
                .zip_with_index()
                .run_collect()
                .unwrap(),
            vec![(5, 0), (6, 1), (7, 2)]
        );

        assert_eq!(
            Source::zip_n([Source::from_iter([1, 2]), Source::from_iter([10, 20])])
                .run_collect()
                .unwrap(),
            vec![vec![1, 10], vec![2, 20]]
        );

        assert_eq!(
            Source::zip_with_n(
                [
                    Source::from_iter([1, 2]),
                    Source::from_iter([10, 20]),
                    Source::from_iter([100, 200]),
                ],
                |values| values.into_iter().sum::<i32>(),
            )
            .run_collect()
            .unwrap(),
            vec![111, 222]
        );

        assert_eq!(
            Source::merge_prioritized_n(
                [
                    (Source::from_iter([1, 2, 3, 4]), 2),
                    (Source::from_iter([10, 11]), 1),
                ],
                false,
            )
            .run_collect()
            .unwrap(),
            vec![1, 2, 10, 3, 4, 11]
        );

        assert_eq!(
            Source::combine(
                Source::from_iter([1, 2, 3]),
                Source::from_iter([10, 11]),
                std::iter::empty::<Source<i32, NotUsed>>(),
                SourceCombineStrategy::Merge {
                    eager_complete: false,
                },
            )
            .run_collect()
            .unwrap(),
            vec![1, 10, 2, 11, 3]
        );

        let combined_sink = Sink::combine(
            Sink::ignore(),
            Sink::ignore(),
            std::iter::empty::<Sink<i32, NotUsed>>(),
            SinkCombineStrategy::Broadcast,
        );
        assert_eq!(
            Source::from_iter([1, 2, 3])
                .run_with(combined_sink)
                .unwrap(),
            NotUsed
        );
    }

    #[test]
    fn sink_combine_broadcast_delivers_every_element_to_every_child() {
        // Regression: the combined children's materialized values were
        // dropped at materialization, tripping cancel-on-drop and silently
        // cancelling every child before it consumed anything (0 deliveries).
        let first_count = StdArc::new(StdAtomicUsize::new(0));
        let second_count = StdArc::new(StdAtomicUsize::new(0));
        let first_counter = StdArc::clone(&first_count);
        let second_counter = StdArc::clone(&second_count);
        let combined = Sink::combine(
            Sink::foreach(move |_: i32| {
                first_counter.fetch_add(1, StdOrdering::SeqCst);
            }),
            Sink::foreach(move |_: i32| {
                second_counter.fetch_add(1, StdOrdering::SeqCst);
            }),
            std::iter::empty::<Sink<i32, NotUsed>>(),
            SinkCombineStrategy::Broadcast,
        );
        assert_eq!(
            Source::from_iter(0..100).run_with(combined).unwrap(),
            NotUsed
        );
        // The rendezvous channels guarantee each child has received every
        // element before the parent send returns. The child callback may still
        // be one scheduler tick behind the parent, so assert with a bounded
        // condition wait instead of an immediate load.
        assert!(wait_until(StdDuration::from_secs(1), || {
            first_count.load(StdOrdering::SeqCst) == 100
                && second_count.load(StdOrdering::SeqCst) == 100
        }));
    }

    #[test]
    fn zip_latest_completes_when_one_side_finishes_without_emitting() {
        // Regression: with eager_complete = false, an empty side left
        // `latest = None` forever while the loop drained the other side —
        // an infinite busy-loop when that side is unbounded.
        assert_eq!(
            Source::from_iter(std::iter::empty::<i32>())
                .zip_latest_with(Source::repeat(10), false, |left, right| left + right)
                .run_collect()
                .unwrap(),
            Vec::<i32>::new()
        );
        assert_eq!(
            Source::repeat(10)
                .zip_latest_with(
                    Source::from_iter(std::iter::empty::<i32>()),
                    false,
                    |left, right| left + right,
                )
                .run_collect()
                .unwrap(),
            Vec::<i32>::new()
        );
    }

    #[test]
    fn zip_family_completion_boundaries_match_expected_results() {
        assert_eq!(
            Source::from_iter([1, 2, 3])
                .zip_with(Source::from_iter([10]), |left, right| left + right)
                .run_collect()
                .unwrap(),
            vec![11]
        );

        assert_eq!(
            Source::from_iter([1, 2, 3])
                .zip_latest_with(Source::from_iter([10]), true, |left, right| left + right)
                .run_collect()
                .unwrap(),
            vec![11, 12]
        );

        assert_eq!(
            Source::zip_n([
                Source::from_iter([1, 2, 3]),
                Source::from_iter([10]),
                Source::from_iter([100, 200, 300]),
            ])
            .run_collect()
            .unwrap(),
            vec![vec![1, 10, 100]]
        );
    }

    #[test]
    fn combine_strategies_follow_merge_concat_and_priority_rules() {
        assert_eq!(
            Source::combine(
                Source::from_iter([1, 2]),
                Source::from_iter([10, 11]),
                [Source::from_iter([100])],
                SourceCombineStrategy::Concat,
            )
            .run_collect()
            .unwrap(),
            vec![1, 2, 10, 11, 100]
        );

        assert_eq!(
            Source::combine(
                Source::from_iter([1, 2, 3, 4]),
                Source::from_iter([10, 11]),
                std::iter::empty::<Source<i32, NotUsed>>(),
                SourceCombineStrategy::Prioritized {
                    priorities: vec![2, 1],
                    eager_complete: false,
                },
            )
            .run_collect()
            .unwrap(),
            vec![1, 2, 10, 3, 4, 11]
        );
    }

    #[test]
    fn concat_lazy_defers_follow_on_source_until_needed() {
        let source_counter = StdArc::new(StdAtomicUsize::new(0));
        let source_counter_clone = StdArc::clone(&source_counter);
        let lazy_source = Source::from_materialized_factory(move |_| {
            source_counter_clone.fetch_add(1, StdOrdering::SeqCst);
            Ok((Box::new(std::iter::once(Ok(99))), NotUsed))
        });
        let source_head = Source::single(1)
            .concat_lazy(lazy_source)
            .run_with(Sink::head());
        assert_eq!(wait(source_head.unwrap()), 1);
        assert_eq!(source_counter.load(StdOrdering::SeqCst), 0);

        let flow_counter = StdArc::new(StdAtomicUsize::new(0));
        let flow_counter_clone = StdArc::clone(&flow_counter);
        let lazy_flow_source = Source::from_materialized_factory(move |_| {
            flow_counter_clone.fetch_add(1, StdOrdering::SeqCst);
            Ok((Box::new(std::iter::once(Ok(99))), NotUsed))
        });
        let flow_head = Source::single(1)
            .via(Flow::identity().concat_lazy(lazy_flow_source))
            .run_with(Sink::head());
        assert_eq!(wait(flow_head.unwrap()), 1);
        assert_eq!(flow_counter.load(StdOrdering::SeqCst), 0);
    }

    #[test]
    fn also_to_completes_when_side_sink_cancels() {
        assert_eq!(
            Source::from_iter([1, 2, 3])
                .also_to(Sink::cancelled())
                .run_collect()
                .unwrap(),
            Vec::<i32>::new()
        );
        assert_eq!(
            Source::from_iter([1, 2, 3])
                .also_to_all([Sink::cancelled(), Sink::cancelled()])
                .run_collect()
                .unwrap(),
            Vec::<i32>::new()
        );
    }

    #[test]
    fn also_to_completes_gracefully_when_side_sink_disconnects() {
        let result = Source::from_iter(0..100)
            .also_to(Sink::head())
            .run_collect()
            .unwrap();
        assert!(!result.is_empty(), "main should emit at least one element");
        assert!(
            result.len() < 100,
            "main should complete early when side disconnects"
        );
    }

    #[test]
    fn also_to_propagates_original_error_when_side_is_disconnected() {
        let err = StreamError::Failed("distinctive-boom".into());
        assert!(matches!(
            Source::<i32>::failed(err.clone())
                .also_to(Sink::cancelled())
                .run_collect(),
            Err(StreamError::Failed(msg)) if msg == "distinctive-boom"
        ));
        assert!(matches!(
            Source::<i32>::failed(err.clone())
                .also_to_all([Sink::cancelled()])
                .run_collect(),
            Err(StreamError::Failed(msg)) if msg == "distinctive-boom"
        ));
        assert!(matches!(
            Source::<i32>::failed(err)
                .divert_to(Sink::cancelled(), |_: &i32| true)
                .run_collect(),
            Err(StreamError::Failed(msg)) if msg == "distinctive-boom"
        ));
    }

    #[test]
    fn divert_to_routes_matching_elements_to_side_sink() {
        let diverted = Source::from_iter([1, 2, 3, 4])
            .divert_to(Sink::ignore(), |item| item % 2 == 0)
            .run_collect()
            .unwrap();
        assert_eq!(diverted, vec![1, 3]);
    }

    #[test]
    fn wire_tap_drops_when_side_sink_backpressures() {
        let tapped = Source::from_iter([1, 2, 3])
            .wire_tap(Sink::head())
            .run_collect()
            .unwrap();
        assert_eq!(tapped, vec![1, 2, 3]);

        let tapped_via_flow = Source::from_iter([1, 2, 3])
            .via(Flow::identity().wire_tap(Sink::head()))
            .run_collect()
            .unwrap();
        assert_eq!(tapped_via_flow, vec![1, 2, 3]);
    }

    #[test]
    fn async_mapping_variants_complete() {
        let ordered = Source::from_iter(0..4)
            .map_async(2, |item| async move { Ok(item * 2) })
            .run_collect()
            .unwrap();
        assert_eq!(ordered, vec![0, 2, 4, 6]);

        let unordered = Source::from_iter(0..4)
            .map_async_unordered(2, |item| async move { Ok(item * 2) })
            .run_collect()
            .unwrap();
        assert_eq!(unordered, vec![0, 2, 4, 6]);

        let partitioned = Source::from_iter(0..4)
            .map_async_partitioned(4, 1, |item| item % 2, |item| async move { Ok(item + 1) })
            .run_collect()
            .unwrap();
        assert_eq!(partitioned, vec![1, 2, 3, 4]);
    }

    #[test]
    fn map_async_ordered_bounds_pulls_behind_stuck_head() {
        let pulls = StdArc::new(StdAtomicUsize::new(0));
        let pulls_for_source = StdArc::clone(&pulls);
        let probe = Source::from_fn_iter(move || {
            let pulls = StdArc::clone(&pulls_for_source);
            std::iter::from_fn(move || {
                let next = pulls.fetch_add(1, StdOrdering::SeqCst);
                Some(next)
            })
        })
        .map_async(2, |item| async move {
            if item == 0 {
                tokio::time::sleep(StdDuration::from_millis(300)).await;
            }
            Ok(item)
        })
        .run_with(TestSink::probe())
        .unwrap();

        probe.request(16);
        thread::sleep(StdDuration::from_millis(100));
        assert!(
            pulls.load(StdOrdering::SeqCst) <= 3,
            "pulled {} elements with parallelism=2 behind a stuck ordered head",
            pulls.load(StdOrdering::SeqCst)
        );
    }

    #[test]
    fn async_mapping_parks_until_woken_future_completes() {
        struct WakeOnceFuture {
            value: Option<u64>,
            ready: StdArc<StdAtomicBool>,
            polls: StdArc<StdAtomicUsize>,
            poll_tx: mpsc::Sender<usize>,
            latest_waker: StdArc<Mutex<Option<std::task::Waker>>>,
        }

        impl std::future::Future for WakeOnceFuture {
            type Output = StreamResult<u64>;

            fn poll(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
                let this = self.as_mut().get_mut();
                let poll = this.polls.fetch_add(1, StdOrdering::SeqCst) + 1;
                if this.ready.load(StdOrdering::SeqCst) {
                    let _ = this.poll_tx.send(poll);
                    return Poll::Ready(Ok(this.value.take().unwrap()));
                }

                *this.latest_waker.lock().expect("latest wake slot mutex") =
                    Some(cx.waker().clone());
                let _ = this.poll_tx.send(poll);
                Poll::Pending
            }
        }

        let ready = StdArc::new(StdAtomicBool::new(false));
        let polls = StdArc::new(StdAtomicUsize::new(0));
        let latest_waker = StdArc::new(Mutex::new(None));
        let (poll_tx, poll_rx) = mpsc::channel();
        let (result_tx, result_rx) = mpsc::channel();
        let ready_for_stage = StdArc::clone(&ready);
        let polls_for_stage = StdArc::clone(&polls);
        let poll_tx_for_stage = poll_tx.clone();
        let latest_waker_for_stage = StdArc::clone(&latest_waker);
        let worker = thread::spawn(move || {
            let result = Source::single(41)
                .map_async(1, move |item| WakeOnceFuture {
                    value: Some(item + 1),
                    ready: StdArc::clone(&ready_for_stage),
                    polls: StdArc::clone(&polls_for_stage),
                    poll_tx: poll_tx_for_stage.clone(),
                    latest_waker: StdArc::clone(&latest_waker_for_stage),
                })
                .run_collect();
            let _ = result_tx.send(result);
        });

        assert_eq!(poll_rx.recv_timeout(StdDuration::from_secs(10)).unwrap(), 1);
        assert!(
            matches!(
                result_rx.try_recv(),
                Err(std::sync::mpsc::TryRecvError::Empty)
            ),
            "stream completed before the future was woken"
        );
        assert_eq!(poll_rx.recv_timeout(StdDuration::from_secs(10)).unwrap(), 2);
        assert!(
            matches!(
                result_rx.try_recv(),
                Err(std::sync::mpsc::TryRecvError::Empty)
            ),
            "stream completed before the future was marked ready"
        );

        ready.store(true, StdOrdering::SeqCst);
        latest_waker
            .lock()
            .expect("latest wake slot mutex")
            .take()
            .expect("pending future should have registered a waker")
            .wake();

        let values = result_rx
            .recv_timeout(StdDuration::from_secs(10))
            .expect("stream should complete after the pending future is woken")
            .unwrap();
        worker.join().expect("stream worker should not panic");
        assert_eq!(values, vec![42]);
        assert_eq!(
            polls.load(StdOrdering::SeqCst),
            3,
            "pending future should be polled once inline, once to register the waker, and once after wake"
        );
    }

    #[test]
    fn async_mapping_emits_before_unbounded_upstream_finishes() {
        let ordered = Source::repeat(1)
            .map_async(2, |item| async move { Ok(item + 1) })
            .take(1)
            .run_collect()
            .unwrap();
        assert_eq!(ordered, vec![2]);

        let unordered = Source::repeat(1)
            .map_async_unordered(2, |item| async move { Ok(item + 1) })
            .take(1)
            .run_collect()
            .unwrap();
        assert_eq!(unordered, vec![2]);

        let partitioned = Source::repeat(1)
            .map_async_partitioned(2, 1, |_| 0_u8, |item| async move { Ok(item + 1) })
            .take(1)
            .run_collect()
            .unwrap();
        assert_eq!(partitioned, vec![2]);
    }

    #[test]
    fn partitioned_async_mapping_limits_same_key_concurrency() {
        let active = StdArc::new(StdAtomicUsize::new(0));
        let max_active = StdArc::new(StdAtomicUsize::new(0));
        let active_for_stage = StdArc::clone(&active);
        let max_for_stage = StdArc::clone(&max_active);

        let values = Source::from_iter(0..6)
            .map_async_partitioned(
                4,
                1,
                |_| 0_u8,
                move |item| {
                    let active = StdArc::clone(&active_for_stage);
                    let max_active = StdArc::clone(&max_for_stage);
                    let current = active.fetch_add(1, StdOrdering::SeqCst) + 1;
                    max_active.fetch_max(current, StdOrdering::SeqCst);
                    async move {
                        thread::sleep(Duration::from_millis(1));
                        active.fetch_sub(1, StdOrdering::SeqCst);
                        Ok(item)
                    }
                },
            )
            .run_collect()
            .unwrap();

        assert_eq!(values, vec![0, 1, 2, 3, 4, 5]);
        assert_eq!(max_active.load(StdOrdering::SeqCst), 1);
    }

    #[test]
    fn partitioned_async_mapping_scans_past_blocked_pending_key() {
        let active = StdArc::new(StdAtomicUsize::new(0));
        let max_active = StdArc::new(StdAtomicUsize::new(0));
        let active_for_stage = StdArc::clone(&active);
        let max_for_stage = StdArc::clone(&max_active);
        let (release_tx, release_rx) = oneshot::channel::<()>();
        let release_rx = StdArc::new(std::sync::Mutex::new(Some(release_rx)));
        let release_rx_for_stage = StdArc::clone(&release_rx);
        let max_for_release = StdArc::clone(&max_active);

        let releaser = thread::spawn(move || {
            let deadline = Instant::now() + StdDuration::from_secs(1);
            while max_for_release.load(StdOrdering::SeqCst) < 2 && Instant::now() < deadline {
                thread::yield_now();
            }
            let _ = release_tx.send(());
        });

        let values = Source::from_iter([0, 2, 1])
            .map_async_partitioned(
                2,
                1,
                |item| item % 2,
                move |item| {
                    let active = StdArc::clone(&active_for_stage);
                    let max_active = StdArc::clone(&max_for_stage);
                    let release_rx = StdArc::clone(&release_rx_for_stage);
                    let current = active.fetch_add(1, StdOrdering::SeqCst) + 1;
                    max_active.fetch_max(current, StdOrdering::SeqCst);
                    async move {
                        if item == 0 {
                            let receiver = release_rx
                                .lock()
                                .expect("release receiver mutex")
                                .take()
                                .expect("release receiver present");
                            let _ = receiver.await;
                        }
                        active.fetch_sub(1, StdOrdering::SeqCst);
                        Ok(item)
                    }
                },
            )
            .run_collect()
            .unwrap();
        releaser.join().unwrap();

        assert_eq!(values, vec![0, 2, 1]);
        assert_eq!(max_active.load(StdOrdering::SeqCst), 2);
    }

    #[test]
    fn partitioned_async_mapping_p1_still_evaluates_partition() {
        let partitions = StdArc::new(StdAtomicUsize::new(0));
        let partitions_for_stage = StdArc::clone(&partitions);

        let values = Source::from_iter(0..8)
            .map_async_partitioned(
                1,
                1,
                move |item| {
                    partitions_for_stage.fetch_add(1, StdOrdering::SeqCst);
                    item % 2
                },
                |item| async move { Ok(item + 1) },
            )
            .run_collect()
            .unwrap();

        assert_eq!(values, (1..9).collect::<Vec<_>>());
        assert_eq!(partitions.load(StdOrdering::SeqCst), 8);
    }

    #[test]
    fn partitioned_async_mapping_handles_many_keys_high_parallelism() {
        let active_by_key =
            StdArc::new((0..16).map(|_| StdAtomicUsize::new(0)).collect::<Vec<_>>());
        let max_by_key = StdArc::new((0..16).map(|_| StdAtomicUsize::new(0)).collect::<Vec<_>>());
        let active_for_stage = StdArc::clone(&active_by_key);
        let max_for_stage = StdArc::clone(&max_by_key);

        let values = Source::from_iter(0..512_usize)
            .map_async_partitioned(
                32,
                1,
                |item| item % 16,
                move |item| {
                    let active = StdArc::clone(&active_for_stage);
                    let max_active = StdArc::clone(&max_for_stage);
                    let key = item % 16;
                    let current = active[key].fetch_add(1, StdOrdering::SeqCst) + 1;
                    max_active[key].fetch_max(current, StdOrdering::SeqCst);
                    async move {
                        active[key].fetch_sub(1, StdOrdering::SeqCst);
                        Ok(item)
                    }
                },
            )
            .run_collect()
            .unwrap();

        assert_eq!(values, (0..512).collect::<Vec<_>>());
        for max_active in max_by_key.iter() {
            assert_eq!(max_active.load(StdOrdering::SeqCst), 1);
        }
    }

    #[test]
    fn error_operators_map_recover_and_complete() {
        let mapped = Source::<i32>::failed(StreamError::Failed("boom".into()))
            .map_error(|_| StreamError::Failed("mapped".into()))
            .run_collect();
        assert_eq!(mapped, Err(StreamError::Failed("mapped".into())));

        let recovered = Source::<i32>::failed(StreamError::Failed("boom".into()))
            .recover(|error| match error {
                StreamError::Failed(_) => Some(42),
                _ => None,
            })
            .run_collect()
            .unwrap();
        assert_eq!(recovered, vec![42]);

        let unrecovered = Source::<i32>::failed(StreamError::Failed("original".into()))
            .recover(|_| None)
            .run_collect();
        assert_eq!(unrecovered, Err(StreamError::Failed("original".into())));

        let recovered_with = Source::<i32>::failed(StreamError::Failed("boom".into()))
            .recover_with_retries(1, |_| Some(Source::from_iter([1, 2])))
            .run_collect()
            .unwrap();
        assert_eq!(recovered_with, vec![1, 2]);

        let declined_recover_with = Source::<i32>::failed(StreamError::Failed("declined".into()))
            .recover_with_retries(1, |_| None)
            .run_collect();
        assert_eq!(
            declined_recover_with,
            Err(StreamError::Failed("declined".into()))
        );

        let completed = Source::from_factory(|| {
            Box::new(vec![Ok(1), Err(StreamError::Failed("ignored".into())), Ok(2)].into_iter())
        })
        .on_error_complete()
        .run_collect()
        .unwrap();
        assert_eq!(completed, vec![1]);
    }

    #[test]
    fn sliding_matches_akka_window_semantics() {
        // Full windows then stop; no spurious trailing partial windows.
        assert_eq!(
            Source::from_iter(1..=4)
                .sliding(3, 1)
                .run_collect()
                .unwrap(),
            vec![vec![1, 2, 3], vec![2, 3, 4]]
        );
        assert_eq!(
            Source::from_iter(1..=4)
                .sliding(2, 1)
                .run_collect()
                .unwrap(),
            vec![vec![1, 2], vec![2, 3], vec![3, 4]]
        );
        // Exact multiple of size leaves nothing to emit at finish.
        assert_eq!(
            Source::from_iter(1..=3)
                .sliding(3, 1)
                .run_collect()
                .unwrap(),
            vec![vec![1, 2, 3]]
        );
        // Short streams emit a single partial window.
        assert_eq!(
            Source::from_iter(1..=2)
                .sliding(3, 1)
                .run_collect()
                .unwrap(),
            vec![vec![1, 2]]
        );
        // Every element in its own window.
        assert_eq!(
            Source::from_iter(1..=3)
                .sliding(1, 1)
                .run_collect()
                .unwrap(),
            vec![vec![1], vec![2], vec![3]]
        );
        // step > size skips the gap between windows.
        assert_eq!(
            Source::from_iter(1..=6)
                .sliding(2, 3)
                .run_collect()
                .unwrap(),
            vec![vec![1, 2], vec![4, 5]]
        );
        // step > size with an in-progress trailing window is dropped (Akka).
        assert_eq!(
            Source::from_iter(1..=3)
                .sliding(2, 4)
                .run_collect()
                .unwrap(),
            vec![vec![1, 2]]
        );
    }

    #[test]
    fn recover_with_retries_indefinitely_like_akka() {
        let attempts = StdArc::new(StdAtomicUsize::new(0));
        let attempts_in_stage = StdArc::clone(&attempts);
        // The first several recovery sources keep failing; `recover_with` must
        // retry past Datum's old single-retry limit to reach the good source.
        let recovered = Source::<i32>::failed(StreamError::Failed("boom".into()))
            .recover_with(move |_error| {
                if attempts_in_stage.fetch_add(1, StdOrdering::SeqCst) < 5 {
                    Some(Source::<i32>::failed(StreamError::Failed("again".into())))
                } else {
                    Some(Source::from_iter([42]))
                }
            })
            .run_collect()
            .unwrap();
        assert_eq!(recovered, vec![42]);
        assert_eq!(attempts.load(StdOrdering::SeqCst), 6);
    }

    #[test]
    fn many_concurrent_streams_do_not_starve_the_pool() {
        // Regression: a fixed-size pool deadlocks once `num_cpus` streams each
        // monopolize a worker. The growing pool must keep serving new streams.
        //
        // Use a small fixed count rather than `available_parallelism() + 2`: the
        // point is that several concurrent never-completing streams cannot starve
        // a fresh one, and a fixed count keeps this from spawning a large number
        // of permanent threads (which, when the whole suite runs under load, can
        // exhaust the thread limit and wedge the shared executor).
        let materializer = Materializer::new();
        let busy = 6_usize;

        let mut held = Vec::with_capacity(busy);
        for _ in 0..busy {
            held.push(
                Source::single(1_u64)
                    .run_with_materializer(Sink::never(), &materializer)
                    .unwrap(),
            );
        }

        for _ in 0..400 {
            if materializer.active_streams() >= busy {
                break;
            }
            thread::sleep(Duration::from_millis(5));
        }
        assert_eq!(materializer.active_streams(), busy);

        // A fresh finite stream must still complete despite every prior worker
        // being occupied by a never-completing sink.
        let sum = Source::from_iter(0_u64..5)
            .run_with_materializer(Sink::fold(0_u64, |acc, item| acc + item), &materializer)
            .unwrap();
        assert_eq!(sum.wait().unwrap(), 10);

        materializer.shutdown();
        for completion in held {
            assert_eq!(completion.wait(), Err(StreamError::AbruptTermination));
        }
    }
}