datum/graph/

executor.rs

use super::*;

// ---------------------------------------------------------------------------
// ExecutorMode — Phase 0 scaffolding (WP-18)
// ---------------------------------------------------------------------------

/// Selects which executor is used to run a fused graph.
///
/// This is an internal test/diagnostic hook — it is **not** part of the public
/// API and must not be exposed through `pub use` or any user-facing surface.
///
/// * `Auto` — default; tries the typed flow plan first, falls back to the
///   erased executor for unsupported graphs.
/// * `ErasedOnly` — always runs the existing erased (`Box<dyn DatumElement>`)
///   executor regardless of graph shape.
/// * `TypedOnly` — always tries the typed flow plan; returns
///   `StreamError::GraphValidation("typed executor does not support this graph
///   shape")` when the plan cannot be built.  **Test/diagnostic only.**
#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
pub(crate) enum ExecutorMode {
    /// Try typed first; fall back to erased for unsupported shapes.
    #[default]
    Auto,
    /// Always use the erased executor.
    ErasedOnly,
    /// Always use typed; error if unsupported.  Test/diagnostic only.
    TypedOnly,
}

impl<In, Out> GraphBlueprint<FlowShape<In, Out>>
where
    In: Clone + Send + 'static,
    Out: Send + 'static,
{
    pub fn run_with_input<I>(&self, input: I) -> StreamResult<Vec<Out>>
    where
        I: IntoIterator<Item = In>,
    {
        Ok(self
            .run_with_input_report(input, FusedExecutionConfig::default())?
            .output)
    }

    pub fn run_with_input_report<I>(
        &self,
        input: I,
        config: FusedExecutionConfig,
    ) -> StreamResult<FusedExecutionReport<Out>>
    where
        I: IntoIterator<Item = In>,
    {
        self.run_with_input_report_mode(input, config, ExecutorMode::Auto)
    }

    pub(crate) fn run_with_input_report_mode<I>(
        &self,
        input: I,
        config: FusedExecutionConfig,
        mode: ExecutorMode,
    ) -> StreamResult<FusedExecutionReport<Out>>
    where
        I: IntoIterator<Item = In>,
    {
        // --- Typed path (Auto / TypedOnly) ---
        if mode != ExecutorMode::ErasedOnly {
            // Phase 2: linear typed flow plan (Identity/Map/AsyncBoundary chains).
            let linear_plan = try_typed_flow_plan::<In, Out>(
                &self.stages,
                &self.edges,
                self.shape.inlet().id(),
                self.shape.outlet().id(),
            );
            if let Some(plan) = linear_plan {
                let input = input.into_iter();
                let mut output = Vec::with_capacity(input.size_hint().0);
                let mut events = 0usize;
                let mut async_boundary_crossings = 0usize;
                for item in input {
                    let out =
                        plan.run_item(item, config, &mut events, &mut async_boundary_crossings)?;
                    output.push(out);
                }
                return Ok(FusedExecutionReport {
                    output,
                    events,
                    async_boundary_crossings,
                });
            }

            // Phase 3a: typed MergeSequence kernel (Unzip → MergeSequence topology).
            let ms_plan = try_typed_merge_sequence_plan::<In, Out>(
                &self.stages,
                &self.edges,
                self.shape.inlet().id(),
                self.shape.outlet().id(),
            );
            if let Some(mut plan) = ms_plan {
                let output = run_typed_merge_sequence(&mut plan, input)?;
                return Ok(FusedExecutionReport {
                    output,
                    events: 0,
                    async_boundary_crossings: 0,
                });
            }

            // Phase 3b: typed MergeLatest kernel (Unzip → MergeLatest topology).
            //
            // Build a FRESH TypedMergeLatestPlan per run — no cached mutable state,
            // no Mutex.  Blueprint reuse and concurrent runs are fully independent.
            // The per-run cost (topology/type checks, a few Arc clones, one small
            // MergeLatestCore alloc, and one Box for the output Vec) is negligible
            // vs a 10k-element run.
            //
            // try_build_typed_merge_latest_dispatch checks topology and element
            // TypeId without consuming input; if it returns Some(runner), we then
            // consume input for the typed path.  Custom `T` not in the bounded
            // 17-type set returns None → falls back to the erased executor (Auto)
            // or errors (TypedOnly).
            let inlet_id = self.shape.inlet().id();
            let outlet_id = self.shape.outlet().id();
            if let Some(runner) = try_build_typed_merge_latest_dispatch::<In, Out>(
                &self.stages,
                &self.edges,
                inlet_id,
                outlet_id,
            ) {
                let mut input_iter = input.into_iter();
                let output = runner(&mut input_iter)?;
                return Ok(FusedExecutionReport {
                    output,
                    events: 0,
                    async_boundary_crossings: 0,
                });
            }

            if mode == ExecutorMode::TypedOnly {
                return Err(StreamError::GraphValidation(
                    "typed executor does not support this graph shape".into(),
                ));
            }
        }

        // --- Erased fallback ---
        let input = input.into_iter();
        let mut executor = FusedExecutor::new(self, config);
        let inlet = self.shape.inlet().id();
        let outlet = self.shape.outlet().id();
        let mut output = Vec::with_capacity(input.size_hint().0);

        {
            let mut output_sink = VecOutputSink {
                output: &mut output,
            };
            executor.request(outlet, outlet, &mut output_sink)?;
            for item in input {
                executor.deliver(inlet, datum(item), outlet, &mut output_sink)?;
            }
            executor.complete(inlet, outlet, &mut output_sink)?;
        }

        Ok(FusedExecutionReport {
            output,
            events: executor.events,
            async_boundary_crossings: executor.async_boundary_crossings,
        })
    }

    pub fn run_count_with_input<I>(&self, input: I) -> StreamResult<usize>
    where
        I: IntoIterator<Item = In>,
    {
        Ok(self
            .run_count_with_input_report(input, FusedExecutionConfig::default())?
            .result)
    }

    pub fn run_count_with_input_report<I>(
        &self,
        input: I,
        config: FusedExecutionConfig,
    ) -> StreamResult<FusedTerminalReport<usize>>
    where
        I: IntoIterator<Item = In>,
    {
        self.run_count_with_input_report_mode(input, config, ExecutorMode::Auto)
    }

    pub(crate) fn run_count_with_input_report_mode<I>(
        &self,
        input: I,
        config: FusedExecutionConfig,
        mode: ExecutorMode,
    ) -> StreamResult<FusedTerminalReport<usize>>
    where
        I: IntoIterator<Item = In>,
    {
        // --- Typed path (Auto / TypedOnly) ---
        if mode != ExecutorMode::ErasedOnly {
            let plan = try_typed_flow_plan::<In, Out>(
                &self.stages,
                &self.edges,
                self.shape.inlet().id(),
                self.shape.outlet().id(),
            );
            if let Some(plan) = plan {
                let mut count = 0usize;
                let mut events = 0usize;
                let mut async_boundary_crossings = 0usize;
                for item in input {
                    plan.run_item_count(item, config, &mut events, &mut async_boundary_crossings)?;
                    count += 1;
                }
                return Ok(FusedTerminalReport {
                    result: count,
                    events,
                    async_boundary_crossings,
                });
            } else if mode == ExecutorMode::TypedOnly {
                return Err(StreamError::GraphValidation(
                    "typed executor does not support this graph shape".into(),
                ));
            }
        }

        // --- Erased fallback ---
        let mut executor = FusedExecutor::new(self, config);
        let inlet = self.shape.inlet().id();
        let outlet = self.shape.outlet().id();
        let mut output_sink = CountOutputSink { count: 0 };

        executor.request::<Out>(outlet, outlet, &mut output_sink)?;
        for item in input {
            executor.deliver::<Out>(inlet, datum(item), outlet, &mut output_sink)?;
        }
        executor.complete::<Out>(inlet, outlet, &mut output_sink)?;

        Ok(FusedTerminalReport {
            result: output_sink.count,
            events: executor.events,
            async_boundary_crossings: executor.async_boundary_crossings,
        })
    }

    pub fn run_fold_with_input<I, Acc, F>(&self, input: I, zero: Acc, fold: F) -> StreamResult<Acc>
    where
        I: IntoIterator<Item = In>,
        F: FnMut(Acc, Out) -> Acc,
    {
        Ok(self
            .run_fold_with_input_report(input, zero, fold, FusedExecutionConfig::default())?
            .result)
    }

    pub fn run_fold_with_input_report<I, Acc, F>(
        &self,
        input: I,
        zero: Acc,
        fold: F,
        config: FusedExecutionConfig,
    ) -> StreamResult<FusedTerminalReport<Acc>>
    where
        I: IntoIterator<Item = In>,
        F: FnMut(Acc, Out) -> Acc,
    {
        self.run_fold_with_input_report_mode(input, zero, fold, config, ExecutorMode::Auto)
    }

    pub(crate) fn run_fold_with_input_report_mode<I, Acc, F>(
        &self,
        input: I,
        zero: Acc,
        mut fold: F,
        config: FusedExecutionConfig,
        mode: ExecutorMode,
    ) -> StreamResult<FusedTerminalReport<Acc>>
    where
        I: IntoIterator<Item = In>,
        F: FnMut(Acc, Out) -> Acc,
    {
        // --- Typed path (Auto / TypedOnly) ---
        if mode != ExecutorMode::ErasedOnly {
            let plan = try_typed_flow_plan::<In, Out>(
                &self.stages,
                &self.edges,
                self.shape.inlet().id(),
                self.shape.outlet().id(),
            );
            if let Some(plan) = plan {
                let mut accumulator = zero;
                let mut events = 0usize;
                let mut async_boundary_crossings = 0usize;
                for item in input {
                    let out =
                        plan.run_item(item, config, &mut events, &mut async_boundary_crossings)?;
                    accumulator = fold(accumulator, out);
                }
                return Ok(FusedTerminalReport {
                    result: accumulator,
                    events,
                    async_boundary_crossings,
                });
            } else if mode == ExecutorMode::TypedOnly {
                return Err(StreamError::GraphValidation(
                    "typed executor does not support this graph shape".into(),
                ));
            }
        }

        // --- Erased fallback ---
        let mut executor = FusedExecutor::new(self, config);
        let inlet = self.shape.inlet().id();
        let outlet = self.shape.outlet().id();
        let mut output_sink = FoldOutputSink {
            accumulator: Some(zero),
            fold,
        };

        executor.request(outlet, outlet, &mut output_sink)?;
        for item in input {
            executor.deliver(inlet, datum(item), outlet, &mut output_sink)?;
        }
        executor.complete(inlet, outlet, &mut output_sink)?;

        Ok(FusedTerminalReport {
            result: output_sink.finish(),
            events: executor.events,
            async_boundary_crossings: executor.async_boundary_crossings,
        })
    }

    // -- Phase 0/Phase 2 mode-aware entry points (pub(crate) / test-hook) ---

    /// Like `run_with_input` but dispatches based on [`ExecutorMode`].
    ///
    /// * `Auto` — tries the typed flow plan; falls back to erased if unsupported.
    /// * `ErasedOnly` — always uses the erased executor.
    /// * `TypedOnly` — errors if the typed plan cannot be built.
    ///
    /// This method is `pub(crate)` and exercised in `#[cfg(test)]` modules.
    #[cfg_attr(not(test), allow(dead_code))]
    pub(crate) fn run_with_input_mode<I>(
        &self,
        input: I,
        mode: ExecutorMode,
    ) -> StreamResult<Vec<Out>>
    where
        I: IntoIterator<Item = In>,
    {
        Ok(self
            .run_with_input_report_mode(input, FusedExecutionConfig::default(), mode)?
            .output)
    }

    /// Like `run_count_with_input` but dispatches based on [`ExecutorMode`].
    ///
    /// `pub(crate)` test/diagnostic hook.
    #[allow(dead_code)]
    pub(crate) fn run_count_with_input_mode<I>(
        &self,
        input: I,
        mode: ExecutorMode,
    ) -> StreamResult<usize>
    where
        I: IntoIterator<Item = In>,
    {
        Ok(self
            .run_count_with_input_report_mode(input, FusedExecutionConfig::default(), mode)?
            .result)
    }

    /// Like `run_fold_with_input` but dispatches based on [`ExecutorMode`].
    ///
    /// `pub(crate)` test/diagnostic hook.
    #[allow(dead_code)]
    pub(crate) fn run_fold_with_input_mode<I, Acc, F>(
        &self,
        input: I,
        zero: Acc,
        fold: F,
        mode: ExecutorMode,
    ) -> StreamResult<Acc>
    where
        I: IntoIterator<Item = In>,
        F: FnMut(Acc, Out) -> Acc,
    {
        Ok(self
            .run_fold_with_input_report_mode(
                input,
                zero,
                fold,
                FusedExecutionConfig::default(),
                mode,
            )?
            .result)
    }
}

impl<T> GraphBlueprint<FlowShape<T, T>>
where
    T: Send + 'static,
{
    pub fn run_typed_linear_with_input<I>(&self, input: I) -> StreamResult<Vec<T>>
    where
        I: IntoIterator<Item = T>,
    {
        Ok(self
            .run_typed_linear_with_input_report(input, FusedExecutionConfig::default())?
            .output)
    }

    pub fn run_typed_linear_with_input_report<I>(
        &self,
        input: I,
        config: FusedExecutionConfig,
    ) -> StreamResult<FusedExecutionReport<T>>
    where
        I: IntoIterator<Item = T>,
    {
        let input = input.into_iter();
        let plan = self.typed_linear_plan()?;
        let mut output = Vec::with_capacity(input.size_hint().0);
        let mut events = 0;
        let mut async_boundary_crossings = 0;

        for item in input {
            let item = plan.run_item(item, config, &mut events, &mut async_boundary_crossings)?;
            output.push(item);
        }

        Ok(FusedExecutionReport {
            output,
            events,
            async_boundary_crossings,
        })
    }

    pub fn run_typed_linear_count_with_input<I>(&self, input: I) -> StreamResult<usize>
    where
        I: IntoIterator<Item = T>,
    {
        Ok(self
            .run_typed_linear_count_with_input_report(input, FusedExecutionConfig::default())?
            .result)
    }

    pub fn run_typed_linear_count_with_input_report<I>(
        &self,
        input: I,
        config: FusedExecutionConfig,
    ) -> StreamResult<FusedTerminalReport<usize>>
    where
        I: IntoIterator<Item = T>,
    {
        let plan = self.typed_linear_plan()?;
        let mut count = 0;
        let mut events = 0;
        let mut async_boundary_crossings = 0;

        for item in input {
            let _ = plan.run_item(item, config, &mut events, &mut async_boundary_crossings)?;
            count += 1;
        }

        Ok(FusedTerminalReport {
            result: count,
            events,
            async_boundary_crossings,
        })
    }

    pub fn run_typed_linear_fold_with_input<I, Acc, F>(
        &self,
        input: I,
        zero: Acc,
        fold: F,
    ) -> StreamResult<Acc>
    where
        I: IntoIterator<Item = T>,
        F: FnMut(Acc, T) -> Acc,
    {
        Ok(self
            .run_typed_linear_fold_with_input_report(
                input,
                zero,
                fold,
                FusedExecutionConfig::default(),
            )?
            .result)
    }

    pub fn run_typed_linear_fold_with_input_report<I, Acc, F>(
        &self,
        input: I,
        zero: Acc,
        mut fold: F,
        config: FusedExecutionConfig,
    ) -> StreamResult<FusedTerminalReport<Acc>>
    where
        I: IntoIterator<Item = T>,
        F: FnMut(Acc, T) -> Acc,
    {
        let plan = self.typed_linear_plan()?;
        let mut accumulator = zero;
        let mut events = 0;
        let mut async_boundary_crossings = 0;

        for item in input {
            let item = plan.run_item(item, config, &mut events, &mut async_boundary_crossings)?;
            accumulator = fold(accumulator, item);
        }

        Ok(FusedTerminalReport {
            result: accumulator,
            events,
            async_boundary_crossings,
        })
    }

    /// Runs the internal async-boundary count path.
    ///
    /// The graph is still runtime-validated through the typed-linear plan, then
    /// split at `AsyncBoundary` stages and connected with bounded handoff.
    /// The concrete boundary executor is intentionally private so Datum does
    /// not expose multiple public runtime backends.
    pub fn run_async_boundary_count_with_input_report<I>(
        &self,
        input: I,
        config: AsyncBoundaryExecutionConfig,
    ) -> StreamResult<FusedTerminalReport<usize>>
    where
        I: IntoIterator<Item = T> + Send,
        I::IntoIter: Send + 'static,
    {
        let segments = self.typed_linear_async_segments()?;
        BoundaryCountExecutor::Ractor.run_count(input, segments, config)
    }

    fn typed_linear_plan(&self) -> StreamResult<TypedLinearPlan<T>> {
        let graph_inlet = self.shape.inlet().id();
        let graph_outlet = self.shape.outlet().id();
        let type_id = TypeId::of::<T>();
        let mut current_inlet = graph_inlet;
        let mut seen = HashSet::new();
        let mut steps = Vec::new();

        loop {
            let stage_index = self
                .stages
                .iter()
                .position(|stage| {
                    stage
                        .spec
                        .inlets
                        .iter()
                        .any(|inlet| inlet.id() == current_inlet)
                })
                .ok_or_else(|| {
                    StreamError::GraphValidation(format!(
                        "typed linear fast path could not find inlet {}",
                        current_inlet.as_usize()
                    ))
                })?;
            if !seen.insert(stage_index) {
                return Err(StreamError::GraphValidation(
                    "typed linear fast path does not support cycles".into(),
                ));
            }

            let stage = &self.stages[stage_index];
            if stage.spec.inlets.len() != 1 || stage.spec.outlets.len() != 1 {
                return Err(StreamError::GraphValidation(format!(
                    "typed linear fast path requires single-inlet single-outlet stages; {} has {} inlet(s) and {} outlet(s)",
                    stage.spec.name(),
                    stage.spec.inlets.len(),
                    stage.spec.outlets.len()
                )));
            }
            let inlet = &stage.spec.inlets[0];
            let outlet = &stage.spec.outlets[0];
            if inlet.type_id() != type_id || outlet.type_id() != type_id {
                return Err(StreamError::GraphValidation(format!(
                    "typed linear fast path requires every port to use {}",
                    type_name::<T>()
                )));
            }

            let step = match &stage.spec.kind {
                StageKind::Identity | StageKind::Opaque => TypedLinearStep::Pass,
                StageKind::AsyncBoundary => TypedLinearStep::AsyncBoundary,
                StageKind::Map(map) => {
                    let mapper = map
                        .typed
                        .as_ref()
                        .downcast_ref::<Arc<dyn Fn(T) -> T + Send + Sync>>()
                        .ok_or_else(|| {
                            StreamError::GraphValidation(format!(
                                "typed linear fast path could not downcast map stage {}",
                                stage.spec.name()
                            ))
                        })?;
                    TypedLinearStep::Map(Arc::clone(mapper))
                }
                _ => {
                    return Err(StreamError::GraphValidation(format!(
                        "typed linear fast path does not support {}",
                        stage.spec.name()
                    )));
                }
            };
            steps.push(step);

            if outlet.id() == graph_outlet {
                break;
            }
            current_inlet = self
                .edges
                .iter()
                .find_map(|edge| (edge.outlet == outlet.id()).then_some(edge.inlet))
                .ok_or_else(|| {
                    StreamError::GraphValidation(format!(
                        "typed linear fast path could not follow outlet {}",
                        outlet.id().as_usize()
                    ))
                })?;
        }

        if seen.len() != self.stages.len() {
            return Err(StreamError::GraphValidation(
                "typed linear fast path requires all stages to be on the result path".into(),
            ));
        }

        Ok(TypedLinearPlan { steps })
    }

    pub(super) fn typed_linear_async_segments(&self) -> StreamResult<TypedLinearSegments<T>> {
        let plan = self.typed_linear_plan()?;
        let mut segments = Vec::new();
        let mut current = Vec::new();

        for step in plan.steps {
            match step {
                TypedLinearStep::AsyncBoundary => {
                    segments.push(current);
                    current = Vec::new();
                }
                step => current.push(step),
            }
        }
        segments.push(current);

        if segments.len() == 1 {
            return Err(StreamError::GraphValidation(
                "async boundary execution requires at least one AsyncBoundary stage".into(),
            ));
        }

        Ok(TypedLinearSegments { segments })
    }
}

pub(super) struct TypedLinearPlan<T> {
    steps: Vec<TypedLinearStep<T>>,
}

pub(super) struct TypedLinearSegments<T> {
    segments: Vec<Vec<TypedLinearStep<T>>>,
}

pub(super) enum TypedLinearStep<T> {
    Pass,
    Map(Arc<dyn Fn(T) -> T + Send + Sync>),
    AsyncBoundary,
}

impl<T> Clone for TypedLinearStep<T> {
    fn clone(&self) -> Self {
        match self {
            Self::Pass => Self::Pass,
            Self::Map(mapper) => Self::Map(Arc::clone(mapper)),
            Self::AsyncBoundary => Self::AsyncBoundary,
        }
    }
}

impl<T> TypedLinearPlan<T> {
    fn run_item(
        &self,
        mut item: T,
        config: FusedExecutionConfig,
        events: &mut usize,
        async_boundary_crossings: &mut usize,
    ) -> StreamResult<T>
    where
        T: Send + 'static,
    {
        for step in &self.steps {
            bump_fused_event(events, config)?;
            match step {
                TypedLinearStep::Pass => {}
                TypedLinearStep::Map(mapper) => {
                    item = mapper(item);
                }
                TypedLinearStep::AsyncBoundary => {
                    *async_boundary_crossings += 1;
                }
            }
            bump_fused_event(events, config)?;
        }
        Ok(item)
    }
}

// ---------------------------------------------------------------------------
// Phase 1 — Typed-port substrate (WP-18)
// ---------------------------------------------------------------------------

/// A typed storage cell for a single value associated with one port.
///
/// Used by typed junction kernels (Phase 3+) to avoid `DatumValue` boxing for
/// per-port buffering.  Phase 2 linear kernels do not need per-port storage
/// because values flow sequentially through the pipeline without buffering.
// Phase 3+ will use TypedSlot actively; allow dead_code until then.
#[allow(dead_code)]
pub(crate) struct TypedSlot<T>(Option<T>);

#[allow(dead_code)]
impl<T> TypedSlot<T> {
    pub(crate) fn empty() -> Self {
        Self(None)
    }

    pub(crate) fn put(&mut self, value: T) {
        self.0 = Some(value);
    }

    pub(crate) fn take(&mut self) -> Option<T> {
        self.0.take()
    }

    pub(crate) fn is_some(&self) -> bool {
        self.0.is_some()
    }
}

/// Heterogeneous map from [`PortId`] to a type-erased [`TypedSlot<T>`].
///
/// Typed junction kernels (Phase 3+) register per-port slots here at plan
/// time, then access them during execution without boxing the values.
// Phase 3+ will use TypedPortRegistry actively; allow dead_code until then.
#[allow(dead_code)]
pub(crate) struct TypedPortRegistry {
    slots: HashMap<PortId, Box<dyn Any + Send>>,
}

#[allow(dead_code)]
impl TypedPortRegistry {
    pub(crate) fn new() -> Self {
        Self {
            slots: HashMap::new(),
        }
    }

    /// Register a typed slot for `port_id`.  Panics if a slot for that port
    /// already exists (programming error — ports must be registered once).
    pub(crate) fn register<T: Any + Send>(&mut self, port_id: PortId) {
        let prev = self
            .slots
            .insert(port_id, Box::new(TypedSlot::<T>::empty()));
        assert!(prev.is_none(), "port {port_id:?} registered twice");
    }

    /// Obtain a mutable reference to the typed slot for `port_id`.
    ///
    /// Returns `None` if the port was not registered or the stored type does
    /// not match `T`.
    pub(crate) fn get_mut<T: Any + Send>(&mut self, port_id: PortId) -> Option<&mut TypedSlot<T>> {
        self.slots.get_mut(&port_id)?.downcast_mut::<TypedSlot<T>>()
    }
}

/// Per-event kernel for a single stage in the typed-port executor.
///
/// A kernel receives a strongly-typed input value and produces a strongly-typed
/// output — **no `DatumValue` boxing, no per-element downcast.**  Kernels are
/// constructed once at plan time via a [`TypedStageFactory`] and reused across
/// all items in the stream.
// Phase 3+ will add typed junction kernels; allow dead_code until then.
#[allow(dead_code)]
pub(crate) trait TypedKernel<In, Out>: Send + Sync {
    fn run(&self, input: In) -> Out;
}

/// Factory that constructs a [`TypedKernel<In, Out>`] for a specific stage at
/// plan time.
///
/// Factories perform **safe `Any` downcasts** on the stage spec's stored
/// type-erased functions once, during planning.  They return `None` when the
/// stage is unsupported (falls back to the erased executor in `Auto` mode).
// Phase 3+ will add typed junction factories; allow dead_code until then.
#[allow(dead_code)]
pub(crate) trait TypedStageFactory<In, Out>: Send + Sync {
    fn try_build(&self, spec: &StageSpec) -> Option<Box<dyn TypedKernel<In, Out>>>;
}

// ---------------------------------------------------------------------------
// Phase 2 — TypedFlowPlan<In, Out> (WP-18)
// ---------------------------------------------------------------------------

/// A typed, per-element step in a linear flow plan.
///
/// `In` here is the **intermediate** element type — all middle steps are typed
/// as `In → In`.  Only the last step may produce a different type `Out`; that
/// is handled separately in [`TypedFlowPlan`].
enum TypedMiddleStep<T: 'static> {
    /// Pass the value through unchanged.
    Pass,
    /// Apply a pure typed map function.
    Map(Arc<dyn Fn(T) -> T + Send + Sync>),
    /// Record an async-boundary crossing but do not move to a different thread.
    AsyncBoundary,
}

/// The terminal step in a [`TypedFlowPlan`], which produces the final `Out`.
///
/// Two variants:
///
/// * `Map` — the last stage has a stored typed `Fn(In) → Out`.  Obtained by a
///   plan-time `downcast_ref` on `StageMapFns::typed`.  Zero per-element
///   allocations.
/// * `Identity` — the last stage is a pass-through.  This variant is only
///   constructed when `TypeId::of::<In>() == TypeId::of::<Out>()`, i.e. the
///   types are the same at runtime.  The stored closure performs a single
///   `Box<dyn Any + Send>::downcast::<Out>()` per element — one heap allocation
///   — to safely reinterpret the `In` value as `Out` without `unsafe`.
enum TypedLastStep<In: 'static, Out: 'static> {
    /// Last stage is a typed map that directly produces `Out`.
    Map(Arc<dyn Fn(In) -> Out + Send + Sync>),
    /// Last stage is a pass-through; `In` and `Out` are the same type at
    /// runtime (verified via [`TypeId`] at plan time).  One allocation per
    /// element to perform the safe `In → Out` reinterpretation.
    Identity(Arc<dyn Fn(In) -> Out + Send + Sync>),
}

/// A typed, fused execution plan for a linear `FlowShape<In, Out>` graph.
///
/// Built once at plan time by [`try_typed_flow_plan`]; reused across items.
///
/// **No `DatumValue` per-element boxing for Map graphs.**  For Identity-only
/// graphs a single `Box<dyn Any + Send>` allocation per element converts the
/// final `In` value to `Out` safely (see [`TypedLastStep::Identity`]).
///
/// For **count** sinks the plan short-circuits before the last step: no output
/// value is produced and no allocation occurs.
pub(crate) struct TypedFlowPlan<In: 'static, Out: 'static> {
    /// Intermediate steps: `In → In`.
    middle_steps: Vec<TypedMiddleStep<In>>,
    /// Terminal step: `In → Out`.
    last_step: TypedLastStep<In, Out>,
    /// Total stage count (= `middle_steps.len() + 1`).
    /// Retained for diagnostic use; not yet read in execution.
    #[allow(dead_code)]
    stage_count: usize,
}

impl<In: Send + 'static, Out: Send + 'static> TypedFlowPlan<In, Out> {
    /// Run a single item through the typed plan, producing `Out`.
    ///
    /// Bumps the fused event counter **twice** per stage (before and after),
    /// matching the erased executor's accounting.
    pub(crate) fn run_item(
        &self,
        item: In,
        config: FusedExecutionConfig,
        events: &mut usize,
        async_boundary_crossings: &mut usize,
    ) -> StreamResult<Out> {
        let mut val = item;
        for step in &self.middle_steps {
            bump_fused_event(events, config)?;
            val = match step {
                TypedMiddleStep::Pass => val,
                TypedMiddleStep::Map(f) => f(val),
                TypedMiddleStep::AsyncBoundary => {
                    *async_boundary_crossings += 1;
                    val
                }
            };
            bump_fused_event(events, config)?;
        }
        // Terminal step.
        bump_fused_event(events, config)?;
        let out = match &self.last_step {
            TypedLastStep::Map(f) => f(val),
            TypedLastStep::Identity(f) => f(val),
        };
        bump_fused_event(events, config)?;
        Ok(out)
    }

    /// Run a single item through all **middle** stages only, counting events
    /// for the terminal stage but **not** producing an `Out` value.
    ///
    /// Used by `run_count_with_input_report` to avoid any allocation when the
    /// last step is an identity pass-through (`TypedLastStep::Identity`).
    /// Map-last graphs call this path too (the terminal map function is still
    /// executed to maintain semantic equivalence with the erased path, but
    /// its result is discarded — pure functions have no observable side effects).
    pub(crate) fn run_item_count(
        &self,
        item: In,
        config: FusedExecutionConfig,
        events: &mut usize,
        async_boundary_crossings: &mut usize,
    ) -> StreamResult<()> {
        let mut val = item;
        for step in &self.middle_steps {
            bump_fused_event(events, config)?;
            val = match step {
                TypedMiddleStep::Pass => val,
                TypedMiddleStep::Map(f) => f(val),
                TypedMiddleStep::AsyncBoundary => {
                    *async_boundary_crossings += 1;
                    val
                }
            };
            bump_fused_event(events, config)?;
        }
        // Terminal stage: bump events and run (discard output).
        bump_fused_event(events, config)?;
        match &self.last_step {
            TypedLastStep::Map(f) => {
                let _ = f(val);
            }
            TypedLastStep::Identity(_) => {
                // Pass-through: no allocation needed.  Drop `val` in place.
                drop(val);
            }
        }
        bump_fused_event(events, config)?;
        Ok(())
    }
}

/// Attempt to build a [`TypedFlowPlan<In, Out>`] for the given linear graph.
///
/// Returns `None` when the graph cannot be executed on the typed path.  This
/// happens for:
/// - Non-linear graphs (junctions, multi-inlet/outlet stages).
/// - Cyclic stage graphs.
/// - Graphs where not all stage port types match `In`/`Out` in a way that
///   allows safe plan-time downcasts.
/// - Stages with `StageKind::Opaque` (custom logic; always falls back to
///   erased).
///
/// On success, returns `Some(plan)` and the callers (`run_with_input_report`
/// etc.) skip the erased executor entirely.
pub(crate) fn try_typed_flow_plan<In, Out>(
    stages: &[super::builder::StageRecord],
    edges: &[super::builder::Edge],
    graph_inlet: PortId,
    graph_outlet: PortId,
) -> Option<TypedFlowPlan<In, Out>>
where
    In: Clone + Send + 'static,
    Out: Send + 'static,
{
    let in_type_id = TypeId::of::<In>();
    let out_type_id = TypeId::of::<Out>();

    let mut current_inlet = graph_inlet;
    let mut seen = HashSet::new();
    // Collect (StageKind, outlet_type_id) pairs as we walk the chain.
    let mut stage_infos: Vec<(&StageKind, TypeId)> = Vec::new();

    loop {
        let stage_index = stages.iter().position(|s| {
            s.spec
                .inlets
                .iter()
                .any(|inlet| inlet.id() == current_inlet)
        })?;
        if !seen.insert(stage_index) {
            // Cycle detected — not supported on typed path.
            return None;
        }
        let stage = &stages[stage_index];
        if stage.spec.inlets.len() != 1 || stage.spec.outlets.len() != 1 {
            // Non-linear stage (junction) — fall back to erased.
            return None;
        }
        let inlet = &stage.spec.inlets[0];
        let outlet = &stage.spec.outlets[0];

        // All intermediate inlets must have type In.
        if inlet.type_id() != in_type_id {
            return None;
        }

        stage_infos.push((&stage.spec.kind, outlet.type_id()));

        if outlet.id() == graph_outlet {
            break;
        }
        // Follow the edge to the next inlet.
        current_inlet = edges
            .iter()
            .find_map(|e| (e.outlet == outlet.id()).then_some(e.inlet))?;
    }

    if seen.len() != stages.len() {
        // Not all stages are on the result path (unreachable stages).
        return None;
    }

    // Build the plan.  All stages except the last are `In → In`; the last
    // stage may produce a different `Out`.
    let (last_kind, last_outlet_type) = stage_infos.last()?;
    // The last outlet must have type `Out`.
    if *last_outlet_type != out_type_id {
        return None;
    }

    let total = stage_infos.len();
    let mut middle_steps: Vec<TypedMiddleStep<In>> = Vec::with_capacity(total.saturating_sub(1));

    for (kind, outlet_type) in &stage_infos[..total.saturating_sub(1)] {
        // Intermediate outlets must all have type `In`.
        if *outlet_type != in_type_id {
            return None;
        }
        let step = match kind {
            StageKind::Identity => TypedMiddleStep::Pass,
            // Opaque stages have custom GraphStageLogic — they can emit multiple
            // values, buffer, or do arbitrary things.  They are never safe to
            // treat as a pass-through on the typed path.
            StageKind::Opaque => return None,
            StageKind::AsyncBoundary => TypedMiddleStep::AsyncBoundary,
            StageKind::Map(map) => {
                // Downcast `typed` to `Arc<Fn(In) -> In>`.
                let f = map
                    .typed
                    .downcast_ref::<Arc<dyn Fn(In) -> In + Send + Sync>>()?;
                TypedMiddleStep::Map(Arc::clone(f))
            }
            _ => return None,
        };
        middle_steps.push(step);
    }

    // Build the terminal step.
    let last_step: TypedLastStep<In, Out> = match last_kind {
        StageKind::Identity => {
            // Only valid when In = Out (same TypeId).
            if in_type_id != out_type_id {
                return None;
            }
            // Safe reinterpretation via Box<dyn Any + Send> downcast.
            // One allocation per element; no `unsafe`.
            TypedLastStep::Identity(Arc::new(|x: In| -> Out {
                let boxed: Box<dyn Any + Send> = Box::new(x);
                *boxed
                    .downcast::<Out>()
                    .expect("TypeId equality verified at plan time")
            }))
        }
        // Opaque stages always fall back to the erased executor.
        StageKind::Opaque => return None,
        StageKind::AsyncBoundary => {
            // Async boundary as last stage: same as Identity (sync pass-through
            // in the non-async fused executor).
            if in_type_id != out_type_id {
                return None;
            }
            TypedLastStep::Identity(Arc::new(|x: In| -> Out {
                let boxed: Box<dyn Any + Send> = Box::new(x);
                *boxed
                    .downcast::<Out>()
                    .expect("TypeId equality verified at plan time")
            }))
        }
        StageKind::Map(map) => {
            // Downcast `typed` to `Arc<Fn(In) -> Out>`.
            let f = map
                .typed
                .downcast_ref::<Arc<dyn Fn(In) -> Out + Send + Sync>>()?;
            TypedLastStep::Map(Arc::clone(f))
        }
        _ => return None,
    };

    Some(TypedFlowPlan {
        middle_steps,
        last_step,
        stage_count: total,
    })
}

// ---------------------------------------------------------------------------
// Phase 3a — Typed MergeSequence kernel (WP-18)
// ---------------------------------------------------------------------------

/// Shared ordering core for `MergeSequence<T>`.
///
/// Holds all mutable state for sequence-ordering a fan-in: the next expected
/// sequence number, a pending buffer for out-of-order arrivals, and an output
/// queue for items whose sequence number has been resolved.
///
/// Used by the **typed** `MergeSequencePlan` execution path.  The erased
/// executor maintains equivalent state inline in [`StageState::MergeSequence`]
/// (shared semantics, forced equivalence via tests — see the `typed_vs_erased`
/// test suite).
pub(crate) struct MergeSequenceCore<T> {
    next_sequence: u64,
    /// Out-of-order items waiting for their sequence slot to open.
    pending: Vec<(u64, T)>,
    /// Items ready to emit, in sequence order.
    output_buffer: VecDeque<T>,
    /// Number of input ports that have completed.
    completed_count: usize,
    /// Total number of input ports.
    input_count: usize,
    /// Whether this stage has finished (all inputs done, all outputs drained).
    completed: bool,
}

impl<T> MergeSequenceCore<T> {
    pub(crate) fn new(input_count: usize) -> Self {
        Self {
            next_sequence: 0,
            pending: Vec::new(),
            output_buffer: VecDeque::new(),
            completed_count: 0,
            input_count,
            completed: false,
        }
    }

    /// Reset state for reuse across benchmark iterations.
    fn reset(&mut self) {
        self.next_sequence = 0;
        self.pending.clear();
        self.output_buffer.clear();
        self.completed_count = 0;
        self.completed = false;
    }

    /// Push a typed value into the core with its extracted sequence number.
    ///
    /// On success, all newly-resolved items (whose sequence is now contiguous
    /// from `next_sequence`) are appended to `output_buffer`.  On duplicate
    /// sequence, returns `StreamError::Failed`.
    fn push_item(&mut self, seq: u64, val: T) -> StreamResult<()> {
        if seq == self.next_sequence {
            self.output_buffer.push_back(val);
            self.next_sequence += 1;
            // Drain any pending items that are now in sequence.
            while let Some(index) = self
                .pending
                .iter()
                .position(|(s, _)| *s == self.next_sequence)
            {
                let (_, item) = self.pending.remove(index);
                self.output_buffer.push_back(item);
                self.next_sequence += 1;
            }
        } else {
            if self.pending.iter().any(|(s, _)| *s == seq) {
                return Err(StreamError::Failed(format!(
                    "duplicate sequence {seq} on merge sequence"
                )));
            }
            self.pending.push((seq, val));
            self.pending.sort_by_key(|(s, _)| *s);
            // Check whether the new item resolves any contiguous run.
            while let Some(index) = self
                .pending
                .iter()
                .position(|(s, _)| *s == self.next_sequence)
            {
                let (_, item) = self.pending.remove(index);
                self.output_buffer.push_back(item);
                self.next_sequence += 1;
            }
        }
        Ok(())
    }

    /// Signal that one input port has completed.
    ///
    /// Returns `true` when all inputs have completed *and* the output buffer
    /// is empty (the stage can now terminate cleanly).
    ///
    /// Returns `Err` when all inputs have completed but there are still items
    /// in `pending` that cannot be resolved — a sequence gap.
    fn on_inlet_complete(&mut self) -> StreamResult<bool> {
        self.completed_count += 1;
        if self.completed_count >= self.input_count && self.output_buffer.is_empty() {
            if !self.pending.is_empty() {
                return Err(StreamError::Failed(format!(
                    "expected sequence {}, but all input ports have pushed or are complete",
                    self.next_sequence,
                )));
            }
            self.completed = true;
            Ok(true)
        } else {
            Ok(false)
        }
    }

    /// Drain all available output items into `out`.
    fn drain_into(&mut self, out: &mut Vec<T>) {
        out.extend(self.output_buffer.drain(..));
    }
}

// ---------------------------------------------------------------------------
// Typed MergeSequence plan: Unzip<A,B> → MergeSequence<T> topology
// ---------------------------------------------------------------------------

/// A typed, fused execution plan for a graph whose only stages are one
/// `Unzip`-like fan-out followed by one `MergeSequence` fan-in, where every
/// port shares the same element type `T` and `In` is the split input type.
///
/// Built once at plan time by [`try_typed_merge_sequence_plan`]; executed by
/// [`run_typed_merge_sequence`].
///
/// **No `DatumValue` boxing per element.**  The split and sequence-extractor
/// functions are called on strongly-typed values.
pub(crate) struct TypedMergeSequencePlan<In, T> {
    /// Splits one `In` value into `n` typed `T` values (one per MergeSequence
    /// inlet).  Currently only n=2 is supported (Unzip → MergeSequence(2)).
    splits: Vec<Arc<dyn Fn(In) -> T + Send + Sync>>,
    /// Extracts the sequence number from a `T` value.
    extract_sequence: Arc<dyn Fn(&T) -> u64 + Send + Sync>,
    /// Mutable ordering state, reset between runs.
    core: MergeSequenceCore<T>,
}

impl<In: Clone + Send + 'static, T: Send + 'static> TypedMergeSequencePlan<In, T> {
    /// Push one input item, extract and order both split values, drain ready
    /// outputs into `out`.
    fn push_item(&mut self, item: In, out: &mut Vec<T>) -> StreamResult<()> {
        for split_fn in &self.splits {
            let val = split_fn(item.clone());
            let seq = (self.extract_sequence)(&val);
            self.core.push_item(seq, val)?;
        }
        self.core.drain_into(out);
        Ok(())
    }

    /// Signal end-of-stream (all inputs exhausted) and check for gaps.
    ///
    /// Each split function simulates one logical inlet completing.
    fn finish(&mut self, out: &mut Vec<T>) -> StreamResult<()> {
        for _ in 0..self.splits.len() {
            self.core.on_inlet_complete()?;
        }
        self.core.drain_into(out);
        Ok(())
    }

    /// Reset internal state so the plan can be reused across benchmark iterations.
    fn reset(&mut self) {
        self.core.reset();
    }
}

/// Attempt to build a [`TypedMergeSequencePlan<In, T>`] for the given graph.
///
/// Succeeds when the graph has exactly these two stages (in topological order):
/// 1. A `StageKind::Unzip` with one inlet (type `In`) and `k` outlets (type `T`).
/// 2. A `StageKind::MergeSequence` with `k` inlets (type `T`) and one outlet
///    (type `T` = `Out`).
///
/// All `k` outlets of the Unzip must connect to the `k` inlets of the
/// MergeSequence, and all port types must match `T`.
///
/// Returns `None` for any other topology (falls back to erased executor).
pub(crate) fn try_typed_merge_sequence_plan<In, Out>(
    stages: &[super::builder::StageRecord],
    edges: &[super::builder::Edge],
    graph_inlet: PortId,
    graph_outlet: PortId,
) -> Option<TypedMergeSequencePlan<In, Out>>
where
    In: Clone + Send + 'static,
    Out: Send + 'static,
{
    // We require exactly two stages.
    if stages.len() != 2 {
        return None;
    }

    let in_type_id = TypeId::of::<In>();
    let out_type_id = TypeId::of::<Out>();

    // Find the Unzip stage (the one whose inlet id matches the graph inlet).
    let unzip_idx = stages
        .iter()
        .position(|s| s.spec.inlets.len() == 1 && s.spec.inlets[0].id() == graph_inlet)?;
    let unzip_stage = &stages[unzip_idx];

    // Must be a StageKind::Unzip.
    let typed_split_any = match &unzip_stage.spec.kind {
        StageKind::Unzip { typed_split, .. } => Arc::clone(typed_split),
        _ => return None,
    };

    // The unzip inlet must have type In.
    if unzip_stage.spec.inlets[0].type_id() != in_type_id {
        return None;
    }

    // All unzip outlets must have type Out.
    let k = unzip_stage.spec.outlets.len();
    if k == 0 {
        return None;
    }
    for outlet in &unzip_stage.spec.outlets {
        if outlet.type_id() != out_type_id {
            return None;
        }
    }

    // Find the MergeSequence stage.
    let ms_idx = 1 - unzip_idx;
    let ms_stage = &stages[ms_idx];

    // Must be a StageKind::MergeSequence with k inlets and 1 outlet.
    let (ms_input_count, typed_extract_any) = match &ms_stage.spec.kind {
        StageKind::MergeSequence {
            input_count,
            typed_extract,
            ..
        } => (*input_count, Arc::clone(typed_extract)),
        _ => return None,
    };

    if ms_stage.spec.inlets.len() != k || ms_stage.spec.outlets.len() != 1 {
        return None;
    }
    if ms_input_count != k {
        return None;
    }
    // All MergeSequence ports must have type Out.
    for inlet in &ms_stage.spec.inlets {
        if inlet.type_id() != out_type_id {
            return None;
        }
    }
    if ms_stage.spec.outlets[0].type_id() != out_type_id {
        return None;
    }

    // MergeSequence outlet must be the graph outlet.
    if ms_stage.spec.outlets[0].id() != graph_outlet {
        return None;
    }

    // Verify edge wiring: each Unzip outlet must connect to one MergeSequence inlet.
    // Build a set of (unzip_outlet_id → ms_inlet_id) edges and verify all k outlets
    // are connected.
    let unzip_outlet_ids: Vec<PortId> =
        unzip_stage.spec.outlets.iter().map(AnyOutlet::id).collect();
    let ms_inlet_ids: Vec<PortId> = ms_stage.spec.inlets.iter().map(AnyInlet::id).collect();

    // For each Unzip outlet, find the edge and verify it leads to one of the MS inlets.
    let mut outlet_to_ms_inlet: Vec<Option<usize>> = vec![None; k];
    for edge in edges {
        if let Some(uo_idx) = unzip_outlet_ids.iter().position(|&id| id == edge.outlet) {
            if let Some(mi_idx) = ms_inlet_ids.iter().position(|&id| id == edge.inlet) {
                outlet_to_ms_inlet[uo_idx] = Some(mi_idx);
            } else {
                return None; // Unzip outlet goes somewhere unexpected.
            }
        }
    }
    if outlet_to_ms_inlet.iter().any(|x| x.is_none()) {
        return None; // Not all Unzip outlets are wired to MergeSequence.
    }

    // Down-cast typed_split: `Arc<StageTypedUnzipFn>` holds an
    // `Arc<dyn Fn(In) -> (Out0, Out1) + Send + Sync>` (for k=2) or
    // `Arc<dyn Fn(In) -> (Out, Out)>` etc.  We need per-outlet extractors.
    //
    // For k=2 (the only Unzip arity we support today): downcast to
    // `Arc<dyn Fn(In) -> (Out, Out) + Send + Sync>`.
    if k != 2 {
        // Only Unzip (2 outputs) is supported right now.
        return None;
    }

    // Downcast typed_split to Arc<dyn Fn(In) -> (Out, Out) + Send + Sync>.
    let typed_split =
        typed_split_any.downcast_ref::<Arc<dyn Fn(In) -> (Out, Out) + Send + Sync>>()?;
    let typed_split = Arc::clone(typed_split);

    // Downcast typed_extract to Arc<dyn Fn(&Out) -> u64 + Send + Sync>.
    let typed_extract =
        typed_extract_any.downcast_ref::<Arc<dyn Fn(&Out) -> u64 + Send + Sync>>()?;
    let typed_extract = Arc::clone(typed_extract);

    // Build per-inlet split closures respecting the wiring order.
    // outlet_to_ms_inlet[0] = which MS inlet Unzip.out0 connects to.
    // outlet_to_ms_inlet[1] = which MS inlet Unzip.out1 connects to.
    //
    // We build `splits[ms_inlet_idx]` = the closure that extracts that value.
    #[allow(clippy::type_complexity)]
    let mut splits: Vec<Option<Arc<dyn Fn(In) -> Out + Send + Sync>>> = vec![None; k];

    let split0 = Arc::clone(&typed_split);
    let split1 = Arc::clone(&typed_split);

    let ms_idx_for_out0 = outlet_to_ms_inlet[0].unwrap();
    let ms_idx_for_out1 = outlet_to_ms_inlet[1].unwrap();

    splits[ms_idx_for_out0] = Some(Arc::new(move |input: In| split0(input).0));
    splits[ms_idx_for_out1] = Some(Arc::new(move |input: In| split1(input).1));

    // All slots must be filled.
    if splits.iter().any(|s| s.is_none()) {
        return None;
    }
    let splits: Vec<Arc<dyn Fn(In) -> Out + Send + Sync>> =
        splits.into_iter().map(|s| s.unwrap()).collect();

    Some(TypedMergeSequencePlan {
        splits,
        extract_sequence: typed_extract,
        core: MergeSequenceCore::new(k),
    })
}

/// Execute a [`TypedMergeSequencePlan`], collecting all outputs.
///
/// Called from [`run_with_input_report_mode`] when the typed plan is available.
pub(crate) fn run_typed_merge_sequence<In, T, I>(
    plan: &mut TypedMergeSequencePlan<In, T>,
    input: I,
) -> StreamResult<Vec<T>>
where
    In: Clone + Send + 'static,
    T: Send + 'static,
    I: IntoIterator<Item = In>,
{
    plan.reset();
    // Pre-allocate with a 2× hint (each input produces 2 outputs).
    let input = input.into_iter();
    let hint = input.size_hint().0;
    let mut output: Vec<T> = Vec::with_capacity(hint * plan.splits.len());
    for item in input {
        plan.push_item(item, &mut output)?;
    }
    plan.finish(&mut output)?;
    Ok(output)
}

// ---------------------------------------------------------------------------
// Phase 3b — Typed MergeLatest kernel (WP-18)
// ---------------------------------------------------------------------------

/// Shared latest-value core for `MergeLatest<T>`.
///
/// Holds per-inlet latest slots, a seen/completed count, and the pending-snapshot
/// queue.  The typed plan allocates `Vec<T>` snapshots directly without boxing.
///
/// Used by the **typed** `MergeLatestPlan` execution path.  The erased executor
/// maintains equivalent state inline in [`StageState::MergeLatest`] (shared
/// semantics, forced equivalence via typed-vs-erased tests).
pub(crate) struct MergeLatestCore<T> {
    /// Latest value seen on each inlet, `None` until first push.
    latest: Vec<Option<T>>,
    /// Number of inlets that have received at least one value.
    seen_count: usize,
    /// Number of inlets that have completed.
    completed_count: usize,
    /// Total number of input inlets.
    input_count: usize,
    /// Snapshots ready to emit (built when all inlets seen at least once).
    pending: VecDeque<Vec<T>>,
    /// Whether this stage has finished.
    completed: bool,
    /// When `true`, complete on the first inlet completion (Akka eager semantics).
    eager_complete: bool,
}

impl<T: Clone> MergeLatestCore<T> {
    pub(crate) fn new(input_count: usize, eager_complete: bool) -> Self {
        Self {
            latest: vec![None; input_count],
            seen_count: 0,
            completed_count: 0,
            input_count,
            pending: VecDeque::new(),
            completed: false,
            eager_complete,
        }
    }

    /// Reset state for reuse across benchmark iterations.
    fn reset(&mut self) {
        for slot in &mut self.latest {
            *slot = None;
        }
        self.seen_count = 0;
        self.completed_count = 0;
        self.pending.clear();
        self.completed = false;
    }

    /// Push a value on `inlet_index`, update latest, and enqueue a snapshot if
    /// all inlets have been seen at least once.
    fn push_item(&mut self, inlet_index: usize, val: T) {
        if self.latest[inlet_index].is_none() {
            self.seen_count += 1;
        }
        self.latest[inlet_index] = Some(val);
        if self.seen_count >= self.input_count {
            // Build a snapshot without boxing: clone each slot directly.
            let snapshot: Vec<T> = self
                .latest
                .iter()
                .map(|s| s.clone().expect("merge-latest typed: slot seen but None"))
                .collect();
            self.pending.push_back(snapshot);
        }
    }

    /// Signal that one inlet has completed.
    ///
    /// Returns `true` if the stage should now terminate (all inlets done, or
    /// `eager_complete` with no pending output).
    fn on_inlet_complete(&mut self) -> bool {
        self.completed_count += 1;
        let all_done = self.completed_count >= self.input_count;
        let eager_done = self.eager_complete && self.pending.is_empty();
        if all_done || eager_done {
            self.completed = true;
            true
        } else {
            false
        }
    }

    /// Drain all pending snapshots into `out`.
    fn drain_into(&mut self, out: &mut Vec<Vec<T>>) {
        out.extend(self.pending.drain(..));
    }
}

// ---------------------------------------------------------------------------
// Typed MergeLatest plan: Unzip<In> → MergeLatest<T> topology
// ---------------------------------------------------------------------------

/// A typed, fused execution plan for a graph whose only stages are one
/// `Unzip`-like fan-out followed by one `MergeLatest` fan-in, where every
/// inlet shares the same element type `T` and the output is `Vec<T>`.
///
/// Built once at plan time by [`try_typed_merge_latest_plan`]; executed by
/// [`run_typed_merge_latest`].
///
/// **No `DatumValue` boxing per element.**  Only the final `Vec<T>` snapshot
/// allocation is unavoidable (it is the genuine output type).
pub(crate) struct TypedMergeLatestPlan<In, T> {
    /// Per-outlet split functions: `splits[i](input)` produces the value for
    /// MergeLatest inlet `i`.
    splits: Vec<Arc<dyn Fn(In) -> T + Send + Sync>>,
    /// Mutable state, reset between runs.
    core: MergeLatestCore<T>,
}

impl<In: Clone + Send + 'static, T: Clone + Send + 'static> TypedMergeLatestPlan<In, T> {
    /// Push one input item, fan it out to all inlets, and drain ready snapshots.
    fn push_item(&mut self, item: In, out: &mut Vec<Vec<T>>) {
        for (idx, split_fn) in self.splits.iter().enumerate() {
            let val = split_fn(item.clone());
            self.core.push_item(idx, val);
        }
        self.core.drain_into(out);
    }

    /// Signal end-of-stream (the upstream Unzip completed — all inlets complete
    /// simultaneously).
    fn finish(&mut self) -> bool {
        // The Unzip completes all its outlets at once; each outlet is one ML inlet.
        for _ in 0..self.splits.len() {
            if self.core.on_inlet_complete() {
                return true;
            }
        }
        true
    }

    /// Reset internal state so the plan can be reused across benchmark iterations.
    fn reset(&mut self) {
        self.core.reset();
    }
}

/// Attempt to build a [`TypedMergeLatestPlan<In, T>`] for the given graph.
///
/// Succeeds when the graph has exactly these two stages (in topological order):
/// 1. A `StageKind::Unzip` with one inlet (type `In`) and `k` outlets (type `T`).
/// 2. A `StageKind::MergeLatest` with `k` inlets (type `T`) and one outlet
///    (type `Vec<T>`).
///
/// All `k` outlets of the Unzip must connect to the `k` inlets of the
/// MergeLatest, and all port types must match.
///
/// **Type parameters**: `In` is the graph inlet type; `T` is the MergeLatest
/// element type (each inlet and each slot has type `T`; the output per snapshot
/// is `Vec<T>`).
///
/// Returns `None` for any other topology (falls back to erased executor).
pub(crate) fn try_typed_merge_latest_plan<In, T>(
    stages: &[super::builder::StageRecord],
    edges: &[super::builder::Edge],
    graph_inlet: PortId,
    graph_outlet: PortId,
) -> Option<TypedMergeLatestPlan<In, T>>
where
    In: Clone + Send + 'static,
    T: Clone + Send + 'static,
{
    // We require exactly two stages.
    if stages.len() != 2 {
        return None;
    }

    let in_type_id = TypeId::of::<In>();
    let elem_type_id = TypeId::of::<T>();
    let vec_type_id = TypeId::of::<Vec<T>>();

    // Find the Unzip stage (the one whose inlet id matches the graph inlet).
    let unzip_idx = stages
        .iter()
        .position(|s| s.spec.inlets.len() == 1 && s.spec.inlets[0].id() == graph_inlet)?;
    let unzip_stage = &stages[unzip_idx];

    // Must be a StageKind::Unzip.
    let typed_split_any = match &unzip_stage.spec.kind {
        StageKind::Unzip { typed_split, .. } => Arc::clone(typed_split),
        _ => return None,
    };

    // The unzip inlet must have type In.
    if unzip_stage.spec.inlets[0].type_id() != in_type_id {
        return None;
    }

    // All unzip outlets must have type T.
    let k = unzip_stage.spec.outlets.len();
    if k == 0 {
        return None;
    }
    for outlet in &unzip_stage.spec.outlets {
        if outlet.type_id() != elem_type_id {
            return None;
        }
    }

    // Find the MergeLatest stage.
    let ml_idx = 1 - unzip_idx;
    let ml_stage = &stages[ml_idx];

    // Must be a StageKind::MergeLatest with k inlets and 1 outlet.
    let (ml_input_count, typed_snapshot_any) = match &ml_stage.spec.kind {
        StageKind::MergeLatest {
            input_count,
            typed_snapshot,
            ..
        } => (*input_count, Arc::clone(typed_snapshot)),
        _ => return None,
    };

    if ml_stage.spec.inlets.len() != k || ml_stage.spec.outlets.len() != 1 {
        return None;
    }
    if ml_input_count != k {
        return None;
    }
    // All MergeLatest inlets must have type T.
    for inlet in &ml_stage.spec.inlets {
        if inlet.type_id() != elem_type_id {
            return None;
        }
    }
    // MergeLatest outlet must have type Vec<T>.
    if ml_stage.spec.outlets[0].type_id() != vec_type_id {
        return None;
    }

    // MergeLatest outlet must be the graph outlet.
    if ml_stage.spec.outlets[0].id() != graph_outlet {
        return None;
    }

    // Verify edge wiring: each Unzip outlet must connect to one MergeLatest inlet.
    let unzip_outlet_ids: Vec<PortId> =
        unzip_stage.spec.outlets.iter().map(AnyOutlet::id).collect();
    let ml_inlet_ids: Vec<PortId> = ml_stage.spec.inlets.iter().map(AnyInlet::id).collect();

    let mut outlet_to_ml_inlet: Vec<Option<usize>> = vec![None; k];
    for edge in edges {
        if let Some(uo_idx) = unzip_outlet_ids.iter().position(|&id| id == edge.outlet) {
            if let Some(mi_idx) = ml_inlet_ids.iter().position(|&id| id == edge.inlet) {
                outlet_to_ml_inlet[uo_idx] = Some(mi_idx);
            } else {
                return None; // Unzip outlet goes somewhere unexpected.
            }
        }
    }
    if outlet_to_ml_inlet.iter().any(|x| x.is_none()) {
        return None; // Not all Unzip outlets are wired to MergeLatest.
    }

    // Only k=2 Unzip is supported (same as MergeSequence typed plan).
    if k != 2 {
        return None;
    }

    // Downcast typed_split to Arc<dyn Fn(In) -> (T, T) + Send + Sync>.
    type SplitFn<A, B> = Arc<dyn Fn(A) -> (B, B) + Send + Sync>;
    let typed_split = typed_split_any.downcast_ref::<SplitFn<In, T>>()?;
    let typed_split = Arc::clone(typed_split);

    // Verify typed_snapshot can be downcast (plan-time only; not called per-element).
    type SnapshotFn<U> = Arc<dyn Fn(&[Option<U>]) -> Vec<U> + Send + Sync>;
    typed_snapshot_any.downcast_ref::<SnapshotFn<T>>()?;

    // Build per-inlet split closures respecting the wiring order.
    #[allow(clippy::type_complexity)]
    let mut splits: Vec<Option<Arc<dyn Fn(In) -> T + Send + Sync>>> = vec![None; k];

    let split0 = Arc::clone(&typed_split);
    let split1 = Arc::clone(&typed_split);

    let ml_idx_for_out0 = outlet_to_ml_inlet[0].unwrap();
    let ml_idx_for_out1 = outlet_to_ml_inlet[1].unwrap();

    splits[ml_idx_for_out0] = Some(Arc::new(move |input: In| split0(input).0));
    splits[ml_idx_for_out1] = Some(Arc::new(move |input: In| split1(input).1));

    if splits.iter().any(|s| s.is_none()) {
        return None;
    }
    let splits: Vec<Arc<dyn Fn(In) -> T + Send + Sync>> =
        splits.into_iter().map(|s| s.unwrap()).collect();

    // Retrieve eager_complete from the MergeLatest stage kind.
    let eager_complete = match &ml_stage.spec.kind {
        StageKind::MergeLatest { eager_complete, .. } => *eager_complete,
        _ => return None,
    };

    Some(TypedMergeLatestPlan {
        splits,
        core: MergeLatestCore::new(k, eager_complete),
    })
}

// ---------------------------------------------------------------------------
// Phase 3b — per-run typed MergeLatest plan dispatcher
// ---------------------------------------------------------------------------

/// Type alias for the boxed per-run MergeLatest runner returned by
/// [`try_build_typed_merge_latest_dispatch`].
///
/// The runner consumes a `&mut dyn Iterator<Item = In>` and returns
/// `StreamResult<Vec<Out>>`.  It is built fresh on every call to
/// [`run_with_input_report_mode`] — no mutable state is shared between runs.
type MergeLatestRunner<In, Out> =
    Box<dyn FnOnce(&mut dyn Iterator<Item = In>) -> StreamResult<Vec<Out>>>;

/// Build a fresh one-shot typed MergeLatest runner for the given graph.
///
/// Inspects the stages to find a `MergeLatest` stage and reads its element
/// `TypeId`.  Then dispatches to the concrete `try_typed_merge_latest_plan`
/// instantiation for the matching type among the **bounded set** of 17 common
/// primitive/`String` types.
///
/// Returns `Some(runner)` if the topology matches and the element type is in
/// the supported set.  Returns `None` otherwise — callers fall back to the
/// erased executor (in `Auto` mode) or return a typed-unsupported error
/// (`TypedOnly`).
///
/// **Bounded set note**: covers all benchmark and typical user types.
/// A custom `T` not in this list falls back to the erased executor in `Auto`
/// mode; there is no silent data-loss.
///
/// **Blueprint independence**: each call builds a new [`TypedMergeLatestPlan`]
/// with its own [`MergeLatestCore`]; running the same blueprint concurrently
/// or sequentially produces independent, correct results.
pub(crate) fn try_build_typed_merge_latest_dispatch<In, Out>(
    stages: &[super::builder::StageRecord],
    edges: &[super::builder::Edge],
    inlet: PortId,
    outlet: PortId,
) -> Option<MergeLatestRunner<In, Out>>
where
    In: Clone + Send + 'static,
    Out: Send + 'static,
{
    // Read the element TypeId from the MergeLatest stage's inlet ports.
    let elem_type_id: TypeId = stages.iter().find_map(|s| {
        if let StageKind::MergeLatest { .. } = &s.spec.kind {
            s.spec.inlets.first().map(|i| i.type_id())
        } else {
            None
        }
    })?;

    // Dispatch to the concrete instantiation for each supported element type.
    // The TypeId guard ensures we only call try_typed_merge_latest_plan when
    // T matches; the plan builder's own checks validate the full topology.
    //
    // The `Box<dyn Any>` round-trip for the output vector (`output` is a
    // `Vec<Vec<$T>>` which is the same as `Vec<Out>` when `Out == Vec<$T>`)
    // is a single allocation of the Vec header (24 bytes); no element data is
    // copied.  This is the minimum necessary to safely reinterpret the output
    // type without `unsafe` code.
    macro_rules! try_elem {
        ($($T:ty),*) => {
            $(
                if elem_type_id == TypeId::of::<$T>() {
                    let mut plan = try_typed_merge_latest_plan::<In, $T>(stages, edges, inlet, outlet)?;
                    // At this point Out == Vec<$T> (enforced by port TypeId checks
                    // inside try_typed_merge_latest_plan).
                    let runner: MergeLatestRunner<In, Out> = Box::new(
                        move |iter: &mut dyn Iterator<Item = In>| {
                            plan.reset();
                            let hint = iter.size_hint().0;
                            let mut output: Vec<Vec<$T>> = Vec::with_capacity(hint);
                            for item in iter {
                                plan.push_item(item, &mut output);
                            }
                            plan.finish();
                            // Vec<Vec<$T>> → Vec<Out>: safe because Out == Vec<$T>
                            // guaranteed by TypeId dispatch above and by
                            // try_typed_merge_latest_plan's port TypeId checks.
                            let boxed: Box<dyn Any + Send> = Box::new(output);
                            boxed
                                .downcast::<Vec<Out>>()
                                .map(|b| *b)
                                .map_err(|_| StreamError::Failed(
                                    "merge-latest typed runner: Out type mismatch".into()
                                ))
                        }
                    );
                    return Some(runner);
                }
            )*
        };
    }

    try_elem!(
        u8, u16, u32, u64, u128, usize, i8, i16, i32, i64, i128, isize, f32, f64, bool, String
    );

    // Element type not in the supported set → fall back to erased executor.
    None
}

// ---------------------------------------------------------------------------

pub(super) enum BoundaryCountExecutor {
    #[cfg(test)]
    Threaded,
    Ractor,
}

impl BoundaryCountExecutor {
    pub(super) fn run_count<I, T>(
        &self,
        input: I,
        segments: TypedLinearSegments<T>,
        config: AsyncBoundaryExecutionConfig,
    ) -> StreamResult<FusedTerminalReport<usize>>
    where
        I: IntoIterator<Item = T> + Send,
        I::IntoIter: Send + 'static,
        T: Send + 'static,
    {
        match self {
            #[cfg(test)]
            Self::Threaded => run_threaded_async_linear_count(input, segments, config),
            Self::Ractor => run_ractor_async_linear_count(input, segments, config),
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[derive(Default)]
    struct BufferedFlowState {
        queued: VecDeque<i32>,
        upstream_closed: bool,
        pull_calls: usize,
        finish_calls: usize,
    }

    struct BufferedFlowOnPull {
        state: Arc<Mutex<BufferedFlowState>>,
    }

    impl GraphStage for BufferedFlowOnPull {
        type Shape = FlowShape<i32, i32>;

        fn name(&self) -> &str {
            "BufferedFlowOnPull"
        }

        fn allocate_shape(&self, _allocator: &mut PortAllocator) -> Self::Shape {
            let first_id = next_port_id_block(2);
            FlowShape::new(
                Inlet::with_id(first_id, "buffered-flow.in"),
                Outlet::with_id(first_id.offset(1), "buffered-flow.out"),
            )
        }

        fn stage_spec(&self, shape: &Self::Shape) -> StageSpec {
            StageSpec::opaque(self.name(), shape.inlets(), shape.outlets())
        }

        fn create_logic(&self, shape: &Self::Shape) -> GraphStageLogic {
            struct In {
                state: Arc<Mutex<BufferedFlowState>>,
            }

            impl InHandler for In {
                fn on_push(
                    &mut self,
                    logic: &mut GraphStageLogic,
                    inlet: AnyInlet,
                ) -> StreamResult<()> {
                    let value: i32 = logic.grab_datum(inlet.id()).and_then(|value| {
                        downcast_datum(value, "grab", || format!("inlet#{}", inlet.id().as_usize()))
                    })?;
                    self.state.lock().unwrap().queued.push_back(value);
                    Ok(())
                }

                fn on_upstream_finish(
                    &mut self,
                    _logic: &mut GraphStageLogic,
                    _inlet: AnyInlet,
                ) -> StreamResult<()> {
                    self.state.lock().unwrap().upstream_closed = true;
                    Ok(())
                }
            }

            struct Out {
                outlet: Outlet<i32>,
                state: Arc<Mutex<BufferedFlowState>>,
            }

            impl OutHandler for Out {
                fn on_pull(
                    &mut self,
                    logic: &mut GraphStageLogic,
                    _outlet: AnyOutlet,
                ) -> StreamResult<()> {
                    let (next, upstream_closed) = {
                        let mut state = self.state.lock().unwrap();
                        state.pull_calls += 1;
                        (state.queued.pop_front(), state.upstream_closed)
                    };
                    if let Some(value) = next {
                        logic.emit(&self.outlet, value)
                    } else if upstream_closed {
                        logic.complete(&self.outlet)
                    } else {
                        Ok(())
                    }
                }

                fn on_downstream_finish(
                    &mut self,
                    logic: &mut GraphStageLogic,
                    _outlet: AnyOutlet,
                ) -> StreamResult<()> {
                    self.state.lock().unwrap().finish_calls += 1;
                    logic.complete_stage()
                }
            }

            let mut logic = GraphStageLogic::new(shape);
            logic
                .set_handler(
                    &shape.inlet(),
                    Box::new(In {
                        state: Arc::clone(&self.state),
                    }),
                )
                .unwrap();
            logic
                .set_out_handler(
                    &shape.outlet(),
                    Box::new(Out {
                        outlet: shape.outlet(),
                        state: Arc::clone(&self.state),
                    }),
                )
                .unwrap();
            logic
        }
    }

    struct EmitMultipleThenFailOnPush;

    impl GraphStage for EmitMultipleThenFailOnPush {
        type Shape = FlowShape<i32, i32>;

        fn name(&self) -> &str {
            "EmitMultipleThenFailOnPush"
        }

        fn allocate_shape(&self, _allocator: &mut PortAllocator) -> Self::Shape {
            let first_id = next_port_id_block(2);
            FlowShape::new(
                Inlet::with_id(first_id, "emit-fail.in"),
                Outlet::with_id(first_id.offset(1), "emit-fail.out"),
            )
        }

        fn stage_spec(&self, shape: &Self::Shape) -> StageSpec {
            StageSpec::opaque(self.name(), shape.inlets(), shape.outlets())
        }

        fn create_logic(&self, shape: &Self::Shape) -> GraphStageLogic {
            struct Handler {
                outlet: Outlet<i32>,
            }

            impl InHandler for Handler {
                fn on_push(
                    &mut self,
                    logic: &mut GraphStageLogic,
                    _inlet: AnyInlet,
                ) -> StreamResult<()> {
                    logic.emit_multiple(&self.outlet, [1, 2])?;
                    Err(StreamError::Failed("emit_multiple boom".into()))
                }
            }

            let mut logic = GraphStageLogic::new(shape);
            logic
                .set_handler(
                    &shape.inlet(),
                    Box::new(Handler {
                        outlet: shape.outlet(),
                    }),
                )
                .unwrap();
            logic
        }
    }

    struct ReadNThenFailOnFinish;

    struct EmitMultipleOnPush;

    impl GraphStage for EmitMultipleOnPush {
        type Shape = FlowShape<i32, i32>;

        fn name(&self) -> &str {
            "EmitMultipleOnPush"
        }

        fn allocate_shape(&self, _allocator: &mut PortAllocator) -> Self::Shape {
            let first_id = next_port_id_block(2);
            FlowShape::new(
                Inlet::with_id(first_id, "emit-multiple.in"),
                Outlet::with_id(first_id.offset(1), "emit-multiple.out"),
            )
        }

        fn stage_spec(&self, shape: &Self::Shape) -> StageSpec {
            StageSpec::opaque(self.name(), shape.inlets(), shape.outlets())
        }

        fn create_logic(&self, shape: &Self::Shape) -> GraphStageLogic {
            struct Handler {
                outlet: Outlet<i32>,
            }

            impl InHandler for Handler {
                fn on_push(
                    &mut self,
                    logic: &mut GraphStageLogic,
                    _inlet: AnyInlet,
                ) -> StreamResult<()> {
                    logic.emit_multiple(&self.outlet, [1, 2])
                }
            }

            let mut logic = GraphStageLogic::new(shape);
            logic
                .set_handler(
                    &shape.inlet(),
                    Box::new(Handler {
                        outlet: shape.outlet(),
                    }),
                )
                .unwrap();
            logic
        }
    }

    impl GraphStage for ReadNThenFailOnFinish {
        type Shape = FlowShape<i32, i32>;

        fn name(&self) -> &str {
            "ReadNThenFailOnFinish"
        }

        fn allocate_shape(&self, _allocator: &mut PortAllocator) -> Self::Shape {
            let first_id = next_port_id_block(2);
            FlowShape::new(
                Inlet::with_id(first_id, "read-n.in"),
                Outlet::with_id(first_id.offset(1), "read-n.out"),
            )
        }

        fn stage_spec(&self, shape: &Self::Shape) -> StageSpec {
            StageSpec::opaque(self.name(), shape.inlets(), shape.outlets())
        }

        fn create_logic(&self, shape: &Self::Shape) -> GraphStageLogic {
            struct Handler {
                inlet: Inlet<i32>,
                armed: bool,
            }

            impl InHandler for Handler {
                fn on_push(
                    &mut self,
                    logic: &mut GraphStageLogic,
                    _inlet: AnyInlet,
                ) -> StreamResult<()> {
                    if !self.armed {
                        self.armed = true;
                        logic.read_n(&self.inlet, 2, |_values| {}, |_values| {})
                    } else {
                        Ok(())
                    }
                }

                fn on_upstream_finish(
                    &mut self,
                    _logic: &mut GraphStageLogic,
                    _inlet: AnyInlet,
                ) -> StreamResult<()> {
                    Err(StreamError::Failed("read_n finish boom".into()))
                }
            }

            let mut logic = GraphStageLogic::new(shape);
            logic
                .set_handler(
                    &shape.inlet(),
                    Box::new(Handler {
                        inlet: shape.inlet(),
                        armed: false,
                    }),
                )
                .unwrap();
            logic
        }
    }

    fn single_opaque_stage_graph<G>(stage: G) -> GraphBlueprint<FlowShape<i32, i32>>
    where
        G: GraphStage<Shape = FlowShape<i32, i32>>,
    {
        GraphDsl::create(|builder| builder.add(stage)).unwrap()
    }

    #[test]
    fn process_push_restores_handler_before_emit_multiple_error_propagates() {
        let graph = single_opaque_stage_graph(EmitMultipleThenFailOnPush);
        let inlet = graph.shape.inlet().id();
        let mut executor = FusedExecutor::new(&graph, FusedExecutionConfig::default());

        let result = executor.process_stage(0, inlet, datum(10));

        assert!(matches!(
            result,
            Err(StreamError::Failed(message)) if message == "emit_multiple boom"
        ));
        assert!(
            executor.opaque_logics[0]
                .as_mut()
                .unwrap()
                .get_in_handler_mut(inlet)
                .is_some()
        );
    }

    #[test]
    fn process_completion_restores_handler_before_read_n_finish_error_propagates() {
        let graph = single_opaque_stage_graph(ReadNThenFailOnFinish);
        let inlet = graph.shape.inlet().id();
        let mut executor = FusedExecutor::new(&graph, FusedExecutionConfig::default());

        executor.process_stage(0, inlet, datum(1)).unwrap();
        executor.process_stage(0, inlet, datum(2)).unwrap();
        let result = executor.process_completion(0, inlet);

        assert!(matches!(
            result,
            Err(StreamError::Failed(message)) if message == "read_n finish boom"
        ));
        assert!(
            executor.opaque_logics[0]
                .as_mut()
                .unwrap()
                .get_in_handler_mut(inlet)
                .is_some()
        );
    }

    #[test]
    fn opaque_request_drives_out_handler_for_buffered_output() {
        let state = Arc::new(Mutex::new(BufferedFlowState::default()));
        let graph = single_opaque_stage_graph(BufferedFlowOnPull {
            state: Arc::clone(&state),
        });
        let inlet = graph.shape.inlet().id();
        let outlet = graph.shape.outlet().id();
        let mut executor = FusedExecutor::new(&graph, FusedExecutionConfig::default());
        let mut output = Vec::<i32>::new();
        let mut output_sink = VecOutputSink {
            output: &mut output,
        };

        executor
            .deliver(inlet, datum(7_i32), outlet, &mut output_sink)
            .unwrap();
        assert!(output_sink.output.is_empty());

        executor.request(outlet, outlet, &mut output_sink).unwrap();

        assert_eq!(&*output_sink.output, &[7]);
        let state = state.lock().unwrap();
        assert_eq!(state.pull_calls, 2);
        assert_eq!(state.finish_calls, 0);
    }

    #[test]
    fn opaque_downstream_finish_before_first_demand_invokes_out_handler() {
        let state = Arc::new(Mutex::new(BufferedFlowState::default()));
        let graph = single_opaque_stage_graph(BufferedFlowOnPull {
            state: Arc::clone(&state),
        });
        let outlet = graph.shape.outlet().id();
        let mut executor = FusedExecutor::new(&graph, FusedExecutionConfig::default());
        let mut output = Vec::<i32>::new();
        let mut output_sink = VecOutputSink {
            output: &mut output,
        };

        executor
            .downstream_finish(outlet, outlet, &mut output_sink)
            .unwrap();
        executor.request(outlet, outlet, &mut output_sink).unwrap();

        assert!(output_sink.output.is_empty());
        let state = state.lock().unwrap();
        assert_eq!(state.pull_calls, 0);
        assert_eq!(state.finish_calls, 1);
    }

    #[test]
    fn opaque_downstream_finish_drops_buffered_output_after_upstream_complete() {
        let state = Arc::new(Mutex::new(BufferedFlowState::default()));
        let graph = single_opaque_stage_graph(BufferedFlowOnPull {
            state: Arc::clone(&state),
        });
        let inlet = graph.shape.inlet().id();
        let outlet = graph.shape.outlet().id();
        let mut executor = FusedExecutor::new(&graph, FusedExecutionConfig::default());
        let mut output = Vec::<i32>::new();
        let mut output_sink = VecOutputSink {
            output: &mut output,
        };

        executor
            .deliver(inlet, datum(11_i32), outlet, &mut output_sink)
            .unwrap();
        executor.complete(inlet, outlet, &mut output_sink).unwrap();
        executor
            .downstream_finish(outlet, outlet, &mut output_sink)
            .unwrap();
        executor.request(outlet, outlet, &mut output_sink).unwrap();

        assert!(output_sink.output.is_empty());
        let state = state.lock().unwrap();
        assert_eq!(state.finish_calls, 1);
    }

    #[test]
    fn broadcast_cancels_upstream_only_after_all_outlets_cancel() {
        let graph = GraphDsl::try_create(|builder| {
            let broadcast = builder.add(Broadcast::<i32>::new(2));
            let merge = builder.add(Merge::<i32>::new(2));
            builder.connect(broadcast.outlet(0)?, merge.inlet(0)?)?;
            builder.connect(broadcast.outlet(1)?, merge.inlet(1)?)?;
            Ok(FlowShape::new(broadcast.inlet(), merge.outlet()))
        })
        .unwrap();

        let mut executor = FusedExecutor::new(&graph, FusedExecutionConfig::default());
        let broadcast_index = *executor
            .stage_by_inlet
            .get(&graph.shape.inlet().id())
            .unwrap();
        let first = graph.stages[broadcast_index].spec.outlets[0].id();
        let second = graph.stages[broadcast_index].spec.outlets[1].id();

        let first_transition = executor
            .process_downstream_finish(broadcast_index, first)
            .unwrap();
        assert!(first_transition.cancelled_inlets.is_empty());

        let second_transition = executor
            .process_downstream_finish(broadcast_index, second)
            .unwrap();
        assert_eq!(
            second_transition.cancelled_inlets,
            vec![graph.stages[broadcast_index].spec.inlets[0].id()]
        );
    }

    #[test]
    fn downstream_finish_propagates_through_merge_and_broadcast() {
        let graph = GraphDsl::try_create(|builder| {
            let broadcast = builder.add(Broadcast::<i32>::new(2));
            let merge = builder.add(Merge::<i32>::new(2));
            builder.connect(broadcast.outlet(0)?, merge.inlet(0)?)?;
            builder.connect(broadcast.outlet(1)?, merge.inlet(1)?)?;
            Ok(FlowShape::new(broadcast.inlet(), merge.outlet()))
        })
        .unwrap();

        let outlet = graph.shape.outlet().id();
        let mut executor = FusedExecutor::new(&graph, FusedExecutionConfig::default());
        let mut output = Vec::<i32>::new();
        let mut output_sink = VecOutputSink {
            output: &mut output,
        };

        executor
            .downstream_finish(outlet, outlet, &mut output_sink)
            .unwrap();

        let broadcast_index = *executor
            .stage_by_inlet
            .get(&graph.shape.inlet().id())
            .unwrap();
        let StageState::Broadcast {
            live_outlets,
            cancelled_outlets,
            ..
        } = &executor.stage_states[broadcast_index]
        else {
            panic!("expected broadcast state");
        };
        assert_eq!(*live_outlets, 0);
        assert_eq!(cancelled_outlets, &vec![true, true]);
    }

    #[test]
    fn partition_holds_routed_element_until_target_outlet_pulls() {
        let graph = GraphDsl::create(|builder| {
            builder.add(Partition::<i32>::new(2, |value| usize::from(*value >= 10)))
        })
        .unwrap();

        let stage_index = 0usize;
        let inlet = graph.shape.inlet().id();
        let out0 = graph.shape.outlet(0).unwrap().id();
        let out1 = graph.shape.outlet(1).unwrap().id();
        let mut executor = FusedExecutor::new(&graph, FusedExecutionConfig::default());

        executor.process_pull(stage_index, out0).unwrap();
        let transition = executor
            .process_stage(stage_index, inlet, datum(11_i32))
            .unwrap();
        assert!(matches!(transition.emissions, StageEmissions::None));

        let pull_transition = executor.process_pull(stage_index, out1).unwrap();
        match pull_transition.emissions {
            StageEmissions::One(port, value) => {
                assert_eq!(port, out1);
                assert_eq!(
                    downcast_datum::<i32, _>(value, "emit", || "Partition.out1").unwrap(),
                    11
                );
            }
            _ => panic!("expected one pending partition emission"),
        }
    }

    #[test]
    fn partition_cancels_upstream_only_after_all_outlets_cancel_when_not_eager() {
        let graph = GraphDsl::create(|builder| {
            builder.add(Partition::<i32>::new(2, |value| usize::from(*value >= 10)))
        })
        .unwrap();

        let stage_index = 0usize;
        let out0 = graph.shape.outlet(0).unwrap().id();
        let out1 = graph.shape.outlet(1).unwrap().id();
        let mut executor = FusedExecutor::new(&graph, FusedExecutionConfig::default());

        let first = executor
            .process_downstream_finish(stage_index, out0)
            .unwrap();
        assert!(first.cancelled_inlets.is_empty());

        let second = executor
            .process_downstream_finish(stage_index, out1)
            .unwrap();
        assert_eq!(second.cancelled_inlets, vec![graph.shape.inlet().id()]);
    }

    #[test]
    fn unzip_continues_emitting_to_live_outlet_after_peer_cancels() {
        let graph =
            GraphDsl::create(|builder| builder.add(Unzip::<i32, &'static str>::new())).unwrap();

        let stage_index = 0usize;
        let inlet = graph.shape.inlet().id();
        let out0 = graph.shape.out0().id();
        let out1 = graph.shape.out1().id();
        let mut executor = FusedExecutor::new(&graph, FusedExecutionConfig::default());

        executor.process_pull(stage_index, out0).unwrap();
        executor.process_pull(stage_index, out1).unwrap();
        let cancel = executor
            .process_downstream_finish(stage_index, out1)
            .unwrap();
        assert!(cancel.cancelled_inlets.is_empty());

        let transition = executor
            .process_stage(stage_index, inlet, datum((7_i32, "seven")))
            .unwrap();
        match transition.emissions {
            StageEmissions::One(port, value) => {
                assert_eq!(port, out0);
                assert_eq!(
                    downcast_datum::<i32, _>(value, "emit", || "Unzip.out0").unwrap(),
                    7
                );
            }
            StageEmissions::Many(values) => {
                assert_eq!(values.len(), 1);
                assert_eq!(values[0].0, out0);
            }
            _ => panic!("expected emission to the remaining live unzip outlet"),
        }
    }

    #[test]
    fn unzip_cancels_upstream_only_after_both_outlets_cancel() {
        let graph =
            GraphDsl::create(|builder| builder.add(Unzip::<i32, &'static str>::new())).unwrap();

        let stage_index = 0usize;
        let out0 = graph.shape.out0().id();
        let out1 = graph.shape.out1().id();
        let mut executor = FusedExecutor::new(&graph, FusedExecutionConfig::default());

        let first = executor
            .process_downstream_finish(stage_index, out0)
            .unwrap();
        assert!(first.cancelled_inlets.is_empty());

        let second = executor
            .process_downstream_finish(stage_index, out1)
            .unwrap();
        assert_eq!(second.cancelled_inlets, vec![graph.shape.inlet().id()]);
    }

    #[test]
    fn opaque_internal_outlet_repulls_after_first_emission() {
        let graph = GraphDsl::try_create(|builder| {
            let opaque = builder.add(EmitMultipleOnPush);
            let identity = builder.add(Identity::<i32>::new());
            builder.connect(opaque.outlet(), identity.inlet())?;
            Ok(FlowShape::new(opaque.inlet(), identity.outlet()))
        })
        .unwrap();

        assert_eq!(graph.run_with_input([10]).unwrap(), vec![1, 2]);
    }

    // -- Phase 0 (WP-18): ExecutorMode scaffolding tests -------------------

    /// Verifies that `Auto` and `ErasedOnly` produce identical results on a
    /// representative junction graph (Broadcast→Merge) and that `TypedOnly`
    /// errors cleanly (junctions are not yet on the typed path).
    ///
    /// Graph topology:
    ///   in → Broadcast(2) → out0 → Merge(2).in0 → out
    ///                      → out1 → Merge(2).in1
    ///
    /// This exercises fan-out + fan-in junctions that cannot hit the typed
    /// linear fast path, so Auto falls back to the erased executor.
    #[test]
    fn executor_mode_auto_erased_identical_typed_errors() {
        let graph = GraphDsl::try_create(|builder| {
            let broadcast = builder.add(Broadcast::<i32>::new(2));
            let merge = builder.add(Merge::<i32>::new(2));
            builder.connect(broadcast.outlet(0)?, merge.inlet(0)?)?;
            builder.connect(broadcast.outlet(1)?, merge.inlet(1)?)?;
            Ok(FlowShape::new(broadcast.inlet(), merge.outlet()))
        })
        .unwrap();

        let input = vec![1, 2, 3];

        let auto_result = graph
            .run_with_input_mode(input.clone(), ExecutorMode::Auto)
            .unwrap();
        let erased_result = graph
            .run_with_input_mode(input.clone(), ExecutorMode::ErasedOnly)
            .unwrap();

        // Auto and ErasedOnly must produce identical output (broadcast
        // duplicates each element onto both merge inlets; merge emits both).
        assert_eq!(auto_result, erased_result);
        // Each element is broadcast to both merge inlets, so output is 2× input.
        assert_eq!(auto_result.len(), input.len() * 2);

        // TypedOnly must error because junctions are not on the typed path yet.
        let typed_result = graph.run_with_input_mode(input.clone(), ExecutorMode::TypedOnly);
        assert!(
            matches!(
                typed_result,
                Err(StreamError::GraphValidation(ref msg))
                if msg.contains("typed executor does not support this graph shape")
            ),
            "expected TypedOnly to error for junction graph, got: {typed_result:?}"
        );
    }

    // -- Phase 2 (WP-18): typed-vs-erased equivalence tests ----------------

    /// Builds an identity chain of `n` stages with the given type.
    fn identity_chain_bp(n: usize) -> GraphBlueprint<FlowShape<i64, i64>> {
        assert!(n >= 1);
        GraphDsl::try_create(|builder| {
            let first = builder.add(Identity::<i64>::new());
            let inlet = first.inlet();
            let mut outlet = first.outlet();
            for _ in 1..n {
                let next = builder.add(Identity::<i64>::new());
                builder.connect(outlet, next.inlet())?;
                outlet = next.outlet();
            }
            Ok(FlowShape::new(inlet, outlet))
        })
        .unwrap()
    }

    /// Builds a map chain of `n` stages that double each value.
    fn map_chain_bp(n: usize) -> GraphBlueprint<FlowShape<i64, i64>> {
        assert!(n >= 1);
        GraphDsl::try_create(|builder| {
            let first = builder.add(MapStage::new(|x: i64| x.wrapping_mul(2)));
            let inlet = first.inlet();
            let mut outlet = first.outlet();
            for _ in 1..n {
                let next = builder.add(MapStage::new(|x: i64| x.wrapping_mul(2)));
                builder.connect(outlet, next.inlet())?;
                outlet = next.outlet();
            }
            Ok(FlowShape::new(inlet, outlet))
        })
        .unwrap()
    }

    /// `ErasedOnly` and `TypedOnly` (via `run_with_input_mode`) produce
    /// identical output on a 5-stage identity graph.
    #[test]
    fn typed_erased_equivalence_identity_collect() {
        let graph = identity_chain_bp(5);
        let input: Vec<i64> = (0..20).collect();

        let erased = graph
            .run_with_input_mode(input.clone(), ExecutorMode::ErasedOnly)
            .unwrap();
        let typed = graph
            .run_with_input_mode(input.clone(), ExecutorMode::TypedOnly)
            .unwrap();

        assert_eq!(
            typed, erased,
            "typed and erased paths disagree on identity×5"
        );
    }

    /// `ErasedOnly` and `TypedOnly` produce identical count on a 5-stage
    /// identity graph.
    #[test]
    fn typed_erased_equivalence_identity_count() {
        let graph = identity_chain_bp(5);
        let input: Vec<i64> = (0..20).collect();
        let config = FusedExecutionConfig::default();

        let erased = graph
            .run_count_with_input_report_mode(input.clone(), config, ExecutorMode::ErasedOnly)
            .unwrap()
            .result;
        let typed = graph
            .run_count_with_input_report_mode(input.clone(), config, ExecutorMode::TypedOnly)
            .unwrap()
            .result;

        assert_eq!(typed, erased, "typed and erased count differ on identity×5");
    }

    /// `ErasedOnly` and `TypedOnly` produce identical fold result on a 5-stage
    /// map graph (each stage doubles the value).
    #[test]
    fn typed_erased_equivalence_map_fold() {
        let graph = map_chain_bp(5);
        let input: Vec<i64> = (1..=10).collect();
        let config = FusedExecutionConfig::default();

        let erased = graph
            .run_fold_with_input_report_mode(
                input.clone(),
                0i64,
                |acc, x| acc.wrapping_add(x),
                config,
                ExecutorMode::ErasedOnly,
            )
            .unwrap()
            .result;
        let typed = graph
            .run_fold_with_input_report_mode(
                input.clone(),
                0i64,
                |acc, x| acc.wrapping_add(x),
                config,
                ExecutorMode::TypedOnly,
            )
            .unwrap()
            .result;

        assert_eq!(typed, erased, "typed and erased fold differ on map×5");
    }

    /// `ErasedOnly` and `TypedOnly` produce identical output on a 5-stage map
    /// graph.
    #[test]
    fn typed_erased_equivalence_map_collect() {
        let graph = map_chain_bp(5);
        let input: Vec<i64> = (0..20).collect();

        let erased = graph
            .run_with_input_mode(input.clone(), ExecutorMode::ErasedOnly)
            .unwrap();
        let typed = graph
            .run_with_input_mode(input.clone(), ExecutorMode::TypedOnly)
            .unwrap();

        assert_eq!(typed, erased, "typed and erased paths disagree on map×5");
    }

    /// `TypedOnly` on a junction graph (Broadcast→Merge) returns a
    /// `GraphValidation` error (junctions not yet on the typed path).
    #[test]
    fn typed_only_errors_on_junction_graph() {
        let graph = GraphDsl::try_create(|builder| {
            let broadcast = builder.add(Broadcast::<i32>::new(2));
            let merge = builder.add(Merge::<i32>::new(2));
            builder.connect(broadcast.outlet(0)?, merge.inlet(0)?)?;
            builder.connect(broadcast.outlet(1)?, merge.inlet(1)?)?;
            Ok(FlowShape::new(broadcast.inlet(), merge.outlet()))
        })
        .unwrap();

        let result = graph.run_with_input_mode(vec![1], ExecutorMode::TypedOnly);
        assert!(
            matches!(
                result,
                Err(StreamError::GraphValidation(ref msg))
                if msg.contains("typed executor does not support this graph shape")
            ),
            "expected TypedOnly to error on junction, got: {result:?}"
        );
    }

    /// `Auto` on a junction graph falls back silently to the erased executor.
    #[test]
    fn auto_falls_back_silently_for_junction_graph() {
        let graph = GraphDsl::try_create(|builder| {
            let broadcast = builder.add(Broadcast::<i32>::new(2));
            let merge = builder.add(Merge::<i32>::new(2));
            builder.connect(broadcast.outlet(0)?, merge.inlet(0)?)?;
            builder.connect(broadcast.outlet(1)?, merge.inlet(1)?)?;
            Ok(FlowShape::new(broadcast.inlet(), merge.outlet()))
        })
        .unwrap();

        let result = graph.run_with_input_mode(vec![1, 2, 3], ExecutorMode::Auto);
        assert!(result.is_ok(), "Auto should succeed (fallback to erased)");
        assert_eq!(result.unwrap().len(), 6); // broadcast × 2
    }

    // -- Phase 3a (WP-18): typed MergeSequence equivalence tests ------------

    /// Build the Unzip → MergeSequence graph used in the benchmark.
    fn merge_sequence_graph() -> GraphBlueprint<FlowShape<(u64, u64), u64>> {
        GraphDsl::try_create(|builder| {
            let unzip = builder.add(Unzip::<u64, u64>::new());
            let merge = builder.add(MergeSequence::<u64>::new(2, |item| *item));
            builder.connect(unzip.out0(), merge.inlet(0)?)?;
            builder.connect(unzip.out1(), merge.inlet(1)?)?;
            Ok(FlowShape::new(unzip.inlet(), merge.outlet()))
        })
        .unwrap()
    }

    /// `TypedOnly` accepts the Unzip → MergeSequence topology (does not
    /// error with "typed executor does not support this graph shape").
    #[test]
    fn typed_only_accepts_merge_sequence_topology() {
        let graph = merge_sequence_graph();
        let result =
            graph.run_with_input_mode(vec![(0u64, 1u64), (2u64, 3u64)], ExecutorMode::TypedOnly);
        assert!(
            result.is_ok(),
            "TypedOnly should accept Unzip→MergeSequence topology, got: {result:?}"
        );
    }

    /// `ErasedOnly` and `TypedOnly` produce identical output on an in-order
    /// Unzip → MergeSequence graph.
    #[test]
    fn typed_erased_equivalence_merge_sequence_in_order() {
        let graph = merge_sequence_graph();
        let input: Vec<(u64, u64)> = (0..10).step_by(2).map(|i| (i, i + 1)).collect();

        let erased = graph
            .run_with_input_mode(input.clone(), ExecutorMode::ErasedOnly)
            .unwrap();
        let typed = graph
            .run_with_input_mode(input.clone(), ExecutorMode::TypedOnly)
            .unwrap();

        assert_eq!(
            typed, erased,
            "typed and erased disagree on in-order merge_sequence"
        );
        // Verify the output is actually sorted.
        let expected: Vec<u64> = (0..10).collect();
        assert_eq!(typed, expected);
    }

    /// `ErasedOnly` and `TypedOnly` produce identical output on an adversarial
    /// (out-of-order) Unzip → MergeSequence graph where pairs are (even, odd)
    /// with `even > odd - 1`, forcing out-of-order arrivals.
    #[test]
    fn typed_erased_equivalence_merge_sequence_out_of_order() {
        let graph = merge_sequence_graph();
        // Pairs (1,0), (3,2), (5,4): out0 gets larger sequence, out1 smaller.
        let input: Vec<(u64, u64)> = vec![(1, 0), (3, 2), (5, 4)];

        let erased = graph
            .run_with_input_mode(input.clone(), ExecutorMode::ErasedOnly)
            .unwrap();
        let typed = graph
            .run_with_input_mode(input.clone(), ExecutorMode::TypedOnly)
            .unwrap();

        assert_eq!(
            typed, erased,
            "typed and erased disagree on out-of-order merge_sequence"
        );
        // Output should be sorted.
        assert_eq!(typed, vec![0u64, 1, 2, 3, 4, 5]);
    }

    /// Both `ErasedOnly` and `TypedOnly` fail with a sequence-gap error when
    /// the pending items can never be resolved at completion.
    ///
    /// This is the #78 regression: `merge_sequence_fails_on_gap_at_completion`.
    #[test]
    fn typed_erased_equivalence_merge_sequence_gap_failure() {
        // Sequences 1 and 2 arrive (via the unzip split of pair (1, 2)), but
        // sequence 0 never does.  Both paths must return an error.
        let graph = merge_sequence_graph();
        let input = vec![(1u64, 2u64)];

        let erased = graph.run_with_input_mode(input.clone(), ExecutorMode::ErasedOnly);
        let typed = graph.run_with_input_mode(input.clone(), ExecutorMode::TypedOnly);

        assert!(
            matches!(&erased, Err(StreamError::Failed(msg)) if msg.contains("expected sequence")),
            "ErasedOnly should fail on gap: {erased:?}"
        );
        assert!(
            matches!(&typed, Err(StreamError::Failed(msg)) if msg.contains("expected sequence")),
            "TypedOnly should fail on gap: {typed:?}"
        );
    }

    /// `ErasedOnly` and `TypedOnly` both complete cleanly with a single-item
    /// in-order input.
    #[test]
    fn typed_erased_equivalence_merge_sequence_completion() {
        let graph = merge_sequence_graph();
        let input = vec![(0u64, 1u64)];

        let erased = graph
            .run_with_input_mode(input.clone(), ExecutorMode::ErasedOnly)
            .unwrap();
        let typed = graph
            .run_with_input_mode(input.clone(), ExecutorMode::TypedOnly)
            .unwrap();

        assert_eq!(typed, erased, "typed and erased disagree on completion");
        assert_eq!(typed, vec![0u64, 1u64]);
    }

    /// `Auto` selects the typed path for the Unzip → MergeSequence topology
    /// and produces the same result as `ErasedOnly`.
    #[test]
    fn auto_selects_typed_for_merge_sequence_topology() {
        let graph = merge_sequence_graph();
        let input: Vec<(u64, u64)> = (0..20).step_by(2).map(|i| (i, i + 1)).collect();

        let auto_result = graph
            .run_with_input_mode(input.clone(), ExecutorMode::Auto)
            .unwrap();
        let erased_result = graph
            .run_with_input_mode(input.clone(), ExecutorMode::ErasedOnly)
            .unwrap();

        assert_eq!(
            auto_result, erased_result,
            "Auto and ErasedOnly disagree on merge_sequence topology"
        );
    }

    // -- Phase 3b (WP-18): typed MergeLatest equivalence tests ---------------

    /// Build the Unzip → MergeLatest graph used in the benchmark.
    fn merge_latest_graph_exec() -> GraphBlueprint<FlowShape<(u64, u64), Vec<u64>>> {
        GraphDsl::try_create(|builder| {
            let unzip = builder.add(Unzip::<u64, u64>::new());
            let merge = builder.add(MergeLatest::<u64>::new(2, false));
            builder.connect(unzip.out0(), merge.inlet(0)?)?;
            builder.connect(unzip.out1(), merge.inlet(1)?)?;
            Ok(FlowShape::new(unzip.inlet(), merge.outlet()))
        })
        .unwrap()
    }

    /// Build a Unzip → MergeLatest graph with `eager_complete = true`.
    fn merge_latest_eager_graph_exec() -> GraphBlueprint<FlowShape<(i32, i32), Vec<i32>>> {
        GraphDsl::try_create(|builder| {
            let unzip = builder.add(Unzip::<i32, i32>::new());
            let merge = builder.add(MergeLatest::<i32>::new(2, true));
            builder.connect(unzip.out0(), merge.inlet(0)?)?;
            builder.connect(unzip.out1(), merge.inlet(1)?)?;
            Ok(FlowShape::new(unzip.inlet(), merge.outlet()))
        })
        .unwrap()
    }

    /// `TypedOnly` accepts the Unzip → MergeLatest topology.
    #[test]
    fn typed_only_accepts_merge_latest_topology() {
        let graph = merge_latest_graph_exec();
        let result =
            graph.run_with_input_mode(vec![(0u64, 1u64), (2u64, 3u64)], ExecutorMode::TypedOnly);
        assert!(
            result.is_ok(),
            "TypedOnly should accept Unzip→MergeLatest topology, got: {result:?}"
        );
    }

    /// `ErasedOnly` and `TypedOnly` produce identical latest-snapshot sequences.
    #[test]
    fn typed_erased_equivalence_merge_latest_snapshot_ordering() {
        let graph = merge_latest_graph_exec();
        // Each pair (a, b) updates both inlets; after the first pair all inlets
        // are seen and every subsequent pair also produces a snapshot.
        let input: Vec<(u64, u64)> = (0..10).map(|i| (i, i + 100)).collect();

        let erased = graph
            .run_with_input_mode(input.clone(), ExecutorMode::ErasedOnly)
            .unwrap();
        let typed = graph
            .run_with_input_mode(input.clone(), ExecutorMode::TypedOnly)
            .unwrap();

        assert_eq!(
            typed, erased,
            "typed and erased disagree on snapshot ordering"
        );
        // All snapshots should have length 2 (one entry per inlet).
        assert!(
            typed.iter().all(|s| s.len() == 2),
            "snapshots must have len 2"
        );
    }

    /// Partial-fill: first item sees only 1 of 2 inlets; no snapshot until both seen.
    #[test]
    fn typed_erased_equivalence_merge_latest_partial_fill() {
        // With a single input pair, both inlets get their first value simultaneously
        // (Unzip splits into two values), so exactly one snapshot is produced.
        let graph = merge_latest_graph_exec();
        let input = vec![(5u64, 42u64)];

        let erased = graph
            .run_with_input_mode(input.clone(), ExecutorMode::ErasedOnly)
            .unwrap();
        let typed = graph
            .run_with_input_mode(input.clone(), ExecutorMode::TypedOnly)
            .unwrap();

        assert_eq!(typed, erased, "typed and erased disagree on partial-fill");
        // Exactly one snapshot, containing [5, 42] in inlet order.
        assert_eq!(typed.len(), 1, "expected exactly one snapshot");
    }

    /// Eager-complete: #78 regression must pass in BOTH `ErasedOnly` and `TypedOnly`.
    ///
    /// With `eager_complete = true`, the graph should complete as soon as any
    /// inlet finishes.  Since the Unzip stage completes all outlets simultaneously,
    /// both paths must produce the same result.
    #[test]
    fn typed_erased_equivalence_merge_latest_eager_complete() {
        let graph_eager = merge_latest_eager_graph_exec();
        let input = vec![(1i32, 10i32)];

        let erased = graph_eager
            .run_with_input_mode(input.clone(), ExecutorMode::ErasedOnly)
            .unwrap();
        let typed = graph_eager
            .run_with_input_mode(input.clone(), ExecutorMode::TypedOnly)
            .unwrap();

        assert_eq!(
            typed, erased,
            "typed and erased disagree on eager-complete behavior"
        );
        assert!(
            !typed.is_empty(),
            "eager-complete graph should produce at least one snapshot"
        );
    }

    /// Completion: a single pair produces one snapshot; both paths complete cleanly.
    #[test]
    fn typed_erased_equivalence_merge_latest_completion() {
        let graph = merge_latest_graph_exec();
        let input = vec![(0u64, 1u64)];

        let erased = graph
            .run_with_input_mode(input.clone(), ExecutorMode::ErasedOnly)
            .unwrap();
        let typed = graph
            .run_with_input_mode(input.clone(), ExecutorMode::TypedOnly)
            .unwrap();

        assert_eq!(typed, erased, "typed and erased disagree on completion");
    }

    /// `Auto` selects the typed path for the Unzip → MergeLatest topology and
    /// produces the same result as `ErasedOnly`.
    #[test]
    fn auto_selects_typed_for_merge_latest_topology() {
        let graph = merge_latest_graph_exec();
        let input: Vec<(u64, u64)> = (0..20).map(|i| (i, i + 1_000)).collect();

        let auto_result = graph
            .run_with_input_mode(input.clone(), ExecutorMode::Auto)
            .unwrap();
        let erased_result = graph
            .run_with_input_mode(input.clone(), ExecutorMode::ErasedOnly)
            .unwrap();

        assert_eq!(
            auto_result, erased_result,
            "Auto and ErasedOnly disagree on merge_latest topology"
        );
    }

    // -- Phase 3b blueprint-independence tests --------------------------------

    /// Running the same MergeLatest blueprint twice (sequentially) produces
    /// independent, correct results — no shared mutable state between runs.
    ///
    /// Pins the fix that removed the `TypedPlanCache` / `MergeLatestRunnerCell`
    /// (which serialised concurrent reuse on a `Mutex` and shared one execution
    /// core across runs).  The typed path must build a fresh `MergeLatestCore`
    /// per call to `run_with_input_report_mode`.
    #[test]
    fn merge_latest_blueprint_sequential_reuse_is_independent() {
        let graph = merge_latest_graph_exec();
        let input_a: Vec<(u64, u64)> = (0..5).map(|i| (i, i + 100)).collect();
        let input_b: Vec<(u64, u64)> = (10..15).map(|i| (i, i + 200)).collect();

        // Run the same blueprint with two different inputs.
        let result_a_typed = graph
            .run_with_input_mode(input_a.clone(), ExecutorMode::TypedOnly)
            .unwrap();
        let result_b_typed = graph
            .run_with_input_mode(input_b.clone(), ExecutorMode::TypedOnly)
            .unwrap();

        // Each run should match its own erased reference.
        let result_a_erased = graph
            .run_with_input_mode(input_a, ExecutorMode::ErasedOnly)
            .unwrap();
        let result_b_erased = graph
            .run_with_input_mode(input_b, ExecutorMode::ErasedOnly)
            .unwrap();

        assert_eq!(
            result_a_typed, result_a_erased,
            "sequential run A: typed and erased disagree"
        );
        assert_eq!(
            result_b_typed, result_b_erased,
            "sequential run B: typed and erased disagree"
        );
        // The two runs must NOT share state — run B must not contain run A's values.
        assert_ne!(
            result_a_typed, result_b_typed,
            "runs A and B should differ (different inputs)"
        );
    }

    /// Running the same MergeLatest blueprint concurrently from two threads
    /// produces independent, correct results — no Mutex serialisation, no
    /// shared execution state.
    ///
    /// This is the concurrency variant of `merge_latest_blueprint_sequential_reuse_is_independent`.
    #[test]
    fn merge_latest_blueprint_concurrent_reuse_is_independent() {
        use std::sync::Arc as StdArc;

        // Wrap in Arc so both threads can share the (immutable) blueprint.
        let graph = StdArc::new(merge_latest_graph_exec());

        let input_a: Vec<(u64, u64)> = (0..50).map(|i| (i, i + 1_000)).collect();
        let input_b: Vec<(u64, u64)> = (100..150).map(|i| (i, i + 2_000)).collect();

        let graph_a = StdArc::clone(&graph);
        let graph_b = StdArc::clone(&graph);
        let ia = input_a.clone();
        let ib = input_b.clone();

        let handle_a =
            std::thread::spawn(move || graph_a.run_with_input_mode(ia, ExecutorMode::TypedOnly));
        let handle_b =
            std::thread::spawn(move || graph_b.run_with_input_mode(ib, ExecutorMode::TypedOnly));

        let result_a = handle_a.join().expect("thread A panicked").unwrap();
        let result_b = handle_b.join().expect("thread B panicked").unwrap();

        // Reference results from the erased executor.
        let ref_a = graph
            .run_with_input_mode(input_a, ExecutorMode::ErasedOnly)
            .unwrap();
        let ref_b = graph
            .run_with_input_mode(input_b, ExecutorMode::ErasedOnly)
            .unwrap();

        assert_eq!(
            result_a, ref_a,
            "concurrent run A: typed and erased disagree"
        );
        assert_eq!(
            result_b, ref_b,
            "concurrent run B: typed and erased disagree"
        );
        // Sanity: the two runs must have produced different outputs.
        assert_ne!(result_a, result_b, "concurrent runs must be independent");
    }
}

enum AsyncLinearMessage<T> {
    Item(T),
    Done,
    Failed(StreamError),
}

enum RactorBoundaryCommand {
    Run,
}

#[cfg(feature = "cluster")]
impl ractor::Message for RactorBoundaryCommand {}

#[cfg(test)]
fn run_threaded_async_linear_count<I, T>(
    input: I,
    segments: TypedLinearSegments<T>,
    config: AsyncBoundaryExecutionConfig,
) -> StreamResult<FusedTerminalReport<usize>>
where
    I: IntoIterator<Item = T> + Send,
    I::IntoIter: Send,
    T: Send + 'static,
{
    let channels = segments.segments.len() + 1;
    let mut senders = Vec::with_capacity(channels);
    let mut receivers = Vec::with_capacity(channels);
    for _ in 0..channels {
        let (sender, receiver) = mpsc::sync_channel(config.buffer_size);
        senders.push(sender);
        receivers.push(Some(receiver));
    }

    let first_sender = senders
        .first()
        .expect("at least one async segment channel")
        .clone();
    let mut final_receiver = Some(
        receivers
            .last_mut()
            .expect("at least one async segment channel")
            .take()
            .expect("final receiver is present"),
    );
    let events = AtomicUsize::new(0);
    let async_boundary_crossings = AtomicUsize::new(0);

    let result = thread::scope(|scope| {
        let source = scope.spawn(move || feed_threaded_async_linear_input(input, first_sender));
        let mut workers = Vec::with_capacity(segments.segments.len());

        for (index, steps) in segments.segments.iter().enumerate() {
            let input = receivers[index].take().expect("worker receiver is present");
            let output = senders[index + 1].clone();
            let has_boundary_after = index + 1 < segments.segments.len();
            let events = &events;
            let async_boundary_crossings = &async_boundary_crossings;
            workers.push(scope.spawn(move || {
                run_threaded_async_linear_segment(
                    steps,
                    input,
                    output,
                    has_boundary_after,
                    events,
                    async_boundary_crossings,
                    config,
                )
            }));
        }
        drop(senders);

        let final_rx = final_receiver.take().expect("final receiver present");
        let mut count = 0;
        let mut terminal_error = None;
        loop {
            match final_rx.recv() {
                Ok(AsyncLinearMessage::Item(_)) => count += 1,
                Ok(AsyncLinearMessage::Done) => break,
                Ok(AsyncLinearMessage::Failed(error)) => {
                    terminal_error = Some(error);
                    break;
                }
                Err(_) => {
                    terminal_error = Some(StreamError::AbruptTermination);
                    break;
                }
            }
        }
        drop(final_rx);

        let mut worker_error = join_threaded_async_linear_worker(source)?;
        for worker in workers {
            if worker_error.is_none() {
                worker_error = join_threaded_async_linear_worker(worker)?;
            } else {
                let _ = join_threaded_async_linear_worker(worker);
            }
        }

        match (terminal_error, worker_error) {
            (Some(error), _) if error != StreamError::AbruptTermination => return Err(error),
            (_, Some(error)) => return Err(error),
            (Some(error), None) => return Err(error),
            (None, None) => {}
        }

        Ok(count)
    });

    Ok(FusedTerminalReport {
        result: result?,
        events: events.load(Ordering::Relaxed),
        async_boundary_crossings: async_boundary_crossings.load(Ordering::Relaxed),
    })
}

fn run_ractor_async_linear_count<I, T>(
    input: I,
    segments: TypedLinearSegments<T>,
    config: AsyncBoundaryExecutionConfig,
) -> StreamResult<FusedTerminalReport<usize>>
where
    I: IntoIterator<Item = T> + Send,
    I::IntoIter: Send + 'static,
    T: Send + 'static,
{
    if config.buffer_size == 0 {
        return Err(StreamError::GraphValidation(
            "ractor async boundary execution requires buffer_size greater than zero".into(),
        ));
    }

    let input = input.into_iter();
    let runtime = ractor_boundary_runtime()?;
    if tokio::runtime::Handle::try_current().is_ok() {
        thread::scope(|scope| {
            let handle = scope.spawn(move || {
                runtime.block_on(run_ractor_async_linear_count_on_runtime(
                    input, segments, config,
                ))
            });
            handle.join().map_err(|_| {
                StreamError::Failed("ractor async boundary runtime thread panicked".into())
            })?
        })
    } else {
        runtime.block_on(run_ractor_async_linear_count_on_runtime(
            input, segments, config,
        ))
    }
}

fn ractor_boundary_runtime() -> StreamResult<&'static tokio::runtime::Runtime> {
    static RUNTIME: OnceLock<Result<tokio::runtime::Runtime, String>> = OnceLock::new();

    match RUNTIME.get_or_init(|| {
        tokio::runtime::Builder::new_multi_thread()
            .build()
            .map_err(|error| format!("ractor async boundary runtime failed to start: {error}"))
    }) {
        Ok(runtime) => Ok(runtime),
        Err(error) => Err(StreamError::Failed(error.clone())),
    }
}

async fn run_ractor_async_linear_count_on_runtime<I, T>(
    input: I,
    segments: TypedLinearSegments<T>,
    config: AsyncBoundaryExecutionConfig,
) -> StreamResult<FusedTerminalReport<usize>>
where
    I: Iterator<Item = T> + Send + 'static,
    T: Send + 'static,
{
    let channels = segments.segments.len() + 1;
    let mut senders = Vec::with_capacity(channels);
    let mut receivers = Vec::with_capacity(channels);
    for _ in 0..channels {
        let (sender, receiver) = ractor::concurrency::mpsc_bounded(config.buffer_size);
        senders.push(sender);
        receivers.push(Some(receiver));
    }

    let first_sender = senders
        .first()
        .expect("at least one async segment channel")
        .clone();
    let mut final_receiver = receivers
        .last_mut()
        .expect("at least one async segment channel")
        .take()
        .expect("final receiver is present");
    let events = Arc::new(AtomicUsize::new(0));
    let async_boundary_crossings = Arc::new(AtomicUsize::new(0));

    let (source_ref, source_handle) = Actor::spawn(
        None,
        RactorLinearSourceActor::<I, T>::new(),
        RactorLinearSourceState {
            input: Some(input),
            output: first_sender,
        },
    )
    .await
    .map_err(ractor_spawn_error)?;

    let mut actors = Vec::with_capacity(segments.segments.len() + 1);
    actors.push((source_ref, source_handle));

    for (index, steps) in segments.segments.into_iter().enumerate() {
        let input = receivers[index].take().expect("worker receiver is present");
        let output = senders[index + 1].clone();
        let has_boundary_after = index + 1 < channels - 1;
        let (worker_ref, worker_handle) = match Actor::spawn(
            None,
            RactorLinearSegmentActor::<T>::new(),
            RactorLinearSegmentState {
                steps,
                input,
                output,
                has_boundary_after,
                events: Arc::clone(&events),
                async_boundary_crossings: Arc::clone(&async_boundary_crossings),
                config,
            },
        )
        .await
        {
            Ok(actor) => actor,
            Err(error) => {
                let error = ractor_spawn_error(error);
                stop_ractor_async_linear_actors(&actors);
                let _ = join_ractor_async_linear_actors(actors).await;
                return Err(error);
            }
        };
        actors.push((worker_ref, worker_handle));
    }
    drop(senders);

    let mut start_error = None;
    for (actor, _) in &actors {
        if actor.send_message(RactorBoundaryCommand::Run).is_err() {
            start_error = Some(StreamError::AbruptTermination);
            break;
        }
    }
    if let Some(error) = start_error {
        stop_ractor_async_linear_actors(&actors);
        let _ = join_ractor_async_linear_actors(actors).await;
        return Err(error);
    }

    let mut count = 0;
    let mut terminal_error = None;
    loop {
        match final_receiver.recv().await {
            Some(AsyncLinearMessage::Item(_)) => count += 1,
            Some(AsyncLinearMessage::Done) => break,
            Some(AsyncLinearMessage::Failed(error)) => {
                terminal_error = Some(error);
                break;
            }
            None => {
                terminal_error = Some(StreamError::AbruptTermination);
                break;
            }
        }
    }
    drop(final_receiver);

    stop_ractor_async_linear_actors(&actors);
    let actor_error = join_ractor_async_linear_actors(actors).await;

    match (terminal_error, actor_error) {
        (Some(error), _) if error != StreamError::AbruptTermination => return Err(error),
        (_, Some(error)) => return Err(error),
        (Some(error), None) => return Err(error),
        (None, None) => {}
    }

    Ok(FusedTerminalReport {
        result: count,
        events: events.load(Ordering::Relaxed),
        async_boundary_crossings: async_boundary_crossings.load(Ordering::Relaxed),
    })
}

struct RactorLinearSourceActor<I, T> {
    _marker: PhantomData<fn() -> (I, T)>,
}

impl<I, T> RactorLinearSourceActor<I, T> {
    fn new() -> Self {
        Self {
            _marker: PhantomData,
        }
    }
}

struct RactorLinearSourceState<I, T> {
    input: Option<I>,
    output: ractor::concurrency::MpscSender<AsyncLinearMessage<T>>,
}

impl<I, T> Actor for RactorLinearSourceActor<I, T>
where
    I: Iterator<Item = T> + Send + 'static,
    T: Send + 'static,
{
    type Msg = RactorBoundaryCommand;
    type State = RactorLinearSourceState<I, T>;
    type Arguments = RactorLinearSourceState<I, T>;

    async fn pre_start(
        &self,
        _myself: ActorRef<Self::Msg>,
        args: Self::Arguments,
    ) -> Result<Self::State, ActorProcessingErr> {
        Ok(args)
    }

    async fn handle(
        &self,
        myself: ActorRef<Self::Msg>,
        message: Self::Msg,
        state: &mut Self::State,
    ) -> Result<(), ActorProcessingErr> {
        match message {
            RactorBoundaryCommand::Run => {
                let input = state.input.take().ok_or_else(|| {
                    actor_processing_error(StreamError::GraphValidation(
                        "ractor async boundary source actor was run more than once".into(),
                    ))
                })?;
                feed_ractor_async_linear_input(input, &state.output)
                    .await
                    .map_err(actor_processing_error)?;
                myself.stop(None);
            }
        }
        Ok(())
    }
}

struct RactorLinearSegmentActor<T> {
    _marker: PhantomData<fn() -> T>,
}

impl<T> RactorLinearSegmentActor<T> {
    fn new() -> Self {
        Self {
            _marker: PhantomData,
        }
    }
}

struct RactorLinearSegmentState<T> {
    steps: Vec<TypedLinearStep<T>>,
    input: ractor::concurrency::MpscReceiver<AsyncLinearMessage<T>>,
    output: ractor::concurrency::MpscSender<AsyncLinearMessage<T>>,
    has_boundary_after: bool,
    events: Arc<AtomicUsize>,
    async_boundary_crossings: Arc<AtomicUsize>,
    config: AsyncBoundaryExecutionConfig,
}

impl<T> Actor for RactorLinearSegmentActor<T>
where
    T: Send + 'static,
{
    type Msg = RactorBoundaryCommand;
    type State = RactorLinearSegmentState<T>;
    type Arguments = RactorLinearSegmentState<T>;

    async fn pre_start(
        &self,
        _myself: ActorRef<Self::Msg>,
        args: Self::Arguments,
    ) -> Result<Self::State, ActorProcessingErr> {
        Ok(args)
    }

    async fn handle(
        &self,
        myself: ActorRef<Self::Msg>,
        message: Self::Msg,
        state: &mut Self::State,
    ) -> Result<(), ActorProcessingErr> {
        match message {
            RactorBoundaryCommand::Run => {
                run_ractor_async_linear_segment(state)
                    .await
                    .map_err(actor_processing_error)?;
                myself.stop(None);
            }
        }
        Ok(())
    }
}

async fn feed_ractor_async_linear_input<I, T>(
    input: I,
    output: &ractor::concurrency::MpscSender<AsyncLinearMessage<T>>,
) -> StreamResult<()>
where
    I: Iterator<Item = T>,
{
    for item in input {
        output
            .send(AsyncLinearMessage::Item(item))
            .await
            .map_err(|_| StreamError::AbruptTermination)?;
    }
    output
        .send(AsyncLinearMessage::Done)
        .await
        .map_err(|_| StreamError::AbruptTermination)
}

async fn run_ractor_async_linear_segment<T>(
    state: &mut RactorLinearSegmentState<T>,
) -> StreamResult<()>
where
    T: Send + 'static,
{
    loop {
        match state.input.recv().await {
            Some(AsyncLinearMessage::Item(item)) => {
                let result =
                    run_async_linear_item(item, &state.steps, &state.events, state.config.fused)
                        .and_then(|item| {
                            if state.has_boundary_after {
                                bump_fused_event_atomic(&state.events, state.config.fused)?;
                                state
                                    .async_boundary_crossings
                                    .fetch_add(1, Ordering::Relaxed);
                                bump_fused_event_atomic(&state.events, state.config.fused)?;
                            }
                            Ok(item)
                        });

                match result {
                    Ok(item) => state
                        .output
                        .send(AsyncLinearMessage::Item(item))
                        .await
                        .map_err(|_| StreamError::AbruptTermination)?,
                    Err(error) => {
                        let _ = state
                            .output
                            .send(AsyncLinearMessage::Failed(error.clone()))
                            .await;
                        return Err(error);
                    }
                }
            }
            Some(AsyncLinearMessage::Done) => {
                state
                    .output
                    .send(AsyncLinearMessage::Done)
                    .await
                    .map_err(|_| StreamError::AbruptTermination)?;
                return Ok(());
            }
            Some(AsyncLinearMessage::Failed(error)) => {
                let _ = state
                    .output
                    .send(AsyncLinearMessage::Failed(error.clone()))
                    .await;
                return Err(error);
            }
            None => return Err(StreamError::AbruptTermination),
        }
    }
}

async fn join_ractor_async_linear_actor(
    handle: ractor::concurrency::JoinHandle<()>,
) -> StreamResult<()> {
    handle.await.map_err(|error| {
        StreamError::Failed(format!("ractor async boundary actor task failed: {error}"))
    })
}

fn stop_ractor_async_linear_actors(
    actors: &[(
        ActorRef<RactorBoundaryCommand>,
        ractor::concurrency::JoinHandle<()>,
    )],
) {
    for (actor, _) in actors {
        actor.stop(None);
    }
}

#[cfg_attr(not(test), allow(dead_code))]
async fn join_ractor_async_linear_actors(
    actors: Vec<(
        ActorRef<RactorBoundaryCommand>,
        ractor::concurrency::JoinHandle<()>,
    )>,
) -> Option<StreamError> {
    let mut actor_error = None;
    for (_, handle) in actors {
        let result = join_ractor_async_linear_actor(handle).await;
        if actor_error.is_some() {
            continue;
        }
        if let Err(error) = result {
            actor_error = Some(error);
        }
    }
    actor_error
}

fn ractor_spawn_error(error: ractor::SpawnErr) -> StreamError {
    StreamError::Failed(format!(
        "ractor async boundary actor failed to spawn: {error}"
    ))
}

fn actor_processing_error(error: StreamError) -> ActorProcessingErr {
    Box::new(error)
}

#[cfg(test)]
fn feed_threaded_async_linear_input<I, T>(
    input: I,
    output: mpsc::SyncSender<AsyncLinearMessage<T>>,
) -> StreamResult<()>
where
    I: IntoIterator<Item = T>,
{
    for item in input {
        output
            .send(AsyncLinearMessage::Item(item))
            .map_err(|_| StreamError::AbruptTermination)?;
    }
    output
        .send(AsyncLinearMessage::Done)
        .map_err(|_| StreamError::AbruptTermination)
}

#[cfg(test)]
fn run_threaded_async_linear_segment<T>(
    steps: &[TypedLinearStep<T>],
    input: mpsc::Receiver<AsyncLinearMessage<T>>,
    output: mpsc::SyncSender<AsyncLinearMessage<T>>,
    has_boundary_after: bool,
    events: &AtomicUsize,
    async_boundary_crossings: &AtomicUsize,
    config: AsyncBoundaryExecutionConfig,
) -> StreamResult<()>
where
    T: Send + 'static,
{
    loop {
        match input.recv().map_err(|_| StreamError::AbruptTermination)? {
            AsyncLinearMessage::Item(item) => {
                let result =
                    run_async_linear_item(item, steps, events, config.fused).and_then(|item| {
                        if has_boundary_after {
                            bump_fused_event_atomic(events, config.fused)?;
                            async_boundary_crossings.fetch_add(1, Ordering::Relaxed);
                            bump_fused_event_atomic(events, config.fused)?;
                        }
                        output
                            .send(AsyncLinearMessage::Item(item))
                            .map_err(|_| StreamError::AbruptTermination)
                    });
                if let Err(error) = result {
                    let _ = output.send(AsyncLinearMessage::Failed(error.clone()));
                    return Err(error);
                }
            }
            AsyncLinearMessage::Done => {
                output
                    .send(AsyncLinearMessage::Done)
                    .map_err(|_| StreamError::AbruptTermination)?;
                return Ok(());
            }
            AsyncLinearMessage::Failed(error) => {
                let _ = output.send(AsyncLinearMessage::Failed(error.clone()));
                return Err(error);
            }
        }
    }
}

fn run_async_linear_item<T>(
    mut item: T,
    steps: &[TypedLinearStep<T>],
    events: &AtomicUsize,
    config: FusedExecutionConfig,
) -> StreamResult<T>
where
    T: Send + 'static,
{
    for step in steps {
        bump_fused_event_atomic(events, config)?;
        match step {
            TypedLinearStep::Pass => {}
            TypedLinearStep::Map(mapper) => {
                item = mapper(item);
            }
            TypedLinearStep::AsyncBoundary => {
                return Err(StreamError::GraphValidation(
                    "async boundary execution expects pre-split linear segments".into(),
                ));
            }
        }
        bump_fused_event_atomic(events, config)?;
    }
    Ok(item)
}

#[cfg(test)]
fn join_threaded_async_linear_worker(
    handle: thread::ScopedJoinHandle<'_, StreamResult<()>>,
) -> StreamResult<Option<StreamError>> {
    match handle.join() {
        Ok(Ok(())) => Ok(None),
        Ok(Err(error)) => Ok(Some(error)),
        Err(_) => Err(StreamError::Failed("async boundary worker panicked".into())),
    }
}

fn bump_fused_event_atomic(events: &AtomicUsize, config: FusedExecutionConfig) -> StreamResult<()> {
    let events = events.fetch_add(1, Ordering::Relaxed) + 1;
    if events > config.event_limit {
        return Err(StreamError::EventLimitExceeded {
            limit: config.event_limit,
        });
    }
    Ok(())
}

impl<T> GraphBlueprint<FanInShape<T, T>>
where
    T: Clone + Send + 'static,
{
    pub fn run_fan_in(&self, inputs: Vec<Vec<T>>) -> StreamResult<Vec<T>> {
        Ok(self
            .run_fan_in_report(inputs, FusedExecutionConfig::default())?
            .output)
    }

    pub fn run_fan_in_report(
        &self,
        inputs: Vec<Vec<T>>,
        config: FusedExecutionConfig,
    ) -> StreamResult<FusedExecutionReport<T>> {
        if inputs.len() != self.shape.inlet_count() {
            return Err(StreamError::GraphValidation(format!(
                "expected {} input streams, got {}",
                self.shape.inlet_count(),
                inputs.len()
            )));
        }

        let output_capacity = inputs.iter().map(Vec::len).sum();
        let mut queues: Vec<_> = inputs.into_iter().map(Vec::into_iter).collect();
        let schedule = self.fan_in_schedule();
        let mut schedule_index = 0;
        let outlet = self.shape.outlet().id();
        let mut executor = FusedExecutor::new(self, config);
        let mut output = Vec::with_capacity(output_capacity);
        let mut completed = vec![false; queues.len()];

        {
            let mut output_sink = VecOutputSink {
                output: &mut output,
            };
            for (index, queue) in queues.iter().enumerate() {
                if queue.len() == 0 {
                    executor.complete(self.shape.inlet(index)?.id(), outlet, &mut output_sink)?;
                    completed[index] = true;
                }
            }
            while queues.iter().any(|queue| queue.len() > 0) {
                let input_index = next_scheduled_input(&queues, &schedule, &mut schedule_index)
                    .ok_or_else(|| {
                        StreamError::GraphValidation("no runnable fan-in input".into())
                    })?;
                let item = queues[input_index]
                    .next()
                    .expect("scheduled input had an item");
                executor.deliver(
                    self.shape.inlet(input_index)?.id(),
                    datum(item),
                    outlet,
                    &mut output_sink,
                )?;
                if queues[input_index].len() == 0 && !completed[input_index] {
                    executor.complete(
                        self.shape.inlet(input_index)?.id(),
                        outlet,
                        &mut output_sink,
                    )?;
                    completed[input_index] = true;
                }
            }
        }

        Ok(FusedExecutionReport {
            output,
            events: executor.events,
            async_boundary_crossings: executor.async_boundary_crossings,
        })
    }

    fn fan_in_schedule(&self) -> Vec<usize> {
        self.stages
            .iter()
            .find_map(|stage| match &stage.spec.kind {
                StageKind::MergePrioritized { weights }
                    if weights.len() == self.shape.inlet_count()
                        && stage.spec.outlets.len() == 1
                        && stage.spec.outlets[0].id() == self.shape.outlet().id()
                        && stage.spec.inlets.iter().map(AnyInlet::id).eq(self
                            .shape
                            .inlets()
                            .iter()
                            .map(|inlet| inlet.id())) =>
                {
                    Some(
                        weights
                            .iter()
                            .enumerate()
                            .flat_map(|(index, weight)| std::iter::repeat_n(index, *weight))
                            .collect(),
                    )
                }
                _ => None,
            })
            .unwrap_or_else(|| (0..self.shape.inlet_count()).collect())
    }

    pub fn run_concat(&self, inputs: Vec<Vec<T>>) -> StreamResult<Vec<T>> {
        Ok(self
            .run_concat_report(inputs, FusedExecutionConfig::default())?
            .output)
    }

    pub fn run_concat_report(
        &self,
        inputs: Vec<Vec<T>>,
        config: FusedExecutionConfig,
    ) -> StreamResult<FusedExecutionReport<T>> {
        if inputs.len() != self.shape.inlet_count() {
            return Err(StreamError::GraphValidation(format!(
                "expected {} input streams, got {}",
                self.shape.inlet_count(),
                inputs.len()
            )));
        }

        let output_capacity = inputs.iter().map(Vec::len).sum();
        let mut queues: Vec<_> = inputs.into_iter().map(Vec::into_iter).collect();
        let outlet = self.shape.outlet().id();
        let mut executor = FusedExecutor::new(self, config);
        let mut output = Vec::with_capacity(output_capacity);

        {
            let mut output_sink = VecOutputSink {
                output: &mut output,
            };
            for (index, queue) in queues.iter_mut().enumerate() {
                for item in queue.by_ref() {
                    executor.deliver(
                        self.shape.inlet(index)?.id(),
                        datum(item),
                        outlet,
                        &mut output_sink,
                    )?;
                }
                executor.complete(self.shape.inlet(index)?.id(), outlet, &mut output_sink)?;
            }
        }

        Ok(FusedExecutionReport {
            output,
            events: executor.events,
            async_boundary_crossings: executor.async_boundary_crossings,
        })
    }

    pub fn run_or_else(&self, primary: Vec<T>, secondary: Vec<T>) -> StreamResult<Vec<T>> {
        Ok(self
            .run_or_else_report(primary, secondary, FusedExecutionConfig::default())?
            .output)
    }

    pub fn run_or_else_report(
        &self,
        primary: Vec<T>,
        secondary: Vec<T>,
        config: FusedExecutionConfig,
    ) -> StreamResult<FusedExecutionReport<T>> {
        if self.shape.inlet_count() != 2 {
            return Err(StreamError::GraphValidation(format!(
                "or-else helper expected 2 inlets, got {}",
                self.shape.inlet_count()
            )));
        }

        let primary = primary.into_iter();
        let secondary = secondary.into_iter();
        let primary_inlet = self.shape.inlet(0)?.id();
        let secondary_inlet = self.shape.inlet(1)?.id();
        let outlet = self.shape.outlet().id();
        let mut executor = FusedExecutor::new(self, config);
        let mut output = Vec::new();
        let mut primary_emitted = false;

        {
            let mut output_sink = VecOutputSink {
                output: &mut output,
            };
            for item in primary {
                primary_emitted = true;
                executor.deliver(primary_inlet, datum(item), outlet, &mut output_sink)?;
            }
            executor.complete(primary_inlet, outlet, &mut output_sink)?;

            if !primary_emitted {
                for item in secondary {
                    executor.deliver(secondary_inlet, datum(item), outlet, &mut output_sink)?;
                }
            }
            executor.complete(secondary_inlet, outlet, &mut output_sink)?;
        }

        Ok(FusedExecutionReport {
            output,
            events: executor.events,
            async_boundary_crossings: executor.async_boundary_crossings,
        })
    }

    pub fn run_or_else_secondary_first(
        &self,
        primary: Vec<T>,
        secondary: Vec<T>,
    ) -> StreamResult<Vec<T>> {
        Ok(self
            .run_or_else_secondary_first_report(
                primary,
                secondary,
                FusedExecutionConfig::default(),
            )?
            .output)
    }

    pub fn run_or_else_secondary_first_report(
        &self,
        primary: Vec<T>,
        secondary: Vec<T>,
        config: FusedExecutionConfig,
    ) -> StreamResult<FusedExecutionReport<T>> {
        if self.shape.inlet_count() != 2 {
            return Err(StreamError::GraphValidation(format!(
                "or-else helper expected 2 inlets, got {}",
                self.shape.inlet_count()
            )));
        }

        let primary_inlet = self.shape.inlet(0)?.id();
        let secondary_inlet = self.shape.inlet(1)?.id();
        let outlet = self.shape.outlet().id();
        let mut executor = FusedExecutor::new(self, config);
        let mut output = Vec::new();

        {
            let mut output_sink = VecOutputSink {
                output: &mut output,
            };
            for item in secondary {
                executor.deliver(secondary_inlet, datum(item), outlet, &mut output_sink)?;
            }
            for item in primary {
                executor.deliver(primary_inlet, datum(item), outlet, &mut output_sink)?;
            }
            executor.complete(primary_inlet, outlet, &mut output_sink)?;
            executor.complete(secondary_inlet, outlet, &mut output_sink)?;
        }

        Ok(FusedExecutionReport {
            output,
            events: executor.events,
            async_boundary_crossings: executor.async_boundary_crossings,
        })
    }

    pub fn run_or_else_secondary_closed_first(&self, secondary: Vec<T>) -> StreamResult<Vec<T>> {
        if self.shape.inlet_count() != 2 {
            return Err(StreamError::GraphValidation(format!(
                "or-else helper expected 2 inlets, got {}",
                self.shape.inlet_count()
            )));
        }

        let primary_inlet = self.shape.inlet(0)?.id();
        let secondary_inlet = self.shape.inlet(1)?.id();
        let outlet = self.shape.outlet().id();
        let mut executor = FusedExecutor::new(self, FusedExecutionConfig::default());
        let mut output = Vec::new();

        {
            let mut output_sink = VecOutputSink {
                output: &mut output,
            };
            for item in secondary {
                executor.deliver(secondary_inlet, datum(item), outlet, &mut output_sink)?;
            }
            executor.complete(secondary_inlet, outlet, &mut output_sink)?;
            executor.complete(primary_inlet, outlet, &mut output_sink)?;
        }

        Ok(output)
    }

    pub fn run_interleave(
        &self,
        inputs: Vec<Vec<T>>,
        segment_size: usize,
        eager_close: bool,
    ) -> StreamResult<Vec<T>> {
        Ok(self
            .run_interleave_report(
                inputs,
                segment_size,
                eager_close,
                FusedExecutionConfig::default(),
            )?
            .output)
    }

    pub fn run_interleave_report(
        &self,
        inputs: Vec<Vec<T>>,
        segment_size: usize,
        eager_close: bool,
        config: FusedExecutionConfig,
    ) -> StreamResult<FusedExecutionReport<T>> {
        if inputs.len() != self.shape.inlet_count() {
            return Err(StreamError::GraphValidation(format!(
                "expected {} input streams, got {}",
                self.shape.inlet_count(),
                inputs.len()
            )));
        }
        if segment_size == 0 {
            return Err(StreamError::GraphValidation(
                "interleave segment size must be greater than zero".into(),
            ));
        }

        let output_capacity = inputs.iter().map(Vec::len).sum();
        let mut queues: Vec<_> = inputs.into_iter().map(Vec::into_iter).collect();
        let mut completed = vec![false; queues.len()];
        let outlet = self.shape.outlet().id();
        let mut executor = FusedExecutor::new(self, config);
        let mut output = Vec::with_capacity(output_capacity);

        {
            let mut output_sink = VecOutputSink {
                output: &mut output,
            };
            for (index, queue) in queues.iter().enumerate() {
                if queue.len() == 0 {
                    executor.complete(self.shape.inlet(index)?.id(), outlet, &mut output_sink)?;
                    completed[index] = true;
                    if eager_close {
                        return Ok(FusedExecutionReport {
                            output,
                            events: executor.events,
                            async_boundary_crossings: executor.async_boundary_crossings,
                        });
                    }
                }
            }

            let mut current = 0usize;
            while completed.iter().any(|done| !done) {
                if completed[current] {
                    current = next_open_index(&completed, current).ok_or_else(|| {
                        StreamError::GraphValidation("no open interleave input".into())
                    })?;
                    continue;
                }

                let mut emitted = 0usize;
                while emitted < segment_size {
                    match queues[current].next() {
                        Some(item) => {
                            executor.deliver(
                                self.shape.inlet(current)?.id(),
                                datum(item),
                                outlet,
                                &mut output_sink,
                            )?;
                            emitted += 1;
                        }
                        None => {
                            executor.complete(
                                self.shape.inlet(current)?.id(),
                                outlet,
                                &mut output_sink,
                            )?;
                            completed[current] = true;
                            if eager_close {
                                return Ok(FusedExecutionReport {
                                    output,
                                    events: executor.events,
                                    async_boundary_crossings: executor.async_boundary_crossings,
                                });
                            }
                            break;
                        }
                    }
                }

                if completed.iter().all(|done| *done) {
                    break;
                }
                current = next_open_index(&completed, current).ok_or_else(|| {
                    StreamError::GraphValidation("no open interleave input".into())
                })?;
            }
        }

        Ok(FusedExecutionReport {
            output,
            events: executor.events,
            async_boundary_crossings: executor.async_boundary_crossings,
        })
    }
}

impl<T> GraphBlueprint<MergePreferredShape<T>>
where
    T: Clone + Send + 'static,
{
    pub fn run_merge_preferred(
        &self,
        preferred: Vec<T>,
        secondary_inputs: Vec<Vec<T>>,
    ) -> StreamResult<Vec<T>> {
        Ok(self
            .run_merge_preferred_report(
                preferred,
                secondary_inputs,
                FusedExecutionConfig::default(),
            )?
            .output)
    }

    pub fn run_merge_preferred_report(
        &self,
        preferred: Vec<T>,
        secondary_inputs: Vec<Vec<T>>,
        config: FusedExecutionConfig,
    ) -> StreamResult<FusedExecutionReport<T>> {
        if secondary_inputs.len() != self.shape.secondary_count() {
            return Err(StreamError::GraphValidation(format!(
                "expected {} secondary input streams, got {}",
                self.shape.secondary_count(),
                secondary_inputs.len()
            )));
        }

        let output_capacity =
            preferred.len() + secondary_inputs.iter().map(Vec::len).sum::<usize>();
        let mut preferred = preferred.into_iter();
        let mut secondary: Vec<_> = secondary_inputs.into_iter().map(Vec::into_iter).collect();
        let secondary_schedule = (0..secondary.len()).collect::<Vec<_>>();
        let mut secondary_index = 0;
        let outlet = self.shape.outlet().id();
        let preferred_inlet = self.shape.preferred().id();
        let mut executor = FusedExecutor::new(self, config);
        let mut output = Vec::with_capacity(output_capacity);
        let mut preferred_completed = false;
        let mut secondary_completed = vec![false; secondary.len()];

        {
            let mut output_sink = VecOutputSink {
                output: &mut output,
            };
            if preferred.len() == 0 {
                executor.complete(preferred_inlet, outlet, &mut output_sink)?;
                preferred_completed = true;
            }
            for (index, queue) in secondary.iter().enumerate() {
                if queue.len() == 0 {
                    executor.complete(
                        self.shape.secondary(index)?.id(),
                        outlet,
                        &mut output_sink,
                    )?;
                    secondary_completed[index] = true;
                }
            }
            while preferred.len() > 0 || secondary.iter().any(|queue| queue.len() > 0) {
                if let Some(item) = preferred.next() {
                    executor.deliver(preferred_inlet, datum(item), outlet, &mut output_sink)?;
                    if preferred.len() == 0 && !preferred_completed {
                        executor.complete(preferred_inlet, outlet, &mut output_sink)?;
                        preferred_completed = true;
                    }
                    continue;
                }

                let input_index =
                    next_scheduled_input(&secondary, &secondary_schedule, &mut secondary_index)
                        .ok_or_else(|| {
                            StreamError::GraphValidation("no runnable secondary input".into())
                        })?;
                let item = secondary[input_index]
                    .next()
                    .expect("scheduled secondary input had an item");
                executor.deliver(
                    self.shape.secondary(input_index)?.id(),
                    datum(item),
                    outlet,
                    &mut output_sink,
                )?;
                if secondary[input_index].len() == 0 && !secondary_completed[input_index] {
                    executor.complete(
                        self.shape.secondary(input_index)?.id(),
                        outlet,
                        &mut output_sink,
                    )?;
                    secondary_completed[input_index] = true;
                }
            }
        }

        Ok(FusedExecutionReport {
            output,
            events: executor.events,
            async_boundary_crossings: executor.async_boundary_crossings,
        })
    }
}

fn next_scheduled_input<I>(
    queues: &[I],
    schedule: &[usize],
    schedule_index: &mut usize,
) -> Option<usize>
where
    I: ExactSizeIterator,
{
    if schedule.is_empty() {
        return queues.iter().position(|queue| queue.len() > 0);
    }

    for _ in 0..schedule.len() {
        let index = schedule[*schedule_index % schedule.len()];
        *schedule_index += 1;
        if queues.get(index).is_some_and(|queue| queue.len() > 0) {
            return Some(index);
        }
    }

    queues.iter().position(|queue| queue.len() > 0)
}

fn next_open_index(completed: &[bool], current: usize) -> Option<usize> {
    if completed.is_empty() {
        return None;
    }
    let len = completed.len();
    for offset in 1..=len {
        let index = (current + offset) % len;
        if !completed[index] {
            return Some(index);
        }
    }
    None
}

// ---------------------------------------------------------------------------
// Output-sink plumbing — pub(crate) so Phase 1+ typed executors can reuse
// ---------------------------------------------------------------------------

/// Typed output receiver used by the fused executor.
///
/// `pub(crate)` so the upcoming typed-port executor (Phase 1+) can reuse the
/// same sink abstractions without duplicating the erased-path infrastructure.
pub(crate) trait FusedOutputSink<Out> {
    fn emit(&mut self, value: Out) -> StreamResult<()>;
}

pub(crate) struct VecOutputSink<'a, Out> {
    pub(crate) output: &'a mut Vec<Out>,
}

impl<Out> FusedOutputSink<Out> for VecOutputSink<'_, Out> {
    fn emit(&mut self, value: Out) -> StreamResult<()> {
        self.output.push(value);
        Ok(())
    }
}

pub(crate) struct CountOutputSink {
    pub(crate) count: usize,
}

impl<Out> FusedOutputSink<Out> for CountOutputSink {
    fn emit(&mut self, _value: Out) -> StreamResult<()> {
        self.count += 1;
        Ok(())
    }
}

pub(crate) struct FoldOutputSink<Acc, F> {
    pub(crate) accumulator: Option<Acc>,
    pub(crate) fold: F,
}

impl<Out, Acc, F> FusedOutputSink<Out> for FoldOutputSink<Acc, F>
where
    F: FnMut(Acc, Out) -> Acc,
{
    fn emit(&mut self, value: Out) -> StreamResult<()> {
        let accumulator = self
            .accumulator
            .take()
            .expect("fold accumulator is present while executor is running");
        self.accumulator = Some((self.fold)(accumulator, value));
        Ok(())
    }
}

impl<Acc, F> FoldOutputSink<Acc, F> {
    pub(crate) fn finish(mut self) -> Acc {
        self.accumulator
            .take()
            .expect("fold accumulator is present after executor finishes")
    }
}

// ---------------------------------------------------------------------------
// Graph-index scaffolding — pub(crate) for Phase 1+ typed executor reuse
// ---------------------------------------------------------------------------

/// Fused graph executor over erased (`Box<dyn DatumElement>`) values.
///
/// The struct itself and its edge/stage-lookup maps are `pub(crate)` so the
/// typed-port executor introduced in Phase 1+ can reuse the graph-index
/// infrastructure (edge maps, stage-index maps) without duplicating it.
/// The per-element hot path remains encapsulated — only the structural pieces
/// are exposed.
#[derive(Debug)]
pub(crate) struct FusedExecutor<'a, S: Shape> {
    graph: &'a GraphBlueprint<S>,
    /// Outlet → connected inlet (internal graph edge lookup).
    pub(crate) edge_by_outlet: HashMap<PortId, PortId>,
    /// Inlet → connected outlet (internal graph edge lookup).
    pub(crate) edge_by_inlet: HashMap<PortId, PortId>,
    /// Inlet port → owning stage index.
    pub(crate) stage_by_inlet: HashMap<PortId, usize>,
    /// Outlet port → owning stage index.
    pub(crate) stage_by_outlet: HashMap<PortId, usize>,
    stage_states: Vec<StageState>,
    pub(crate) opaque_logics: Vec<Option<GraphStageLogic>>,
    config: FusedExecutionConfig,
    pub(crate) events: usize,
    pub(crate) async_boundary_crossings: usize,
    cancelled_outlets: HashSet<PortId>,
}

#[derive(Debug)]
enum StageState {
    Stateless,
    Broadcast {
        fast_path: Option<BroadcastZipFastPath>,
        cancelled_outlets: Vec<bool>,
        live_outlets: usize,
    },
    Balance {
        fast_path: Option<BalanceMergeFastPath>,
        next: usize,
        cancelled_outlets: Vec<bool>,
        live_outlets: usize,
    },
    Merge {
        open_inputs: usize,
        eager_complete: bool,
        completed: bool,
    },
    OrElse {
        primary_emitted: bool,
        buffer: VecDeque<DatumValue>,
        primary_closed: bool,
        secondary_closed: bool,
        completed: bool,
    },
    Zip {
        left_inlet: PortId,
        right_inlet: PortId,
        left: VecDeque<DatumValue>,
        right: VecDeque<DatumValue>,
        left_pending_complete: bool,
        right_pending_complete: bool,
        completed: bool,
    },
    Unzip {
        fast_path: Option<UnzipFanInFastPath>,
        zip_fast_path: Option<UnzipZipFastPath>,
        demand: [bool; 2],
        cancelled: [bool; 2],
        upstream_closed: bool,
    },
    MergeSorted {
        left: VecDeque<DatumValue>,
        right: VecDeque<DatumValue>,
        left_closed: bool,
        right_closed: bool,
        pending: VecDeque<DatumValue>,
        completed: bool,
    },
    MergeSequence {
        next_sequence: u64,
        pending: Vec<(u64, DatumValue)>,
        completed_count: usize,
        output_buffer: VecDeque<DatumValue>,
        completed: bool,
    },
    MergeLatest {
        latest: Vec<Option<DatumValue>>,
        seen_count: usize,
        completed_count: usize,
        pending: VecDeque<DatumValue>,
        completed: bool,
    },
    Partition {
        fast_path: Option<PartitionMergeFastPath>,
        pending: Option<(usize, DatumValue)>,
        upstream_closed: bool,
        demand: Vec<bool>,
        cancelled: Vec<bool>,
        output_count: usize,
        completed: bool,
    },
}

#[derive(Clone, Copy, Debug)]
struct BroadcastZipFastPath {
    zip_stage_index: usize,
    left_uses_clone: bool,
}

#[derive(Clone, Copy, Debug)]
struct BalanceMergeFastPath {
    merge_stage_index: usize,
}

#[derive(Clone, Copy, Debug)]
struct UnzipFanInFastPath {
    fan_in_stage_index: usize,
    /// Which inlet indices of the fan-in stage the two Unzip outlets connect to.
    /// `target_inlet_indices[0]` is the index of the inlet connected to `out0`,
    /// and `target_inlet_indices[1]` is the index connected to `out1`.
    target_inlet_indices: [usize; 2],
}

#[derive(Clone, Copy, Debug)]
struct UnzipZipFastPath {
    zip_stage_index: usize,
}

#[derive(Clone, Copy, Debug)]
struct PartitionMergeFastPath {
    merge_stage_index: usize,
}

enum StageEmissions {
    None,
    One(PortId, DatumValue),
    Two((PortId, DatumValue), (PortId, DatumValue)),
    Many(Vec<(PortId, DatumValue)>),
}

struct StageTransition {
    emissions: StageEmissions,
    completed_outlets: Vec<PortId>,
    cancelled_inlets: Vec<PortId>,
}

impl StageTransition {
    fn none() -> Self {
        Self {
            emissions: StageEmissions::None,
            completed_outlets: Vec::new(),
            cancelled_inlets: Vec::new(),
        }
    }

    fn emit(emissions: StageEmissions) -> Self {
        Self {
            emissions,
            completed_outlets: Vec::new(),
            cancelled_inlets: Vec::new(),
        }
    }

    fn with_completion(mut self, completed_outlets: Vec<PortId>) -> Self {
        self.completed_outlets = completed_outlets;
        self
    }

    fn with_cancellations(mut self, cancelled_inlets: Vec<PortId>) -> Self {
        self.cancelled_inlets = cancelled_inlets;
        self
    }
}

impl<'a, S: Shape> FusedExecutor<'a, S> {
    fn is_outlet_cancelled(&self, outlet: PortId) -> bool {
        !self.cancelled_outlets.is_empty() && self.cancelled_outlets.contains(&outlet)
    }

    fn new(graph: &'a GraphBlueprint<S>, config: FusedExecutionConfig) -> Self {
        let edge_by_outlet = graph
            .edges
            .iter()
            .map(|edge| (edge.outlet, edge.inlet))
            .collect();
        let edge_by_inlet = graph
            .edges
            .iter()
            .map(|edge| (edge.inlet, edge.outlet))
            .collect();
        let mut stage_by_inlet = HashMap::new();
        let mut stage_by_outlet = HashMap::new();
        for (index, stage) in graph.stages.iter().enumerate() {
            for inlet in &stage.spec.inlets {
                stage_by_inlet.insert(inlet.id(), index);
            }
            for outlet in &stage.spec.outlets {
                stage_by_outlet.insert(outlet.id(), index);
            }
        }

        let opaque_logics: Vec<_> = graph
            .stages
            .iter()
            .map(|stage| stage.logic_factory.as_ref().map(|factory| factory()))
            .collect();

        let stage_states = graph
            .stages
            .iter()
            .map(|stage| match stage.spec.kind {
                StageKind::Broadcast => StageState::Broadcast {
                    fast_path: broadcast_zip_fast_path(
                        stage,
                        graph,
                        &edge_by_outlet,
                        &stage_by_inlet,
                    ),
                    cancelled_outlets: vec![false; stage.spec.outlets.len()],
                    live_outlets: stage.spec.outlets.len(),
                },
                StageKind::Balance => StageState::Balance {
                    fast_path: balance_merge_fast_path(
                        stage,
                        graph,
                        &edge_by_outlet,
                        &stage_by_inlet,
                    ),
                    next: 0,
                    cancelled_outlets: vec![false; stage.spec.outlets.len()],
                    live_outlets: stage.spec.outlets.len(),
                },
                StageKind::Merge => StageState::Merge {
                    open_inputs: stage.spec.inlets.len(),
                    eager_complete: false,
                    completed: false,
                },
                StageKind::MergePreferred => StageState::Merge {
                    open_inputs: stage.spec.inlets.len(),
                    eager_complete: false,
                    completed: false,
                },
                StageKind::MergePrioritized { .. } => StageState::Merge {
                    open_inputs: stage.spec.inlets.len(),
                    eager_complete: false,
                    completed: false,
                },
                StageKind::Concat => StageState::Merge {
                    open_inputs: stage.spec.inlets.len(),
                    eager_complete: false,
                    completed: false,
                },
                StageKind::Interleave { eager_close, .. } => StageState::Merge {
                    open_inputs: stage.spec.inlets.len(),
                    eager_complete: eager_close,
                    completed: false,
                },
                StageKind::OrElse { primary_inlet: _ } => StageState::OrElse {
                    primary_emitted: false,
                    buffer: VecDeque::new(),
                    primary_closed: false,
                    secondary_closed: false,
                    completed: false,
                },
                StageKind::Zip(_) => {
                    if let [left, right] = stage.spec.inlets.as_slice() {
                        StageState::Zip {
                            left_inlet: left.id(),
                            right_inlet: right.id(),
                            left: VecDeque::new(),
                            right: VecDeque::new(),
                            left_pending_complete: false,
                            right_pending_complete: false,
                            completed: false,
                        }
                    } else {
                        StageState::Stateless
                    }
                }
                StageKind::Unzip { .. } => StageState::Unzip {
                    fast_path: unzip_fan_in_fast_path(
                        stage,
                        graph,
                        &edge_by_outlet,
                        &stage_by_inlet,
                    ),
                    zip_fast_path: unzip_zip_fast_path(
                        stage,
                        graph,
                        &edge_by_outlet,
                        &stage_by_inlet,
                    ),
                    demand: [false; 2],
                    cancelled: [false; 2],
                    upstream_closed: false,
                },
                StageKind::MergeSorted(_) => StageState::MergeSorted {
                    left: VecDeque::new(),
                    right: VecDeque::new(),
                    left_closed: false,
                    right_closed: false,
                    pending: VecDeque::new(),
                    completed: false,
                },
                StageKind::MergeSequence { .. } => StageState::MergeSequence {
                    next_sequence: 0,
                    pending: Vec::new(),
                    completed_count: 0,
                    output_buffer: VecDeque::new(),
                    completed: false,
                },
                StageKind::MergeLatest { input_count, .. } => {
                    let mut latest = Vec::with_capacity(input_count);
                    for _ in 0..input_count {
                        latest.push(None);
                    }
                    StageState::MergeLatest {
                        latest,
                        seen_count: 0,
                        completed_count: 0,
                        pending: VecDeque::new(),
                        completed: false,
                    }
                }
                StageKind::Partition { output_count, .. } => StageState::Partition {
                    fast_path: partition_merge_fast_path(
                        stage,
                        graph,
                        &edge_by_outlet,
                        &stage_by_inlet,
                    ),
                    pending: None,
                    upstream_closed: false,
                    demand: vec![false; output_count],
                    cancelled: vec![false; output_count],
                    output_count,
                    completed: false,
                },
                _ => StageState::Stateless,
            })
            .collect();

        let mut executor = Self {
            graph,
            edge_by_outlet,
            edge_by_inlet,
            stage_by_inlet,
            stage_by_outlet,
            stage_states,
            opaque_logics,
            config,
            events: 0,
            async_boundary_crossings: 0,
            cancelled_outlets: HashSet::new(),
        };
        executor.prime_connected_demands();
        executor
    }

    fn deliver<Out>(
        &mut self,
        inlet: PortId,
        value: DatumValue,
        graph_outlet: PortId,
        output: &mut impl FusedOutputSink<Out>,
    ) -> StreamResult<()>
    where
        Out: Send + 'static,
    {
        self.bump_event()?;
        let stage_index = *self.stage_by_inlet.get(&inlet).ok_or_else(|| {
            StreamError::GraphValidation(format!("inlet {} has no owning stage", inlet.as_usize()))
        })?;

        let transition = self.process_stage(stage_index, inlet, value)?;
        self.emit_all(transition.emissions, graph_outlet, output)?;
        self.complete_all(transition.completed_outlets, graph_outlet, output)?;
        self.cancel_all(transition.cancelled_inlets, graph_outlet, output)
    }

    fn complete<Out>(
        &mut self,
        inlet: PortId,
        graph_outlet: PortId,
        output: &mut impl FusedOutputSink<Out>,
    ) -> StreamResult<()>
    where
        Out: Send + 'static,
    {
        self.bump_event()?;
        let stage_index = *self.stage_by_inlet.get(&inlet).ok_or_else(|| {
            StreamError::GraphValidation(format!("inlet {} has no owning stage", inlet.as_usize()))
        })?;
        let transition = self.process_completion(stage_index, inlet)?;
        self.emit_all(transition.emissions, graph_outlet, output)?;
        self.complete_all(transition.completed_outlets, graph_outlet, output)?;
        self.cancel_all(transition.cancelled_inlets, graph_outlet, output)
    }

    fn request<Out>(
        &mut self,
        outlet: PortId,
        graph_outlet: PortId,
        output: &mut impl FusedOutputSink<Out>,
    ) -> StreamResult<()>
    where
        Out: Send + 'static,
    {
        if self.is_outlet_cancelled(outlet) {
            return Ok(());
        }
        self.bump_event()?;
        let Some(stage_index) = self.stage_by_outlet.get(&outlet).copied() else {
            return Ok(());
        };
        let transition = self.process_pull(stage_index, outlet)?;
        self.emit_all(transition.emissions, graph_outlet, output)?;
        self.complete_all(transition.completed_outlets, graph_outlet, output)?;
        self.cancel_all(transition.cancelled_inlets, graph_outlet, output)
    }

    fn downstream_finish<Out>(
        &mut self,
        outlet: PortId,
        graph_outlet: PortId,
        output: &mut impl FusedOutputSink<Out>,
    ) -> StreamResult<()>
    where
        Out: Send + 'static,
    {
        if !self.cancelled_outlets.insert(outlet) {
            return Ok(());
        }
        self.bump_event()?;
        let stage_index = *self.stage_by_outlet.get(&outlet).ok_or_else(|| {
            StreamError::GraphValidation(format!(
                "outlet {} has no owning stage",
                outlet.as_usize()
            ))
        })?;
        let transition = self.process_downstream_finish(stage_index, outlet)?;
        self.emit_all(transition.emissions, graph_outlet, output)?;
        self.complete_all(transition.completed_outlets, graph_outlet, output)?;
        self.cancel_all(transition.cancelled_inlets, graph_outlet, output)
    }

    fn emit_all<Out>(
        &mut self,
        emitted: StageEmissions,
        graph_outlet: PortId,
        output: &mut impl FusedOutputSink<Out>,
    ) -> StreamResult<()>
    where
        Out: Send + 'static,
    {
        match emitted {
            StageEmissions::None => Ok(()),
            StageEmissions::One(outlet, value) => self.emit(outlet, value, graph_outlet, output),
            StageEmissions::Two(first, second) => {
                self.emit(first.0, first.1, graph_outlet, output)?;
                self.emit(second.0, second.1, graph_outlet, output)
            }
            StageEmissions::Many(emitted) => {
                for (outlet, value) in emitted {
                    self.emit(outlet, value, graph_outlet, output)?;
                }
                Ok(())
            }
        }
    }

    fn complete_all<Out>(
        &mut self,
        completed_outlets: Vec<PortId>,
        graph_outlet: PortId,
        output: &mut impl FusedOutputSink<Out>,
    ) -> StreamResult<()>
    where
        Out: Send + 'static,
    {
        for outlet in completed_outlets {
            if self.is_outlet_cancelled(outlet) {
                continue;
            }
            self.complete_outlet(outlet, graph_outlet, output)?;
        }
        Ok(())
    }

    fn cancel_all<Out>(
        &mut self,
        cancelled_inlets: Vec<PortId>,
        graph_outlet: PortId,
        output: &mut impl FusedOutputSink<Out>,
    ) -> StreamResult<()>
    where
        Out: Send + 'static,
    {
        for inlet in cancelled_inlets {
            if let Some(outlet) = self.edge_by_inlet.get(&inlet).copied() {
                self.downstream_finish(outlet, graph_outlet, output)?;
            }
        }
        Ok(())
    }

    fn emit<Out>(
        &mut self,
        outlet: PortId,
        value: DatumValue,
        graph_outlet: PortId,
        output: &mut impl FusedOutputSink<Out>,
    ) -> StreamResult<()>
    where
        Out: Send + 'static,
    {
        self.bump_event()?;
        if self.is_outlet_cancelled(outlet) {
            return Ok(());
        }
        if outlet == graph_outlet {
            output.emit(downcast_datum(value, "emit", || {
                format!("outlet#{}", outlet.as_usize())
            })?)?;
            self.request(outlet, graph_outlet, output)?;
            return Ok(());
        }

        let Some(inlet) = self.edge_by_outlet.get(&outlet).copied() else {
            return Err(StreamError::GraphValidation(format!(
                "outlet {} is neither connected nor graph output",
                outlet.as_usize()
            )));
        };
        self.deliver(inlet, value, graph_outlet, output)?;
        if self
            .stage_by_outlet
            .get(&outlet)
            .copied()
            .is_some_and(|stage_index| {
                matches!(
                    self.graph.stages[stage_index].spec.kind,
                    StageKind::Opaque | StageKind::Unzip { .. } | StageKind::Partition { .. }
                )
            })
        {
            self.request(outlet, graph_outlet, output)?;
        }
        Ok(())
    }

    fn complete_outlet<Out>(
        &mut self,
        outlet: PortId,
        graph_outlet: PortId,
        output: &mut impl FusedOutputSink<Out>,
    ) -> StreamResult<()>
    where
        Out: Send + 'static,
    {
        self.bump_event()?;
        if outlet == graph_outlet {
            return Ok(());
        }
        let Some(inlet) = self.edge_by_outlet.get(&outlet).copied() else {
            return Err(StreamError::GraphValidation(format!(
                "outlet {} is neither connected nor graph output",
                outlet.as_usize()
            )));
        };
        self.complete(inlet, graph_outlet, output)
    }

    fn process_stage(
        &mut self,
        stage_index: usize,
        inlet: PortId,
        value: DatumValue,
    ) -> StreamResult<StageTransition> {
        let stage = &self.graph.stages[stage_index];
        match &stage.spec.kind {
            StageKind::Identity => Ok(StageTransition::emit(StageEmissions::One(
                single_outlet(stage)?,
                value,
            ))),
            StageKind::Opaque => {
                if let Some(logic) = self
                    .opaque_logics
                    .get_mut(stage_index)
                    .and_then(|l| l.as_mut())
                {
                    logic.drain_async_callbacks();
                    let inlet_ref = stage.spec.inlets.iter().find(|i| i.id() == inlet).cloned();
                    if let Some(inlet_ref) = inlet_ref {
                        logic.offer_datum(inlet, value)?;
                        let mut handler = logic.take_in_handler(inlet);
                        let result = if let Some(ref mut handler) = handler {
                            let inlet_any = inlet_ref;
                            handler.on_push(logic, inlet_any)
                        } else {
                            Ok(())
                        };
                        if let Some(handler) = handler {
                            logic.restore_in_handler(inlet, handler);
                        }
                        result?;
                    }
                    self.collect_opaque_emissions(stage, stage_index)
                } else {
                    Ok(StageTransition::emit(StageEmissions::One(
                        single_outlet(stage)?,
                        value,
                    )))
                }
            }
            StageKind::Map(map) => Ok(StageTransition::emit(StageEmissions::One(
                single_outlet(stage)?,
                (map.erased)(value)?,
            ))),
            StageKind::AsyncBoundary => {
                self.async_boundary_crossings += 1;
                Ok(StageTransition::emit(StageEmissions::One(
                    single_outlet(stage)?,
                    value,
                )))
            }
            StageKind::Broadcast => {
                let fast_path = match &self.stage_states[stage_index] {
                    StageState::Broadcast { fast_path, .. } => *fast_path,
                    _ => None,
                };
                if let Some(fast_path) = fast_path {
                    self.broadcast_zip_emissions(fast_path, value)
                        .map(StageTransition::emit)
                } else {
                    broadcast_emissions(&stage.spec.outlets, value).map(StageTransition::emit)
                }
            }
            StageKind::Balance => {
                let outlets = &stage.spec.outlets;
                if outlets.is_empty() {
                    return Err(StreamError::GraphValidation(
                        "balance has no outlets".into(),
                    ));
                }
                let StageState::Balance {
                    fast_path,
                    next,
                    cancelled_outlets,
                    live_outlets,
                } = &mut self.stage_states[stage_index]
                else {
                    return Err(StreamError::GraphValidation(
                        "balance state is missing".into(),
                    ));
                };
                if *live_outlets == 0 {
                    return Ok(StageTransition::none());
                }
                if let Some(fast_path) = *fast_path {
                    return self
                        .balance_merge_emissions(fast_path, value)
                        .map(StageTransition::emit);
                }
                let mut selected = None;
                for offset in 0..outlets.len() {
                    let index = (*next + offset) % outlets.len();
                    if !cancelled_outlets[index] {
                        selected = Some(index);
                        break;
                    }
                }
                let index = selected.ok_or_else(|| {
                    StreamError::GraphValidation("balance has no live outlets".into())
                })?;
                let outlet = outlets[index].id();
                *next = (index + 1) % outlets.len();
                Ok(StageTransition::emit(StageEmissions::One(outlet, value)))
            }
            StageKind::Merge | StageKind::MergePreferred | StageKind::MergePrioritized { .. } => {
                let StageState::Merge { completed, .. } = &self.stage_states[stage_index] else {
                    return Err(StreamError::GraphValidation(
                        "merge state is missing".into(),
                    ));
                };
                if *completed {
                    return Ok(StageTransition::none());
                }
                Ok(StageTransition::emit(StageEmissions::One(
                    single_outlet(stage)?,
                    value,
                )))
            }
            StageKind::Concat | StageKind::Interleave { .. } => {
                let StageState::Merge { completed, .. } = &self.stage_states[stage_index] else {
                    return Err(StreamError::GraphValidation(
                        "fan-in state is missing".into(),
                    ));
                };
                if *completed {
                    return Ok(StageTransition::none());
                }
                Ok(StageTransition::emit(StageEmissions::One(
                    single_outlet(stage)?,
                    value,
                )))
            }
            StageKind::OrElse { primary_inlet } => {
                let StageState::OrElse {
                    primary_emitted,
                    buffer,
                    primary_closed,
                    completed,
                    ..
                } = &mut self.stage_states[stage_index]
                else {
                    return Err(StreamError::GraphValidation(
                        "or-else state is missing".into(),
                    ));
                };
                if *completed {
                    return Ok(StageTransition::none());
                }
                if inlet == *primary_inlet {
                    *primary_emitted = true;
                    buffer.clear();
                    Ok(StageTransition::emit(StageEmissions::One(
                        single_outlet(stage)?,
                        value,
                    )))
                } else if *primary_emitted {
                    Ok(StageTransition::none())
                } else if *primary_closed {
                    Ok(StageTransition::emit(StageEmissions::One(
                        single_outlet(stage)?,
                        value,
                    )))
                } else {
                    buffer.push_back(value);
                    Ok(StageTransition::none())
                }
            }
            StageKind::Zip(zip) => {
                if stage.spec.inlets.len() != 2 {
                    return Err(StreamError::GraphValidation(format!(
                        "zip stage {} expected 2 inlets, got {}",
                        stage.spec.name(),
                        stage.spec.inlets.len()
                    )));
                }

                let ready = {
                    let StageState::Zip {
                        left_inlet,
                        right_inlet,
                        left,
                        right,
                        left_pending_complete,
                        right_pending_complete,
                        completed,
                    } = &mut self.stage_states[stage_index]
                    else {
                        return Err(StreamError::GraphValidation("zip state is missing".into()));
                    };

                    if *completed {
                        return Ok(StageTransition::none());
                    }

                    // Bounded per-inlet buffering: an inlet may receive several
                    // elements before its pair arrives (e.g. an asymmetric upstream
                    // topology), so queue them and pair in arrival order.
                    if inlet == *left_inlet {
                        left.push_back(value);
                    } else if inlet == *right_inlet {
                        right.push_back(value);
                    } else {
                        return Err(StreamError::GraphValidation(format!(
                            "zip inlet {} is not part of the stage",
                            inlet.as_usize()
                        )));
                    }

                    match (left.front().is_some(), right.front().is_some()) {
                        (true, true) => {
                            let left_item =
                                left.pop_front().expect("zip left buffer had an element");
                            let right_item =
                                right.pop_front().expect("zip right buffer had an element");
                            let should_complete = (*left_pending_complete && left.is_empty())
                                || (*right_pending_complete && right.is_empty());
                            Some((left_item, right_item, should_complete))
                        }
                        _ => None,
                    }
                };

                if let Some((left, right, should_complete)) = ready {
                    let outlet = single_outlet(stage)?;
                    if should_complete {
                        let StageState::Zip { completed, .. } = &mut self.stage_states[stage_index]
                        else {
                            return Err(StreamError::GraphValidation(
                                "zip state is missing".into(),
                            ));
                        };
                        *completed = true;
                    }
                    let mut transition =
                        StageTransition::emit(StageEmissions::One(outlet, zip(left, right)?));
                    if should_complete {
                        transition.completed_outlets.push(outlet);
                    }
                    Ok(transition)
                } else {
                    Ok(StageTransition::none())
                }
            }
            StageKind::Unzip { split, .. } => {
                let (fan_in, zip_fast) = match &self.stage_states[stage_index] {
                    StageState::Unzip {
                        fast_path,
                        zip_fast_path,
                        ..
                    } => (*fast_path, *zip_fast_path),
                    _ => (None, None),
                };
                if let Some(zip_fast) = zip_fast {
                    let (out0_val, out1_val) = split(value);
                    let zip_stage = &self.graph.stages[zip_fast.zip_stage_index];
                    let StageKind::Zip(zip) = &zip_stage.spec.kind else {
                        return Err(StreamError::GraphValidation(
                            "unzip-zip fast path references a non-zip stage".into(),
                        ));
                    };
                    let outlet = single_outlet(zip_stage)?;
                    let zipped = zip(out0_val, out1_val)?;
                    return Ok(StageTransition::emit(StageEmissions::One(outlet, zipped)));
                }
                if let Some(fast_path) = fan_in {
                    let (out0_val, out1_val) = split(value);
                    let target = &self.graph.stages[fast_path.fan_in_stage_index];
                    match &target.spec.kind {
                        StageKind::MergeSorted(compare) => {
                            let result = {
                                let StageState::MergeSorted {
                                    left,
                                    right,
                                    left_closed,
                                    right_closed,
                                    pending,
                                    completed,
                                } = &mut self.stage_states[fast_path.fan_in_stage_index]
                                else {
                                    return Err(StreamError::GraphValidation(
                                        "merge-sorted state is missing".into(),
                                    ));
                                };
                                if *completed {
                                    return Ok(StageTransition::none());
                                }
                                // Route each Unzip output to the correct left/right queue
                                // based on which inlet index each outlet connects to.
                                if fast_path.target_inlet_indices[0] == 0 {
                                    left.push_back(out0_val);
                                    right.push_back(out1_val);
                                } else {
                                    left.push_back(out1_val);
                                    right.push_back(out0_val);
                                }

                                loop {
                                    let next = match (left.front(), right.front()) {
                                        (Some(l), Some(r)) => {
                                            if compare(l, r) != std::cmp::Ordering::Greater {
                                                left.pop_front()
                                            } else {
                                                right.pop_front()
                                            }
                                        }
                                        (Some(_), None) if *right_closed => left.pop_front(),
                                        (None, Some(_)) if *left_closed => right.pop_front(),
                                        _ => break,
                                    };
                                    if let Some(val) = next {
                                        pending.push_back(val);
                                    } else {
                                        break;
                                    }
                                }

                                if let Some(output) = pending.pop_front() {
                                    let outlet = single_outlet(target)?;
                                    let all_done = *left_closed
                                        && *right_closed
                                        && left.is_empty()
                                        && right.is_empty()
                                        && pending.is_empty();
                                    if all_done {
                                        *completed = true;
                                        StageTransition::emit(StageEmissions::One(outlet, output))
                                            .with_completion(vec![outlet])
                                    } else {
                                        StageTransition::emit(StageEmissions::One(outlet, output))
                                    }
                                } else {
                                    StageTransition::none()
                                }
                            };
                            Ok(result)
                        }
                        StageKind::MergeSequence {
                            extract_sequence,
                            input_count,
                            ..
                        } => {
                            let result = {
                                let StageState::MergeSequence {
                                    next_sequence,
                                    pending,
                                    completed_count,
                                    output_buffer,
                                    completed,
                                } = &mut self.stage_states[fast_path.fan_in_stage_index]
                                else {
                                    return Err(StreamError::GraphValidation(
                                        "merge-sequence state is missing".into(),
                                    ));
                                };
                                if *completed {
                                    return Ok(StageTransition::none());
                                }
                                for val in [out0_val, out1_val] {
                                    let seq = extract_sequence(&val);
                                    if seq == *next_sequence {
                                        output_buffer.push_back(val);
                                        *next_sequence += 1;
                                        while let Some(index) =
                                            pending.iter().position(|(s, _)| *s == *next_sequence)
                                        {
                                            let (_, item) = pending.remove(index);
                                            output_buffer.push_back(item);
                                            *next_sequence += 1;
                                        }
                                    } else {
                                        if pending.iter().any(|(s, _)| *s == seq) {
                                            return Err(StreamError::Failed(format!(
                                                "duplicate sequence {seq} on merge sequence"
                                            )));
                                        }
                                        pending.push((seq, val));
                                        pending.sort_by_key(|(s, _)| *s);
                                        while let Some(index) =
                                            pending.iter().position(|(s, _)| *s == *next_sequence)
                                        {
                                            let (_, item) = pending.remove(index);
                                            output_buffer.push_back(item);
                                            *next_sequence += 1;
                                        }
                                    }
                                }

                                if !output_buffer.is_empty() {
                                    let outlet = single_outlet(target)?;
                                    let all_done = *completed_count >= *input_count;
                                    let emissions: Vec<_> =
                                        output_buffer.drain(..).map(|v| (outlet, v)).collect();
                                    if all_done {
                                        *completed = true;
                                        StageTransition::emit(StageEmissions::Many(emissions))
                                            .with_completion(vec![outlet])
                                    } else {
                                        StageTransition::emit(StageEmissions::Many(emissions))
                                    }
                                } else {
                                    StageTransition::none()
                                }
                            };
                            Ok(result)
                        }
                        StageKind::MergeLatest {
                            input_count,
                            build_snapshot,
                            ..
                        } => {
                            let result = {
                                let StageState::MergeLatest {
                                    latest,
                                    seen_count,
                                    completed_count,
                                    pending,
                                    completed,
                                } = &mut self.stage_states[fast_path.fan_in_stage_index]
                                else {
                                    return Err(StreamError::GraphValidation(
                                        "merge-latest state is missing".into(),
                                    ));
                                };
                                if *completed {
                                    return Ok(StageTransition::none());
                                }
                                let inlets = &target.spec.inlets;
                                for (idx, val) in fast_path
                                    .target_inlet_indices
                                    .into_iter()
                                    .zip([out0_val, out1_val])
                                {
                                    if idx < inlets.len() && latest[idx].is_none() {
                                        *seen_count += 1;
                                    }
                                    latest[idx] = Some(val);
                                    if *seen_count >= *input_count {
                                        let values: Vec<&DatumValue> =
                                            latest.iter().filter_map(|v| v.as_ref()).collect();
                                        let snapshot = build_snapshot(&values);
                                        pending.push_back(snapshot);
                                    }
                                }

                                if !pending.is_empty() {
                                    let outlet = single_outlet(target)?;
                                    let all_done = *completed_count >= *input_count;
                                    let emissions: Vec<_> =
                                        pending.drain(..).map(|v| (outlet, v)).collect();
                                    if all_done {
                                        *completed = true;
                                        StageTransition::emit(StageEmissions::Many(emissions))
                                            .with_completion(vec![outlet])
                                    } else {
                                        StageTransition::emit(StageEmissions::Many(emissions))
                                    }
                                } else {
                                    StageTransition::none()
                                }
                            };
                            Ok(result)
                        }
                        _ => {
                            let transition = {
                                let StageState::Unzip { cancelled, .. } =
                                    &self.stage_states[stage_index]
                                else {
                                    return Err(StreamError::GraphValidation(
                                        "unzip state is missing".into(),
                                    ));
                                };
                                let out0 = stage.spec.outlets.first().map(AnyOutlet::id);
                                let out1 = stage.spec.outlets.get(1).map(AnyOutlet::id);
                                let live0 = out0.is_some() && !cancelled[0];
                                let live1 = out1.is_some() && !cancelled[1];
                                if live0 && live1 {
                                    StageTransition::emit(StageEmissions::Two(
                                        (out0.unwrap(), out0_val),
                                        (out1.unwrap(), out1_val),
                                    ))
                                } else if live0 {
                                    StageTransition::emit(StageEmissions::One(
                                        out0.unwrap(),
                                        out0_val,
                                    ))
                                } else if live1 {
                                    StageTransition::emit(StageEmissions::One(
                                        out1.unwrap(),
                                        out1_val,
                                    ))
                                } else {
                                    StageTransition::none()
                                }
                            };
                            Ok(transition)
                        }
                    }
                } else {
                    let transition = {
                        let StageState::Unzip { cancelled, .. } = &self.stage_states[stage_index]
                        else {
                            return Err(StreamError::GraphValidation(
                                "unzip state is missing".into(),
                            ));
                        };
                        let (out0_val, out1_val) = split(value);
                        let out0 = stage.spec.outlets.first().map(AnyOutlet::id);
                        let out1 = stage.spec.outlets.get(1).map(AnyOutlet::id);
                        let live0 = out0.is_some() && !cancelled[0];
                        let live1 = out1.is_some() && !cancelled[1];
                        if live0 && live1 {
                            StageTransition::emit(StageEmissions::Two(
                                (out0.unwrap(), out0_val),
                                (out1.unwrap(), out1_val),
                            ))
                        } else if live0 {
                            StageTransition::emit(StageEmissions::One(out0.unwrap(), out0_val))
                        } else if live1 {
                            StageTransition::emit(StageEmissions::One(out1.unwrap(), out1_val))
                        } else {
                            StageTransition::none()
                        }
                    };
                    Ok(transition)
                }
            }
            StageKind::MergeSorted(compare) => {
                let result = {
                    let StageState::MergeSorted {
                        left,
                        right,
                        left_closed,
                        right_closed,
                        pending,
                        completed,
                    } = &mut self.stage_states[stage_index]
                    else {
                        return Err(StreamError::GraphValidation(
                            "merge-sorted state is missing".into(),
                        ));
                    };
                    if *completed {
                        return Ok(StageTransition::none());
                    }
                    let is_left = stage.spec.inlets.first().is_some_and(|i| i.id() == inlet);
                    if is_left {
                        left.push_back(value);
                    } else {
                        right.push_back(value);
                    }

                    loop {
                        let next = match (left.front(), right.front()) {
                            (Some(l), Some(r)) => {
                                if compare(l, r) != std::cmp::Ordering::Greater {
                                    left.pop_front()
                                } else {
                                    right.pop_front()
                                }
                            }
                            (Some(_), None) if *right_closed => left.pop_front(),
                            (None, Some(_)) if *left_closed => right.pop_front(),
                            _ => break,
                        };
                        if let Some(val) = next {
                            pending.push_back(val);
                        } else {
                            break;
                        }
                    }

                    if let Some(output) = pending.pop_front() {
                        let outlet = single_outlet(stage)?;
                        let all_done = *left_closed
                            && *right_closed
                            && left.is_empty()
                            && right.is_empty()
                            && pending.is_empty();
                        if all_done {
                            *completed = true;
                            StageTransition::emit(StageEmissions::One(outlet, output))
                                .with_completion(vec![outlet])
                        } else {
                            StageTransition::emit(StageEmissions::One(outlet, output))
                        }
                    } else {
                        StageTransition::none()
                    }
                };
                Ok(result)
            }
            StageKind::MergeSequence {
                input_count,
                extract_sequence,
                ..
            } => {
                let result = {
                    let StageState::MergeSequence {
                        next_sequence,
                        pending,
                        completed_count,
                        output_buffer,
                        completed,
                    } = &mut self.stage_states[stage_index]
                    else {
                        return Err(StreamError::GraphValidation(
                            "merge-sequence state is missing".into(),
                        ));
                    };
                    if *completed {
                        return Ok(StageTransition::none());
                    }
                    let seq = extract_sequence(&value);
                    if seq == *next_sequence {
                        output_buffer.push_back(value);
                        *next_sequence += 1;
                        while let Some(index) =
                            pending.iter().position(|(s, _)| *s == *next_sequence)
                        {
                            let (_, item) = pending.remove(index);
                            output_buffer.push_back(item);
                            *next_sequence += 1;
                        }
                    } else {
                        if pending.iter().any(|(s, _)| *s == seq) {
                            return Err(StreamError::Failed(format!(
                                "duplicate sequence {seq} on merge sequence"
                            )));
                        }
                        pending.push((seq, value));
                        pending.sort_by_key(|(s, _)| *s);
                        while let Some(index) =
                            pending.iter().position(|(s, _)| *s == *next_sequence)
                        {
                            let (_, item) = pending.remove(index);
                            output_buffer.push_back(item);
                            *next_sequence += 1;
                        }
                    }

                    if !output_buffer.is_empty() {
                        let outlet = single_outlet(stage)?;
                        let all_done = *completed_count >= *input_count;
                        let emissions: Vec<_> =
                            output_buffer.drain(..).map(|v| (outlet, v)).collect();
                        if all_done {
                            *completed = true;
                            StageTransition::emit(StageEmissions::Many(emissions))
                                .with_completion(vec![outlet])
                        } else {
                            StageTransition::emit(StageEmissions::Many(emissions))
                        }
                    } else {
                        StageTransition::none()
                    }
                };
                Ok(result)
            }
            StageKind::MergeLatest {
                input_count,
                build_snapshot,
                ..
            } => {
                let result = {
                    let StageState::MergeLatest {
                        latest,
                        seen_count,
                        completed_count,
                        pending,
                        completed,
                    } = &mut self.stage_states[stage_index]
                    else {
                        return Err(StreamError::GraphValidation(
                            "merge-latest state is missing".into(),
                        ));
                    };
                    if *completed {
                        return Ok(StageTransition::none());
                    }
                    let inlet_index = stage
                        .spec
                        .inlets
                        .iter()
                        .position(|i| i.id() == inlet)
                        .ok_or_else(|| {
                            StreamError::GraphValidation(format!(
                                "merge-latest inlet {} not part of stage",
                                inlet.as_usize()
                            ))
                        })?;
                    if latest[inlet_index].is_none() {
                        *seen_count += 1;
                    }
                    latest[inlet_index] = Some(value);
                    if *seen_count >= *input_count {
                        let values: Vec<&DatumValue> =
                            latest.iter().filter_map(|v| v.as_ref()).collect();
                        let snapshot = build_snapshot(&values);
                        pending.push_back(snapshot);
                    }

                    if !pending.is_empty() {
                        let outlet = single_outlet(stage)?;
                        let all_done = *completed_count >= *input_count;
                        let emissions: Vec<_> = pending.drain(..).map(|v| (outlet, v)).collect();
                        if all_done {
                            *completed = true;
                            StageTransition::emit(StageEmissions::Many(emissions))
                                .with_completion(vec![outlet])
                        } else {
                            StageTransition::emit(StageEmissions::Many(emissions))
                        }
                    } else {
                        StageTransition::none()
                    }
                };
                Ok(result)
            }
            StageKind::Partition {
                output_count,
                partitioner,
                ..
            } => {
                let fast_path = match &self.stage_states[stage_index] {
                    StageState::Partition { fast_path, .. } => *fast_path,
                    _ => None,
                };
                if let Some(fast_path) = fast_path {
                    let idx = partitioner(&value);
                    if idx >= *output_count {
                        return Err(StreamError::Failed(format!(
                            "partitioner returned out-of-bounds index {idx} for {output_count} outputs"
                        )));
                    }
                    let merge_stage = &self.graph.stages[fast_path.merge_stage_index];
                    let StageState::Merge { completed, .. } =
                        &self.stage_states[fast_path.merge_stage_index]
                    else {
                        return Err(StreamError::GraphValidation(
                            "partition-merge fast path: merge state missing".into(),
                        ));
                    };
                    if *completed {
                        return Ok(StageTransition::none());
                    }
                    return Ok(StageTransition::emit(StageEmissions::One(
                        single_outlet(merge_stage)?,
                        value,
                    )));
                }
                let result = {
                    let StageState::Partition {
                        pending,
                        upstream_closed: _,
                        demand,
                        cancelled,
                        output_count: _,
                        completed,
                        ..
                    } = &mut self.stage_states[stage_index]
                    else {
                        return Err(StreamError::GraphValidation(
                            "partition state is missing".into(),
                        ));
                    };
                    if *completed {
                        return Ok(StageTransition::none());
                    }
                    let idx = partitioner(&value);
                    if idx >= *output_count {
                        return Err(StreamError::Failed(format!(
                            "partitioner returned out-of-bounds index {idx} for {output_count} outputs"
                        )));
                    }
                    if cancelled[idx] {
                        return Ok(StageTransition::none());
                    }
                    if demand[idx] {
                        demand[idx] = false;
                        let outlet = stage.spec.outlets[idx].id();
                        StageTransition::emit(StageEmissions::One(outlet, value))
                    } else {
                        *pending = Some((idx, value));
                        StageTransition::none()
                    }
                };
                Ok(result)
            }
        }
    }

    fn process_completion(
        &mut self,
        stage_index: usize,
        inlet: PortId,
    ) -> StreamResult<StageTransition> {
        let stage = &self.graph.stages[stage_index];
        match &stage.spec.kind {
            StageKind::Identity | StageKind::Map(_) | StageKind::AsyncBoundary => {
                Ok(StageTransition::emit(StageEmissions::None)
                    .with_completion(vec![single_outlet(stage)?]))
            }
            StageKind::Opaque => {
                if let Some(logic) = self
                    .opaque_logics
                    .get_mut(stage_index)
                    .and_then(|l| l.as_mut())
                {
                    logic.drain_async_callbacks();
                    logic.complete_inlet_by_id(inlet)?;
                    let inlet_ref = stage.spec.inlets.iter().find(|i| i.id() == inlet).cloned();
                    if let Some(inlet_ref) = inlet_ref {
                        let mut handler = logic.take_in_handler(inlet);
                        let result = if let Some(ref mut handler) = handler {
                            let inlet_any = inlet_ref;
                            handler.on_upstream_finish(logic, inlet_any)
                        } else {
                            Ok(())
                        };
                        if let Some(handler) = handler {
                            logic.restore_in_handler(inlet, handler);
                        }
                        result?;
                    }
                    self.collect_opaque_emissions(stage, stage_index)
                } else {
                    Ok(StageTransition::emit(StageEmissions::None)
                        .with_completion(vec![single_outlet(stage)?]))
                }
            }
            StageKind::Broadcast | StageKind::Balance => {
                Ok(StageTransition::emit(StageEmissions::None)
                    .with_completion(stage.spec.outlets.iter().map(AnyOutlet::id).collect()))
            }
            StageKind::Merge | StageKind::MergePreferred | StageKind::MergePrioritized { .. } => {
                let StageState::Merge {
                    open_inputs,
                    eager_complete,
                    completed,
                } = &mut self.stage_states[stage_index]
                else {
                    return Err(StreamError::GraphValidation(
                        "merge state is missing".into(),
                    ));
                };

                if *completed {
                    return Ok(StageTransition::none());
                }
                if *open_inputs == 0 {
                    return Ok(StageTransition::none());
                }
                *open_inputs -= 1;
                if *eager_complete || *open_inputs == 0 {
                    *completed = true;
                    Ok(StageTransition::emit(StageEmissions::None)
                        .with_completion(vec![single_outlet(stage)?]))
                } else {
                    Ok(StageTransition::none())
                }
            }
            StageKind::Concat | StageKind::Interleave { .. } => {
                let StageState::Merge {
                    open_inputs,
                    eager_complete,
                    completed,
                } = &mut self.stage_states[stage_index]
                else {
                    return Err(StreamError::GraphValidation(
                        "fan-in state is missing".into(),
                    ));
                };

                if *completed {
                    return Ok(StageTransition::none());
                }
                if *open_inputs == 0 {
                    return Ok(StageTransition::none());
                }
                *open_inputs -= 1;
                if *eager_complete || *open_inputs == 0 {
                    *completed = true;
                    Ok(StageTransition::emit(StageEmissions::None)
                        .with_completion(vec![single_outlet(stage)?]))
                } else {
                    Ok(StageTransition::none())
                }
            }
            StageKind::OrElse { primary_inlet } => {
                let StageState::OrElse {
                    primary_emitted,
                    buffer,
                    primary_closed,
                    secondary_closed,
                    completed,
                    ..
                } = &mut self.stage_states[stage_index]
                else {
                    return Err(StreamError::GraphValidation(
                        "or-else state is missing".into(),
                    ));
                };
                if *completed {
                    return Ok(StageTransition::none());
                }
                if inlet == *primary_inlet {
                    *primary_closed = true;
                    if *primary_emitted {
                        *completed = true;
                        buffer.clear();
                        Ok(StageTransition::emit(StageEmissions::None)
                            .with_completion(vec![single_outlet(stage)?]))
                    } else {
                        let outlet = single_outlet(stage)?;
                        let emissions: Vec<_> = buffer.drain(..).map(|v| (outlet, v)).collect();
                        if *secondary_closed {
                            *completed = true;
                            let transition = if emissions.is_empty() {
                                StageTransition::emit(StageEmissions::None)
                            } else {
                                StageTransition::emit(StageEmissions::Many(emissions))
                            };
                            Ok(transition.with_completion(vec![outlet]))
                        } else {
                            if emissions.is_empty() {
                                Ok(StageTransition::none())
                            } else {
                                Ok(StageTransition::emit(StageEmissions::Many(emissions)))
                            }
                        }
                    }
                } else {
                    *secondary_closed = true;
                    if *primary_closed && !*primary_emitted {
                        let outlet = single_outlet(stage)?;
                        let emissions: Vec<_> = buffer.drain(..).map(|v| (outlet, v)).collect();
                        *completed = true;
                        if emissions.is_empty() {
                            Ok(StageTransition::emit(StageEmissions::None)
                                .with_completion(vec![outlet]))
                        } else {
                            Ok(StageTransition::emit(StageEmissions::Many(emissions))
                                .with_completion(vec![outlet]))
                        }
                    } else {
                        Ok(StageTransition::none())
                    }
                }
            }
            StageKind::Zip(_) => {
                let StageState::Zip {
                    left_inlet,
                    right_inlet,
                    left,
                    right,
                    left_pending_complete,
                    right_pending_complete,
                    completed,
                } = &mut self.stage_states[stage_index]
                else {
                    return Err(StreamError::GraphValidation("zip state is missing".into()));
                };
                if *completed {
                    return Ok(StageTransition::none());
                }
                let finishes_left = inlet == *left_inlet;
                let finishes_right = inlet == *right_inlet;
                if (finishes_left && left.is_empty()) || (finishes_right && right.is_empty()) {
                    *completed = true;
                    Ok(StageTransition::emit(StageEmissions::None)
                        .with_completion(vec![single_outlet(stage)?]))
                } else {
                    if finishes_left {
                        *left_pending_complete = true;
                    }
                    if finishes_right {
                        *right_pending_complete = true;
                    }
                    Ok(StageTransition::none())
                }
            }
            StageKind::Unzip { .. } => {
                let (fan_in, zip_fast) = match &self.stage_states[stage_index] {
                    StageState::Unzip {
                        fast_path,
                        zip_fast_path,
                        ..
                    } => (*fast_path, *zip_fast_path),
                    _ => (None, None),
                };
                if let Some(zip_fast) = zip_fast {
                    let StageState::Unzip {
                        upstream_closed, ..
                    } = &mut self.stage_states[stage_index]
                    else {
                        return Err(StreamError::GraphValidation(
                            "unzip state is missing".into(),
                        ));
                    };
                    *upstream_closed = true;
                    let zip_stage = &self.graph.stages[zip_fast.zip_stage_index];
                    return Ok(StageTransition::emit(StageEmissions::None)
                        .with_completion(vec![single_outlet(zip_stage)?]));
                }
                if let Some(fast_path) = fan_in {
                    let StageState::Unzip {
                        upstream_closed, ..
                    } = &mut self.stage_states[stage_index]
                    else {
                        return Err(StreamError::GraphValidation(
                            "unzip state is missing".into(),
                        ));
                    };
                    *upstream_closed = true;
                    let target_stage = &self.graph.stages[fast_path.fan_in_stage_index];
                    // Only notify the two inlets that the Unzip outlets are wired to —
                    // do not call process_completion on unrelated inlets of the fan-in
                    // stage (e.g. a third inlet fed by a separate source).
                    let target_inlets = [
                        target_stage.spec.inlets[fast_path.target_inlet_indices[0]].id(),
                        target_stage.spec.inlets[fast_path.target_inlet_indices[1]].id(),
                    ];
                    let mut combined = StageTransition::none();
                    for target_inlet in target_inlets {
                        let t =
                            self.process_completion(fast_path.fan_in_stage_index, target_inlet)?;
                        combined.emissions = merge_emissions(combined.emissions, t.emissions);
                        combined.completed_outlets.extend(t.completed_outlets);
                        combined.cancelled_inlets.extend(t.cancelled_inlets);
                    }
                    Ok(combined)
                } else {
                    let StageState::Unzip {
                        upstream_closed, ..
                    } = &mut self.stage_states[stage_index]
                    else {
                        return Err(StreamError::GraphValidation(
                            "unzip state is missing".into(),
                        ));
                    };
                    *upstream_closed = true;
                    Ok(StageTransition::emit(StageEmissions::None)
                        .with_completion(stage.spec.outlets.iter().map(AnyOutlet::id).collect()))
                }
            }
            StageKind::MergeSorted(compare) => {
                let result = {
                    let StageState::MergeSorted {
                        left,
                        right,
                        left_closed,
                        right_closed,
                        pending,
                        completed,
                    } = &mut self.stage_states[stage_index]
                    else {
                        return Err(StreamError::GraphValidation(
                            "merge-sorted state is missing".into(),
                        ));
                    };
                    if *completed {
                        return Ok(StageTransition::none());
                    }
                    let is_left = stage.spec.inlets.first().is_some_and(|i| i.id() == inlet);
                    if is_left {
                        *left_closed = true;
                    } else {
                        *right_closed = true;
                    }

                    loop {
                        let next = match (left.front(), right.front()) {
                            (Some(l), Some(r)) => {
                                if compare(l, r) != std::cmp::Ordering::Greater {
                                    left.pop_front()
                                } else {
                                    right.pop_front()
                                }
                            }
                            (Some(_), None) if *right_closed => left.pop_front(),
                            (None, Some(_)) if *left_closed => right.pop_front(),
                            _ => break,
                        };
                        if let Some(val) = next {
                            pending.push_back(val);
                        } else {
                            break;
                        }
                    }

                    if pending.is_empty() {
                        let all_done =
                            *left_closed && *right_closed && left.is_empty() && right.is_empty();
                        if all_done {
                            *completed = true;
                            StageTransition::emit(StageEmissions::None)
                                .with_completion(vec![single_outlet(stage)?])
                        } else {
                            StageTransition::none()
                        }
                    } else {
                        let outlet = single_outlet(stage)?;
                        let emissions: Vec<_> = pending.drain(..).map(|v| (outlet, v)).collect();
                        let all_done =
                            *left_closed && *right_closed && left.is_empty() && right.is_empty();
                        if all_done {
                            *completed = true;
                            StageTransition::emit(StageEmissions::Many(emissions))
                                .with_completion(vec![outlet])
                        } else {
                            StageTransition::emit(StageEmissions::Many(emissions))
                        }
                    }
                };
                Ok(result)
            }
            StageKind::MergeSequence { input_count, .. } => {
                let result = {
                    let StageState::MergeSequence {
                        next_sequence,
                        pending,
                        completed_count,
                        output_buffer,
                        completed,
                    } = &mut self.stage_states[stage_index]
                    else {
                        return Err(StreamError::GraphValidation(
                            "merge-sequence state is missing".into(),
                        ));
                    };
                    if *completed {
                        return Ok(StageTransition::none());
                    }
                    *completed_count += 1;
                    if *completed_count >= *input_count && output_buffer.is_empty() {
                        if !pending.is_empty() {
                            // All inputs have completed but there are buffered elements
                            // whose sequence numbers do not include `next_sequence`.
                            // This is a gap — fail exactly as the GraphStage logic does.
                            return Err(StreamError::Failed(format!(
                                "expected sequence {next_sequence}, but all input ports have pushed or are complete",
                            )));
                        }
                        *completed = true;
                        StageTransition::emit(StageEmissions::None)
                            .with_completion(vec![single_outlet(stage)?])
                    } else {
                        StageTransition::none()
                    }
                };
                Ok(result)
            }
            StageKind::MergeLatest {
                input_count,
                eager_complete,
                ..
            } => {
                let result = {
                    let StageState::MergeLatest {
                        completed_count,
                        pending,
                        completed,
                        ..
                    } = &mut self.stage_states[stage_index]
                    else {
                        return Err(StreamError::GraphValidation(
                            "merge-latest state is missing".into(),
                        ));
                    };
                    if *completed {
                        return Ok(StageTransition::none());
                    }
                    *completed_count += 1;
                    // Complete when all inputs are done, OR when eager_complete is set
                    // and there is no pending output (matches Akka ZipLatestWith semantics:
                    // complete as soon as any upstream completes if eager is true and the
                    // pending queue is drained).
                    let all_done = *completed_count >= *input_count;
                    let eager_done = *eager_complete && pending.is_empty();
                    if all_done || eager_done {
                        *completed = true;
                        StageTransition::emit(StageEmissions::None)
                            .with_completion(vec![single_outlet(stage)?])
                    } else {
                        StageTransition::none()
                    }
                };
                Ok(result)
            }
            StageKind::Partition { .. } => {
                let fast_path = match &self.stage_states[stage_index] {
                    StageState::Partition { fast_path, .. } => *fast_path,
                    _ => None,
                };
                let result = {
                    let StageState::Partition {
                        pending,
                        upstream_closed,
                        completed,
                        ..
                    } = &mut self.stage_states[stage_index]
                    else {
                        return Err(StreamError::GraphValidation(
                            "partition state is missing".into(),
                        ));
                    };
                    if *completed {
                        return Ok(StageTransition::none());
                    }
                    *upstream_closed = true;
                    if pending.is_none() {
                        *completed = true;
                        if let Some(fast_path) = fast_path {
                            let merge_stage = &self.graph.stages[fast_path.merge_stage_index];
                            StageTransition::emit(StageEmissions::None)
                                .with_completion(vec![single_outlet(merge_stage)?])
                        } else {
                            StageTransition::emit(StageEmissions::None).with_completion(
                                stage.spec.outlets.iter().map(AnyOutlet::id).collect(),
                            )
                        }
                    } else {
                        StageTransition::none()
                    }
                };
                Ok(result)
            }
        }
    }

    fn process_pull(
        &mut self,
        stage_index: usize,
        outlet: PortId,
    ) -> StreamResult<StageTransition> {
        let stage = &self.graph.stages[stage_index];
        match &stage.spec.kind {
            StageKind::Opaque => {
                if let Some(logic) = self
                    .opaque_logics
                    .get_mut(stage_index)
                    .and_then(|l| l.as_mut())
                {
                    logic.drain_async_callbacks();
                    logic.set_demand_by_id(outlet)?;
                    let outlet_ref = stage
                        .spec
                        .outlets
                        .iter()
                        .find(|o| o.id() == outlet)
                        .cloned();
                    if let Some(outlet_ref) = outlet_ref {
                        let mut handler = logic.take_out_handler(outlet);
                        let result = if let Some(ref mut handler) = handler {
                            handler.on_pull(logic, outlet_ref)
                        } else {
                            Ok(())
                        };
                        if let Some(handler) = handler
                            && handler.keep_handler()
                            && logic.get_out_handler_mut(outlet).is_none()
                        {
                            logic.restore_out_handler(outlet, handler);
                        }
                        result?;
                    }
                    self.collect_opaque_emissions(stage, stage_index)
                } else {
                    Ok(StageTransition::none())
                }
            }
            StageKind::Unzip { .. } => {
                let StageState::Unzip {
                    demand, cancelled, ..
                } = &mut self.stage_states[stage_index]
                else {
                    return Ok(StageTransition::none());
                };
                let Some(idx) = stage.spec.outlets.iter().position(|o| o.id() == outlet) else {
                    return Ok(StageTransition::none());
                };
                if idx < 2 && !cancelled[idx] {
                    demand[idx] = true;
                }
                Ok(StageTransition::none())
            }
            StageKind::Partition { .. } => {
                let result = {
                    let StageState::Partition {
                        pending,
                        upstream_closed,
                        demand,
                        cancelled,
                        completed,
                        ..
                    } = &mut self.stage_states[stage_index]
                    else {
                        return Ok(StageTransition::none());
                    };
                    if *completed {
                        return Ok(StageTransition::none());
                    }
                    let Some(idx) = stage.spec.outlets.iter().position(|o| o.id() == outlet) else {
                        return Ok(StageTransition::none());
                    };
                    if cancelled[idx] {
                        return Ok(StageTransition::none());
                    }

                    if let Some((p_idx, p_val)) = pending.take() {
                        if p_idx == idx {
                            let out = stage.spec.outlets[idx].id();
                            if *upstream_closed {
                                *completed = true;
                                StageTransition::emit(StageEmissions::One(out, p_val))
                                    .with_completion(
                                        stage.spec.outlets.iter().map(AnyOutlet::id).collect(),
                                    )
                            } else {
                                StageTransition::emit(StageEmissions::One(out, p_val))
                            }
                        } else {
                            *pending = Some((p_idx, p_val));
                            demand[idx] = true;
                            StageTransition::none()
                        }
                    } else {
                        demand[idx] = true;
                        StageTransition::none()
                    }
                };
                Ok(result)
            }
            _ => Ok(StageTransition::none()),
        }
    }

    fn process_downstream_finish(
        &mut self,
        stage_index: usize,
        outlet: PortId,
    ) -> StreamResult<StageTransition> {
        let stage = &self.graph.stages[stage_index];
        match &stage.spec.kind {
            StageKind::Broadcast => {
                let StageState::Broadcast {
                    cancelled_outlets,
                    live_outlets,
                    ..
                } = &mut self.stage_states[stage_index]
                else {
                    return Err(StreamError::GraphValidation(
                        "broadcast state is missing".into(),
                    ));
                };
                let index = stage
                    .spec
                    .outlets
                    .iter()
                    .position(|candidate| candidate.id() == outlet)
                    .ok_or_else(|| {
                        StreamError::GraphValidation(format!(
                            "broadcast outlet {} is not part of the stage",
                            outlet.as_usize()
                        ))
                    })?;
                if cancelled_outlets[index] {
                    return Ok(StageTransition::none());
                }
                cancelled_outlets[index] = true;
                *live_outlets -= 1;
                if *live_outlets == 0 {
                    Ok(StageTransition::none()
                        .with_cancellations(stage.spec.inlets.iter().map(AnyInlet::id).collect()))
                } else {
                    Ok(StageTransition::none())
                }
            }
            StageKind::Balance => {
                let StageState::Balance {
                    cancelled_outlets,
                    live_outlets,
                    ..
                } = &mut self.stage_states[stage_index]
                else {
                    return Err(StreamError::GraphValidation(
                        "balance state is missing".into(),
                    ));
                };
                let index = stage
                    .spec
                    .outlets
                    .iter()
                    .position(|candidate| candidate.id() == outlet)
                    .ok_or_else(|| {
                        StreamError::GraphValidation(format!(
                            "balance outlet {} is not part of the stage",
                            outlet.as_usize()
                        ))
                    })?;
                if cancelled_outlets[index] {
                    return Ok(StageTransition::none());
                }
                cancelled_outlets[index] = true;
                *live_outlets -= 1;
                if *live_outlets == 0 {
                    Ok(StageTransition::none()
                        .with_cancellations(stage.spec.inlets.iter().map(AnyInlet::id).collect()))
                } else {
                    Ok(StageTransition::none())
                }
            }
            StageKind::Unzip { .. } => {
                let StageState::Unzip { cancelled, .. } = &mut self.stage_states[stage_index]
                else {
                    return Err(StreamError::GraphValidation(
                        "unzip state is missing".into(),
                    ));
                };
                let idx = stage
                    .spec
                    .outlets
                    .iter()
                    .position(|o| o.id() == outlet)
                    .unwrap_or(0);
                if idx < 2 && !cancelled[idx] {
                    cancelled[idx] = true;
                    let all_cancelled = cancelled.iter().all(|c| *c);
                    if all_cancelled {
                        Ok(StageTransition::none().with_cancellations(
                            stage.spec.inlets.iter().map(AnyInlet::id).collect(),
                        ))
                    } else {
                        Ok(StageTransition::none())
                    }
                } else {
                    Ok(StageTransition::none())
                }
            }
            StageKind::MergeSorted(_)
            | StageKind::MergeSequence { .. }
            | StageKind::MergeLatest { .. } => Ok(StageTransition::none()
                .with_cancellations(stage.spec.inlets.iter().map(AnyInlet::id).collect())),
            StageKind::Partition { eager_cancel, .. } => {
                let result = {
                    let StageState::Partition {
                        pending,
                        cancelled,
                        completed,
                        ..
                    } = &mut self.stage_states[stage_index]
                    else {
                        return Err(StreamError::GraphValidation(
                            "partition state is missing".into(),
                        ));
                    };
                    if *completed {
                        return Ok(StageTransition::none());
                    }
                    let Some(idx) = stage.spec.outlets.iter().position(|o| o.id() == outlet) else {
                        return Ok(StageTransition::none());
                    };
                    if cancelled[idx] {
                        return Ok(StageTransition::none());
                    }
                    cancelled[idx] = true;
                    // If the pending element was routed to this outlet, discard it
                    if let Some((p_idx, _)) = pending
                        && *p_idx == idx
                    {
                        *pending = None;
                    }
                    let all_cancelled = cancelled.iter().all(|c| *c);
                    if all_cancelled || *eager_cancel {
                        *completed = true;
                        StageTransition::none().with_cancellations(
                            stage.spec.inlets.iter().map(AnyInlet::id).collect(),
                        )
                    } else {
                        StageTransition::none()
                    }
                };
                Ok(result)
            }
            StageKind::Opaque => {
                let no_cancelled_outlets = self.cancelled_outlets.is_empty();
                if let Some(logic) = self
                    .opaque_logics
                    .get_mut(stage_index)
                    .and_then(|l| l.as_mut())
                {
                    logic.drain_async_callbacks();
                    logic.downstream_finish_by_id(outlet, "downstream_finish")?;
                    let outlet_ref = stage
                        .spec
                        .outlets
                        .iter()
                        .find(|o| o.id() == outlet)
                        .cloned();
                    if let Some(outlet_ref) = outlet_ref {
                        let mut handler = logic.take_out_handler(outlet);
                        let result = if let Some(ref mut handler) = handler {
                            handler.on_downstream_finish(logic, outlet_ref)
                        } else {
                            Ok(())
                        };
                        if let Some(handler) = handler
                            && handler.keep_handler()
                            && logic.get_out_handler_mut(outlet).is_none()
                        {
                            logic.restore_out_handler(outlet, handler);
                        }
                        result?;
                    }
                    let all_outlets_closed = stage.spec.outlets.iter().all(|candidate| {
                        logic.is_closed_by_id(candidate.id())
                            || (!no_cancelled_outlets
                                && self.cancelled_outlets.contains(&candidate.id()))
                    });
                    let mut transition = self.collect_opaque_emissions(stage, stage_index)?;
                    if all_outlets_closed {
                        transition.cancelled_inlets =
                            stage.spec.inlets.iter().map(AnyInlet::id).collect();
                    }
                    Ok(transition)
                } else {
                    Ok(StageTransition::none()
                        .with_cancellations(stage.spec.inlets.iter().map(AnyInlet::id).collect()))
                }
            }
            _ => Ok(StageTransition::none()
                .with_cancellations(stage.spec.inlets.iter().map(AnyInlet::id).collect())),
        }
    }

    fn collect_opaque_emissions(
        &mut self,
        stage: &StageRecord,
        stage_index: usize,
    ) -> StreamResult<StageTransition> {
        if let Some(logic) = self
            .opaque_logics
            .get_mut(stage_index)
            .and_then(|l| l.as_mut())
        {
            let emissions_slots = std::mem::take(&mut logic.pending_emissions);
            let completions = std::mem::take(&mut logic.pending_completions);
            let has_stage_failed = logic.stage_error().is_some();

            let emissions = if emissions_slots.is_empty() {
                StageEmissions::None
            } else if emissions_slots.len() == 1 {
                let (port, val) = emissions_slots.into_iter().next().unwrap();
                StageEmissions::One(port, val)
            } else {
                StageEmissions::Many(emissions_slots)
            };

            if has_stage_failed {
                let _ = has_stage_failed;
            }

            Ok(StageTransition {
                emissions,
                completed_outlets: completions,
                cancelled_inlets: Vec::new(),
            })
        } else {
            Ok(StageTransition::emit(StageEmissions::None)
                .with_completion(vec![single_outlet(stage)?]))
        }
    }

    fn bump_event(&mut self) -> StreamResult<()> {
        bump_fused_event(&mut self.events, self.config)
    }

    fn broadcast_zip_emissions(
        &mut self,
        fast_path: BroadcastZipFastPath,
        value: DatumValue,
    ) -> StreamResult<StageEmissions> {
        // 4 = broadcast fan-out (2) + zip pair (~2) stage events the unfused path would bump.
        for _ in 0..4 {
            self.bump_event()?;
        }

        let zip_stage = &self.graph.stages[fast_path.zip_stage_index];
        let StageKind::Zip(zip) = &zip_stage.spec.kind else {
            return Err(StreamError::GraphValidation(
                "broadcast zip fast path references a non-zip stage".into(),
            ));
        };

        let cloned = value.clone_box();
        let (left, right) = if fast_path.left_uses_clone {
            (cloned, value)
        } else {
            (value, cloned)
        };

        Ok(StageEmissions::One(
            single_outlet(zip_stage)?,
            zip(left, right)?,
        ))
    }

    fn balance_merge_emissions(
        &mut self,
        fast_path: BalanceMergeFastPath,
        value: DatumValue,
    ) -> StreamResult<StageEmissions> {
        // 4 = balance route (~2) + merge handoff (~2) stage events the unfused path would bump.
        for _ in 0..4 {
            self.bump_event()?;
        }

        let merge_stage = &self.graph.stages[fast_path.merge_stage_index];
        let StageKind::Merge = merge_stage.spec.kind else {
            return Err(StreamError::GraphValidation(
                "balance merge fast path references a non-merge stage".into(),
            ));
        };

        Ok(StageEmissions::One(single_outlet(merge_stage)?, value))
    }

    fn prime_connected_demands(&mut self) {
        for (stage_index, stage) in self.graph.stages.iter().enumerate() {
            match &stage.spec.kind {
                StageKind::Opaque => {
                    let Some(logic) = self
                        .opaque_logics
                        .get_mut(stage_index)
                        .and_then(|logic| logic.as_mut())
                    else {
                        continue;
                    };
                    for outlet in &stage.spec.outlets {
                        if self.edge_by_outlet.contains_key(&outlet.id()) {
                            let _ = logic.set_demand_by_id(outlet.id());
                        }
                    }
                }
                StageKind::Unzip { .. } => {
                    let StageState::Unzip { demand, .. } = &mut self.stage_states[stage_index]
                    else {
                        continue;
                    };
                    for (idx, outlet) in stage.spec.outlets.iter().enumerate() {
                        if self.edge_by_outlet.contains_key(&outlet.id()) {
                            demand[idx] = true;
                        }
                    }
                }
                StageKind::Partition { .. } => {
                    let StageState::Partition {
                        demand,
                        output_count,
                        ..
                    } = &mut self.stage_states[stage_index]
                    else {
                        continue;
                    };
                    for (idx, demand_slot) in demand.iter_mut().enumerate().take(*output_count) {
                        if idx < stage.spec.outlets.len()
                            && self
                                .edge_by_outlet
                                .contains_key(&stage.spec.outlets[idx].id())
                        {
                            *demand_slot = true;
                        }
                    }
                }
                _ => {}
            }
        }
    }
}

impl<Left, Right> GraphBlueprint<ZipShape<Left, Right>>
where
    Left: Clone + Send + 'static,
    Right: Clone + Send + 'static,
{
    pub fn run_zip(&self, left: Vec<Left>, right: Vec<Right>) -> StreamResult<Vec<(Left, Right)>> {
        Ok(self
            .run_zip_report(left, right, FusedExecutionConfig::default())?
            .output)
    }

    pub fn run_zip_report(
        &self,
        left: Vec<Left>,
        right: Vec<Right>,
        config: FusedExecutionConfig,
    ) -> StreamResult<FusedExecutionReport<(Left, Right)>> {
        let mut left = left.into_iter();
        let mut right = right.into_iter();
        let left_inlet = self.shape.in0().id();
        let right_inlet = self.shape.in1().id();
        let outlet = self.shape.outlet().id();
        let mut executor = FusedExecutor::new(self, config);
        let mut output = Vec::with_capacity(left.len().min(right.len()));
        let mut left_completed = false;
        let mut right_completed = false;

        {
            let mut output_sink = VecOutputSink {
                output: &mut output,
            };
            if left.len() == 0 {
                executor.complete(left_inlet, outlet, &mut output_sink)?;
                left_completed = true;
            }
            if right.len() == 0 {
                executor.complete(right_inlet, outlet, &mut output_sink)?;
                right_completed = true;
            }

            while left.len() > 0 || right.len() > 0 {
                if let Some(item) = left.next() {
                    executor.deliver(left_inlet, datum(item), outlet, &mut output_sink)?;
                    if left.len() == 0 && !left_completed {
                        executor.complete(left_inlet, outlet, &mut output_sink)?;
                        left_completed = true;
                    }
                }
                if let Some(item) = right.next() {
                    executor.deliver(right_inlet, datum(item), outlet, &mut output_sink)?;
                    if right.len() == 0 && !right_completed {
                        executor.complete(right_inlet, outlet, &mut output_sink)?;
                        right_completed = true;
                    }
                }
            }
        }

        Ok(FusedExecutionReport {
            output,
            events: executor.events,
            async_boundary_crossings: executor.async_boundary_crossings,
        })
    }
}

fn broadcast_zip_fast_path<S: Shape>(
    stage: &StageRecord,
    graph: &GraphBlueprint<S>,
    edge_by_outlet: &HashMap<PortId, PortId>,
    stage_by_inlet: &HashMap<PortId, usize>,
) -> Option<BroadcastZipFastPath> {
    let [first_outlet, second_outlet] = stage.spec.outlets.as_slice() else {
        return None;
    };

    let first_inlet = edge_by_outlet.get(&first_outlet.id()).copied()?;
    let second_inlet = edge_by_outlet.get(&second_outlet.id()).copied()?;
    let first_stage = stage_by_inlet.get(&first_inlet).copied()?;
    let second_stage = stage_by_inlet.get(&second_inlet).copied()?;
    if first_stage != second_stage {
        return None;
    }

    let zip_stage = &graph.stages[first_stage];
    if !matches!(zip_stage.spec.kind, StageKind::Zip(_)) {
        return None;
    }
    let [left_inlet, right_inlet] = zip_stage.spec.inlets.as_slice() else {
        return None;
    };

    if first_inlet == left_inlet.id() && second_inlet == right_inlet.id() {
        Some(BroadcastZipFastPath {
            zip_stage_index: first_stage,
            left_uses_clone: true,
        })
    } else if first_inlet == right_inlet.id() && second_inlet == left_inlet.id() {
        Some(BroadcastZipFastPath {
            zip_stage_index: first_stage,
            left_uses_clone: false,
        })
    } else {
        None
    }
}

fn balance_merge_fast_path<S: Shape>(
    stage: &StageRecord,
    graph: &GraphBlueprint<S>,
    edge_by_outlet: &HashMap<PortId, PortId>,
    stage_by_inlet: &HashMap<PortId, usize>,
) -> Option<BalanceMergeFastPath> {
    let outlets = &stage.spec.outlets;
    let first_inlet = edge_by_outlet.get(&outlets.first()?.id()).copied()?;
    let merge_stage_index = *stage_by_inlet.get(&first_inlet)?;
    let merge_stage = graph.stages.get(merge_stage_index)?;
    if !matches!(merge_stage.spec.kind, StageKind::Merge) {
        return None;
    }
    if merge_stage.spec.inlets.len() != outlets.len() {
        return None;
    }
    for (outlet, merge_inlet) in outlets.iter().zip(merge_stage.spec.inlets.iter()) {
        if edge_by_outlet.get(&outlet.id()).copied() != Some(merge_inlet.id()) {
            return None;
        }
    }
    Some(BalanceMergeFastPath { merge_stage_index })
}

fn unzip_fan_in_fast_path<S: Shape>(
    stage: &StageRecord,
    graph: &GraphBlueprint<S>,
    edge_by_outlet: &HashMap<PortId, PortId>,
    stage_by_inlet: &HashMap<PortId, usize>,
) -> Option<UnzipFanInFastPath> {
    let outlets = &stage.spec.outlets;
    if outlets.len() != 2 {
        return None;
    }
    let inlet0 = edge_by_outlet.get(&outlets[0].id()).copied()?;
    let inlet1 = edge_by_outlet.get(&outlets[1].id()).copied()?;
    let stage0 = *stage_by_inlet.get(&inlet0)?;
    let stage1 = *stage_by_inlet.get(&inlet1)?;
    if stage0 != stage1 {
        return None;
    }
    let target = graph.stages.get(stage0)?;
    if !matches!(
        target.spec.kind,
        StageKind::MergeSorted(_) | StageKind::MergeSequence { .. } | StageKind::MergeLatest { .. }
    ) {
        return None;
    }
    // Resolve the exact inlet indices so the fast path routes each Unzip output
    // to the correct slot in the fan-in stage, regardless of wiring order.
    let idx0 = target.spec.inlets.iter().position(|i| i.id() == inlet0)?;
    let idx1 = target.spec.inlets.iter().position(|i| i.id() == inlet1)?;
    Some(UnzipFanInFastPath {
        fan_in_stage_index: stage0,
        target_inlet_indices: [idx0, idx1],
    })
}

fn unzip_zip_fast_path<S: Shape>(
    stage: &StageRecord,
    graph: &GraphBlueprint<S>,
    edge_by_outlet: &HashMap<PortId, PortId>,
    stage_by_inlet: &HashMap<PortId, usize>,
) -> Option<UnzipZipFastPath> {
    let outlets = &stage.spec.outlets;
    if outlets.len() != 2 {
        return None;
    }
    let inlet0 = edge_by_outlet.get(&outlets[0].id()).copied()?;
    let inlet1 = edge_by_outlet.get(&outlets[1].id()).copied()?;
    let stage0 = *stage_by_inlet.get(&inlet0)?;
    let stage1 = *stage_by_inlet.get(&inlet1)?;
    if stage0 != stage1 {
        return None;
    }
    let target = graph.stages.get(stage0)?;
    if !matches!(target.spec.kind, StageKind::Zip(_)) {
        return None;
    }
    Some(UnzipZipFastPath {
        zip_stage_index: stage0,
    })
}

fn partition_merge_fast_path<S: Shape>(
    stage: &StageRecord,
    graph: &GraphBlueprint<S>,
    edge_by_outlet: &HashMap<PortId, PortId>,
    stage_by_inlet: &HashMap<PortId, usize>,
) -> Option<PartitionMergeFastPath> {
    let outlets = &stage.spec.outlets;
    let first_inlet = edge_by_outlet.get(&outlets.first()?.id()).copied()?;
    let merge_stage_index = *stage_by_inlet.get(&first_inlet)?;
    let merge_stage = graph.stages.get(merge_stage_index)?;
    if !matches!(merge_stage.spec.kind, StageKind::Merge) {
        return None;
    }
    if merge_stage.spec.inlets.len() != outlets.len() {
        return None;
    }
    for (outlet, merge_inlet) in outlets.iter().zip(merge_stage.spec.inlets.iter()) {
        if edge_by_outlet.get(&outlet.id()).copied() != Some(merge_inlet.id()) {
            return None;
        }
    }
    Some(PartitionMergeFastPath { merge_stage_index })
}

/// Increments the event counter and returns an error if the configured limit
/// is exceeded.
///
/// `pub(crate)` so the typed-port executor (Phase 1+) can reuse the same
/// event-budget enforcement without duplicating the check.
pub(crate) fn bump_fused_event(
    events: &mut usize,
    config: FusedExecutionConfig,
) -> StreamResult<()> {
    *events += 1;
    if *events > config.event_limit {
        return Err(StreamError::EventLimitExceeded {
            limit: config.event_limit,
        });
    }
    Ok(())
}

fn broadcast_emissions(outlets: &[AnyOutlet], value: DatumValue) -> StreamResult<StageEmissions> {
    match outlets {
        [] => Err(StreamError::GraphValidation(
            "broadcast has no outlets".into(),
        )),
        [outlet] => Ok(StageEmissions::One(outlet.id(), value)),
        [first, second] => Ok(StageEmissions::Two(
            (first.id(), value.clone_box()),
            (second.id(), value),
        )),
        outlets => {
            let mut emitted = Vec::with_capacity(outlets.len());
            for outlet in &outlets[..outlets.len() - 1] {
                emitted.push((outlet.id(), value.clone_box()));
            }
            emitted.push((outlets[outlets.len() - 1].id(), value));
            Ok(StageEmissions::Many(emitted))
        }
    }
}

fn single_outlet(stage: &StageRecord) -> StreamResult<PortId> {
    stage
        .spec
        .outlets
        .first()
        .map(AnyOutlet::id)
        .ok_or_else(|| {
            StreamError::GraphValidation(format!("stage {} has no outlet", stage.spec.name()))
        })
}

fn merge_emissions(first: StageEmissions, second: StageEmissions) -> StageEmissions {
    match (first, second) {
        (StageEmissions::None, other) | (other, StageEmissions::None) => other,
        (StageEmissions::One(p1, v1), StageEmissions::One(p2, v2)) => {
            StageEmissions::Many(vec![(p1, v1), (p2, v2)])
        }
        (StageEmissions::One(p, v), StageEmissions::Many(mut vec))
        | (StageEmissions::Many(mut vec), StageEmissions::One(p, v)) => {
            vec.push((p, v));
            StageEmissions::Many(vec)
        }
        (StageEmissions::Many(mut v1), StageEmissions::Many(v2)) => {
            v1.extend(v2);
            StageEmissions::Many(v1)
        }
        (a, b) => {
            let mut all = Vec::new();
            push_emissions(&mut all, a);
            push_emissions(&mut all, b);
            StageEmissions::Many(all)
        }
    }
}

fn push_emissions(out: &mut Vec<(PortId, DatumValue)>, emissions: StageEmissions) {
    match emissions {
        StageEmissions::None => {}
        StageEmissions::One(p, v) => out.push((p, v)),
        StageEmissions::Two((p1, v1), (p2, v2)) => {
            out.push((p1, v1));
            out.push((p2, v2));
        }
        StageEmissions::Many(vec) => out.extend(vec),
    }
}