camel_processor/

aggregator.rs

use std::collections::HashMap;
use std::future::Future;
use std::pin::Pin;
use std::sync::{Arc, Mutex};
use std::task::{Context, Poll};
use std::time::{Duration, Instant};

use tokio::sync::mpsc;
use tokio::task::JoinHandle;
use tokio_util::sync::CancellationToken;
use tower::Service;

use async_trait::async_trait;
use camel_api::{
    CamelError, MetricsCollector, StepLifecycle, StepShutdownReason,
    aggregator::{
        AggregationStrategy, AggregatorConfig, CompletionCondition, CompletionMode,
        CompletionReason, CorrelationStrategy,
    },
    body::Body,
    exchange::Exchange,
    message::Message,
};
use camel_component_api::InFlightClaim;
use camel_language_api::Language;

pub type SharedLanguageRegistry = Arc<std::sync::Mutex<HashMap<String, Arc<dyn Language>>>>;

/// A bucket emission delivered outside the submitting call (timeout
/// fire, `force_complete_all`): the aggregated exchange plus every
/// stashed drainclaim it completes. The consumer of the late channel
/// MUST hold `claims` until the emission's continuation pipeline (or
/// reply) completes — dropping them IS the release.
pub struct AggregateEmission {
    pub exchange: Exchange,
    pub claims: Vec<Option<InFlightClaim>>,
}

/// Result of [`AggregatorService::submit_with_claim`].
///
/// Claim-propagation contract (drainclaim): on a pending stash the
/// claim stays inside the bucket; on sync completion the bucket's
/// claims (the submitter's included) are returned here so the caller
/// can hold them across the aggregated exchange's continuation; on
/// error the submitted claim was dropped inside (rejected = released).
pub struct AggregationReceipt {
    /// The pending-marked ack or the aggregated exchange — the same
    /// shapes the tower `Service::call` returns.
    pub reply: Result<Exchange, CamelError>,
    /// Claims released with this receipt. Empty for a pending stash
    /// (claims stay in the bucket) and for errors.
    pub claims: Vec<Option<InFlightClaim>>,
}

/// Sampling cadence for the metrics-only sweep (no `bucket_ttl` configured):
/// matches the 250ms cadence of the other dashboard-observability T3.3
/// queue-depth samplers.
const QUEUE_DEPTH_SAMPLE_INTERVAL: Duration = Duration::from_millis(250);

pub const CAMEL_AGGREGATOR_PENDING: &str = "CamelAggregatorPending";
pub const CAMEL_AGGREGATED_SIZE: &str = "CamelAggregatedSize";
pub const CAMEL_AGGREGATED_KEY: &str = "CamelAggregatedKey";
pub const CAMEL_AGGREGATED_COMPLETION_REASON: &str = "CamelAggregatedCompletionReason";

/// Internal bucket structure with timestamp tracking for TTL eviction.
///
/// drainclaim: `claims` runs parallel to `exchanges` (same length
/// invariant, maintained by [`Bucket::push`] and [`Bucket::into_parts`])
/// — one stashed claim per buffered exchange, moved in by the route
/// loop via [`AggregatorService::submit_with_claim`]. Whenever a bucket
/// leaves the map without emitting (TTL eviction, unarmed-bucket
/// release, discard-on-timeout), the dropped claims release.
struct Bucket {
    exchanges: Vec<Exchange>,
    claims: Vec<Option<InFlightClaim>>,
    last_updated: Instant,
}

impl Bucket {
    fn new() -> Self {
        Self {
            exchanges: Vec::new(),
            claims: Vec::new(),
            last_updated: Instant::now(),
        }
    }

    fn push(&mut self, exchange: Exchange, claim: Option<InFlightClaim>) {
        self.exchanges.push(exchange);
        self.claims.push(claim);
        self.last_updated = Instant::now();
    }

    /// Split the bucket into its exchanges and their stashed claims
    /// (same order, same length).
    fn into_parts(self) -> (Vec<Exchange>, Vec<Option<InFlightClaim>>) {
        (self.exchanges, self.claims)
    }

    fn is_expired(&self, ttl: Duration) -> bool {
        Instant::now().duration_since(self.last_updated) >= ttl
    }
}

#[derive(Clone)]
pub struct AggregatorService {
    config: AggregatorConfig,
    buckets: Arc<Mutex<HashMap<String, Bucket>>>,
    timeout_tasks: Arc<Mutex<HashMap<String, CancellationToken>>>,
    timeout_handles: Arc<Mutex<HashMap<String, JoinHandle<()>>>>,
    late_tx: mpsc::Sender<AggregateEmission>,
    language_registry: SharedLanguageRegistry,
    /// Swappable cell holding the cancellation token for the background
    /// maintenance sweep task. `StepLifecycle::start` replaces this with a fresh token
    /// so the sweep respawns on the next `poll_ready`; `shutdown` cancels the
    /// current token to terminate the task. The value is initially seeded from
    /// the route token — cancelling the route token cancels the sweep (the
    /// primary shutdown path). This replaces the plain `route_cancel` field.
    sweep_cancel: Arc<Mutex<CancellationToken>>,
    /// Handle to the background maintenance sweep task. `None` until the
    /// first `poll_ready`. The sweep exists when `bucket_ttl` OR
    /// `queue_metrics` is configured (TTL eviction and/or queue-depth
    /// sampling); with neither set there is nothing to sweep and no task is
    /// spawned. The task binds to the current `sweep_cancel` token via
    /// `select!`; `StepLifecycle::start` clears this to `None` (aborting the
    /// old task) so `poll_ready` respawns with the fresh token.
    sweep_handle: Arc<Mutex<Option<JoinHandle<()>>>>,
    /// Queue-depth reporting (collector + `aggregator:<route>` label). When
    /// set, the background maintenance sweep publishes the buffered group
    /// count as `camel_queue_depth{queue}` — with or without a
    /// `bucket_ttl`. `None` when no collector was injected at compile time.
    queue_metrics: Option<(Arc<dyn MetricsCollector>, String)>,
    /// Strong-count cell backing the last-owner check in [`Drop`]: transient
    /// clones (the pipeline clones the service per `poll_ready`) must not
    /// abort the shared sweep task; only the final owner's drop may.
    clone_guard: Arc<()>,
    /// Memoized correlation-key serialization: the last STRING key value
    /// paired with its `serde_json::to_string` form. Split-aggregate streams
    /// re-send the same constant key for every fragment; the memo turns
    /// per-fragment serialization into one string equality check. Shared via
    /// `Arc` (cloned into every `call` future) so all clones of the service
    /// share one memo slot. ONLY string keys are memoizable: their `Value`
    /// equality exactly tracks their serialization. Every other shape
    /// serializes per fragment (see `call` — e.g. float `0.0`/`-0.0` compare
    /// equal but serialize differently, so memoizing numbers could merge
    /// buckets that serde keeps distinct).
    cached_key: Arc<Mutex<Option<(serde_json::Value, String)>>>,
    /// Test-only count of correlation-key serializations actually performed
    /// (fast-path misses + per-fragment object bypass). Asserted by the memo
    /// tests; compiles out of non-test builds.
    #[cfg(test)]
    key_serializations: Arc<std::sync::atomic::AtomicUsize>,
}

impl std::fmt::Debug for AggregatorService {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.debug_struct("AggregatorService").finish_non_exhaustive()
    }
}

impl Drop for AggregatorService {
    fn drop(&mut self) {
        // Defense-in-depth: abort the background sweep on drop so it cannot
        // leak even if the route owner forgets to call `shutdown`. The
        // primary shutdown path is now `StepLifecycle::shutdown` which
        // cancels `sweep_cancel` and aborts `sweep_handle`. `abort()` on an
        // already-finished task is a no-op.
        //
        // Last-owner guard: `Clone` shares this cell, so a transient clone
        // dropping (the pipeline clones the service on every `poll_ready`)
        // does NOT abort the sweep — only the final owner's drop does.
        if Arc::strong_count(&self.clone_guard) > 1 {
            return;
        }
        if let Some(handle) = self
            .sweep_handle
            .lock()
            .unwrap_or_else(|e| e.into_inner())
            .take()
        {
            handle.abort();
        }
    }
}

impl AggregatorService {
    /// Lifecycle invariant: `sweep_cancel` is initially seeded from the
    /// route's cancellation token and wrapped in a swappable `Arc<Mutex<...>>`
    /// cell. `StepLifecycle::start` replaces it with a fresh token so the
    /// sweep respawns after a restart. Construction is runtime-free (no
    /// `tokio::spawn` here). The maintenance sweep task (TTL eviction when
    /// `bucket_ttl` is set, queue-depth sampling when `queue_metrics` is
    /// set) is spawned LAZILY on the first `poll_ready` and bound to the
    /// current `sweep_cancel` token via `select!`.
    /// The route owner MUST call `shutdown` to cancel it;
    /// `Drop` also aborts it as defense-in-depth.
    pub fn new(
        config: AggregatorConfig,
        late_tx: mpsc::Sender<AggregateEmission>,
        language_registry: SharedLanguageRegistry,
        route_cancel: CancellationToken,
    ) -> Self {
        // R3-M2: at least one memory-release bound is mandatory.
        config.validate().expect(
            // allow-unwrap
            // fail-closed startup invariant (ADR-0033); a config without a bound is a programmer/operator error.
            "AggregatorService::new: config failed validation \
             (need max_buckets, completionTimeout, or bucket_ttl)",
        );

        // R3-M2 advisory: Size/Predicate-only completion (no Timeout) with no
        // bucket_ttl means buckets accumulate until max_buckets is hit. Bounded
        // by max_buckets (validated above) but worth surfacing to operators.
        let has_timeout = has_timeout_condition(&config.completion);
        if !has_timeout && config.bucket_ttl.is_none() {
            tracing::warn!(
                "Aggregator configured with Size/Predicate completion and no \
                 bucket_ttl: buckets accumulate until max_buckets is reached"
            );
        }

        // Build the shared buckets map up front so the sweep task can
        // share it via Arc::clone.
        let buckets: Arc<Mutex<HashMap<String, Bucket>>> = Arc::new(Mutex::new(HashMap::new()));

        // R3-C1 Batch 1: the TTL-sweep is NOT spawned here — construction
        // must stay runtime-free so callers that build an AggregatorService
        // outside a tokio runtime (route-spec tests) do not panic on
        // `tokio::spawn`. The sweep is spawned lazily on the first
        // `poll_ready`. The inline `guard.retain(...)` per `call` evicts
        // expired buckets regardless, so the cap + TTL invariants hold
        // whether or not the sweep has started yet.

        Self {
            config,
            buckets,
            timeout_tasks: Arc::new(Mutex::new(HashMap::new())),
            timeout_handles: Arc::new(Mutex::new(HashMap::new())),
            late_tx,
            language_registry,
            sweep_cancel: Arc::new(Mutex::new(route_cancel)),
            sweep_handle: Arc::new(Mutex::new(None)),
            queue_metrics: None,
            clone_guard: Arc::new(()),
            cached_key: Arc::new(Mutex::new(None)),
            #[cfg(test)]
            key_serializations: Arc::new(std::sync::atomic::AtomicUsize::new(0)),
        }
    }

    /// Inject queue-depth reporting: the TTL-sweep maintenance pass publishes
    /// the buffered group count (open correlation buckets) as
    /// `camel_queue_depth{queue = label}`. `label` is the route-scoped
    /// closed-set identifier (`aggregator:<route>`).
    pub fn with_queue_metrics(
        mut self,
        metrics: Arc<dyn MetricsCollector>,
        label: impl Into<String>,
    ) -> Self {
        self.queue_metrics = Some((metrics, label.into()));
        self
    }

    pub fn config(&self) -> &AggregatorConfig {
        &self.config
    }

    pub fn has_timeout(&self) -> bool {
        has_timeout_condition(&self.config.completion)
    }

    pub fn force_complete_all(&self) {
        let mut buckets_guard = self.buckets.lock().unwrap_or_else(|e| e.into_inner());
        let keys: Vec<String> = buckets_guard.keys().cloned().collect();

        for key in keys {
            if let Some(bucket) = buckets_guard.remove(&key) {
                if self.config.force_completion_on_stop {
                    cancel_timeout_task_with_handle(
                        &key,
                        &self.timeout_tasks,
                        &self.timeout_handles,
                    );
                    let (exchanges, claims) = bucket.into_parts();
                    match aggregate(exchanges, &self.config.strategy) {
                        Ok(mut result) => {
                            result.set_property(
                                CAMEL_AGGREGATED_COMPLETION_REASON,
                                serde_json::json!(CompletionReason::Stop.as_str()),
                            );
                            if self
                                .late_tx
                                .try_send(AggregateEmission {
                                    exchange: result,
                                    claims,
                                })
                                .is_err()
                            {
                                tracing::warn!(
                                    key = %key,
                                    "aggregator force-complete emit dropped: late channel full"
                                );
                            }
                        }
                        Err(e) => {
                            // log-policy: handler-owned
                            tracing::warn!(
                                key = %key,
                                error = %e,
                                "aggregation failed in force_complete_all"
                            );
                        }
                    }
                } else {
                    cancel_timeout_task_with_handle(
                        &key,
                        &self.timeout_tasks,
                        &self.timeout_handles,
                    );
                }
            }
        }
    }

    /// Consumer-exit release (bd rc-iioeq): discard buckets with NO armed
    /// timeout task, leave armed buckets untouched.
    ///
    /// After a natural consumer exit no further exchanges arrive, so a
    /// bucket without an armed timeout task can never complete on its
    /// own — its buffered exchanges are released eagerly (the same
    /// discard semantics `force_complete_all` applies with
    /// `force_completion_on_stop=false`). A bucket WITH an armed timeout
    /// task is left alone: that task owns its completion and emits
    /// through the late channel when the timeout fires.
    ///
    /// Arming is decided per bucket in `call`: when the
    /// `max_timeout_tasks` cap is reached, a timeout-configured bucket
    /// stays unarmed. Without this release such a bucket would be
    /// orphaned forever — the `bucket_ttl` sweep only runs inside `call`
    /// on the next exchange, which never arrives after consumer exit.
    pub fn release_unarmed_buckets(&self) {
        let mut buckets_guard = self.buckets.lock().unwrap_or_else(|e| e.into_inner());
        let timeout_guard = self.timeout_tasks.lock().unwrap_or_else(|e| e.into_inner());
        let before = buckets_guard.len();
        buckets_guard.retain(|key, _| timeout_guard.contains_key(key));
        let released = before - buckets_guard.len();
        if released > 0 {
            tracing::debug!(
                released,
                armed = buckets_guard.len(),
                "aggregator released unarmed buckets on consumer exit"
            );
        }
    }

    /// Graceful shutdown: cancel all outstanding timeout tasks and await their
    /// JoinHandles (with a 5s deadline) so that no tasks are leaked.
    pub(crate) async fn shutdown_inner(&self) {
        // Cancel all timeout cancellation tokens.
        {
            let mut guard = self.timeout_tasks.lock().unwrap_or_else(|e| e.into_inner());
            for token in guard.values() {
                token.cancel();
            }
            guard.clear();
        };

        // Remove and collect all JoinHandles.
        let handles: Vec<JoinHandle<()>> = {
            let mut guard = self
                .timeout_handles
                .lock()
                .unwrap_or_else(|e| e.into_inner());
            guard.drain().map(|(_, handle)| handle).collect()
        };

        if handles.is_empty() {
            return;
        }

        // Await all handles with a deadline.
        let _ = tokio::time::timeout(Duration::from_secs(5), async {
            for handle in handles {
                let _ = handle.await;
            }
        })
        .await;
    }
}

#[async_trait]
impl StepLifecycle for AggregatorService {
    fn name(&self) -> &'static str {
        "aggregator"
    }

    async fn shutdown(&self, reason: StepShutdownReason) -> Result<(), CamelError> {
        tracing::debug!(reason = ?reason, "Aggregator shutdown via StepLifecycle");

        // Cancel the sweep token so the background task's `select!` cancel
        // branch fires and the task exits.
        self.sweep_cancel
            .lock()
            .unwrap_or_else(|e| e.into_inner())
            .cancel();

        // Abort any running sweep handle (defense-in-depth alongside the
        // token cancel above).
        if let Some(handle) = self
            .sweep_handle
            .lock()
            .unwrap_or_else(|e| e.into_inner())
            .take()
        {
            handle.abort();
        }

        self.shutdown_inner().await;
        Ok(())
    }

    /// Resets the sweep for a fresh lifecycle: replaces the cancellation token
    /// with a new one and aborts any existing sweep handle so the next
    /// `poll_ready` respawns the sweep bound to the new token.
    async fn start(&self) -> Result<(), CamelError> {
        *self.sweep_cancel.lock().unwrap_or_else(|e| e.into_inner()) = CancellationToken::new();

        if let Some(handle) = self
            .sweep_handle
            .lock()
            .unwrap_or_else(|e| e.into_inner())
            .take()
        {
            handle.abort();
        }

        Ok(())
    }
}

pub fn has_timeout_condition(mode: &CompletionMode) -> bool {
    match mode {
        CompletionMode::Single(CompletionCondition::Timeout(_)) => true,
        CompletionMode::Any(conditions) => conditions
            .iter()
            .any(|c| matches!(c, CompletionCondition::Timeout(_))),
        _ => false,
    }
}

impl Service<Exchange> for AggregatorService {
    type Response = Exchange;
    type Error = CamelError;
    type Future = Pin<Box<dyn Future<Output = Result<Exchange, CamelError>> + Send>>;

    fn poll_ready(&mut self, _cx: &mut Context<'_>) -> Poll<Result<(), CamelError>> {
        // R3-C1: lazily spawn the maintenance sweep on the first readiness
        // poll. A tokio runtime is guaranteed here (the runtime driving the
        // service), unlike at construction. Single-spawn via the lock +
        // is_none check. The sweep exists when EITHER release mechanism is
        // configured: `bucket_ttl` drives eviction, `queue_metrics` drives
        // queue-depth sampling — a size-only aggregator (bucket_ttl = None)
        // still samples (T3.3 review F2); eviction itself stays gated on
        // the TTL.
        let ttl = self.config.bucket_ttl;
        if ttl.is_none() && self.queue_metrics.is_none() {
            return Poll::Ready(Ok(()));
        }
        let mut g = self.sweep_handle.lock().unwrap_or_else(|e| e.into_inner());
        if g.is_none() {
            let interval = match ttl {
                Some(ttl) => std::cmp::max(ttl / 2, Duration::from_millis(50)),
                None => QUEUE_DEPTH_SAMPLE_INTERVAL,
            };
            let buckets = Arc::clone(&self.buckets);
            let cancel = self
                .sweep_cancel
                .lock()
                .unwrap_or_else(|e| e.into_inner())
                .clone();
            let queue_metrics = self.queue_metrics.clone();
            *g = Some(tokio::spawn(async move {
                loop {
                    tokio::select! {
                        _ = cancel.cancelled() => break,
                        _ = tokio::time::sleep(interval) => {
                            let mut guard =
                                buckets.lock().unwrap_or_else(|e| e.into_inner());
                            if let Some(ttl) = ttl {
                                guard.retain(|_, b| !b.is_expired(ttl));
                            }
                            // Maintenance-pass sampling: publish the
                            // buffered group count after eviction.
                            if let Some((metrics, label)) = queue_metrics.as_ref() {
                                metrics.set_queue_depth(label, guard.len());
                            }
                        }
                    }
                }
            }));
        }
        Poll::Ready(Ok(()))
    }

    fn call(&mut self, exchange: Exchange) -> Self::Future {
        let svc = self.clone();
        Box::pin(async move {
            // claimfamily (rc-qbigm): pipeline-embedded submissions carry
            // their drain-site split sibling ON the exchange — take it as
            // the stash sidecar so a pending bucket keeps the exchange
            // counted after the embedding pipeline's ack resolves. A sync
            // completion re-attaches ONE of the bucket's claims to the
            // aggregated output (its downstream continuation stays
            // counted; the drain site takes it back if the output
            // completes in-band); the remaining claims drop — their
            // input exchanges were consumed into the aggregate. Every
            // rejection path drops the submitted claim inside
            // (rejected = released).
            let mut exchange = exchange;
            let claim = exchange.in_flight_claim.take();
            let AggregationReceipt { reply, claims } = svc.submit_with_claim(exchange, claim).await;
            match reply {
                Ok(mut result) => {
                    result.in_flight_claim = claims.into_iter().flatten().next();
                    Ok(result)
                }
                Err(e) => Err(e),
            }
        })
    }
}

impl AggregatorService {
    /// Submit one exchange with its drainclaim sidecar (drainclaim
    /// claim propagation) — the route loop's aggregate entry point.
    ///
    /// Behaviorally identical to the tower `Service::call` this is
    /// extracted from, except the claim travels with the exchange: a
    /// pending stash moves it into the bucket, a sync completion
    /// returns it (with the rest of the bucket's claims) to the caller
    /// to hold across the aggregated exchange's continuation, and every
    /// rejection path (bad correlation key, bucket caps) drops it —
    /// rejected means released.
    pub async fn submit_with_claim(
        &self,
        exchange: Exchange,
        claim: Option<InFlightClaim>,
    ) -> AggregationReceipt {
        let config = self.config.clone();
        let buckets = Arc::clone(&self.buckets);
        let timeout_tasks = Arc::clone(&self.timeout_tasks);
        let timeout_handles = Arc::clone(&self.timeout_handles);
        let late_tx = self.late_tx.clone();
        let language_registry = Arc::clone(&self.language_registry);
        let cached_key = Arc::clone(&self.cached_key);
        #[cfg(test)]
        let key_serializations = Arc::clone(&self.key_serializations);

        let inner = async move {
            let key_value =
                extract_correlation_key(&exchange, &config.correlation, &language_registry).await?;

            // Constant-key fast path: STRING keys only are memoized — one
            // string equality check replaces per-fragment serde, and string
            // equality exactly tracks its `serde_json::to_string` output.
            // Every other shape serializes per fragment exactly as before:
            // `Value` equality can diverge from serialization (float
            // `0.0` and `-0.0` compare equal but serialize as "0.0" and
            // "-0.0"; object/array `Value` equality is key-order-insensitive
            // while output is not), so memoizing them could merge buckets
            // that serde keeps distinct. Bool/Null serialize to a constant,
            // so their per-fragment serde cost is trivial.
            let key_is_memoizable = matches!(key_value, serde_json::Value::String(_));
            let key_str = if key_is_memoizable {
                let mut memo = cached_key.lock().unwrap_or_else(|e| e.into_inner());
                if let Some((v, s)) = memo.as_ref()
                    && *v == key_value
                {
                    s.clone()
                } else {
                    #[cfg(test)]
                    key_serializations.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
                    let s = serde_json::to_string(&key_value)
                        .map_err(|e| CamelError::ProcessorError(e.to_string()))?;
                    *memo = Some((key_value.clone(), s.clone()));
                    s
                }
            } else {
                #[cfg(test)]
                key_serializations.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
                serde_json::to_string(&key_value)
                    .map_err(|e| CamelError::ProcessorError(e.to_string()))?
            };

            // Pre-evaluate the completion predicate (if any) BEFORE taking the
            // buckets lock — expression evaluation is async + uses the language
            // registry, which cannot run under the !Send buckets MutexGuard.
            let predicate_satisfied =
                evaluate_completion_predicate(&config.completion, &exchange, &language_registry)
                    .await?;

            let completed_bucket = {
                let mut guard = buckets.lock().unwrap_or_else(|e| e.into_inner());

                if let Some(ttl) = config.bucket_ttl {
                    guard.retain(|_, bucket| !bucket.is_expired(ttl));
                }

                if let Some(max) = config.max_buckets
                    && !guard.contains_key(&key_str)
                    && guard.len() >= max
                {
                    tracing::warn!(
                        max_buckets = max,
                        correlation_key = %key_str,
                        "Aggregator reached max buckets limit, rejecting new correlation key"
                    );
                    return Err(CamelError::ProcessorError(format!(
                        "Aggregator reached maximum {} buckets",
                        max
                    )));
                }

                // F6-2 (audit 2026-08-31): per-bucket accumulation bound. A
                // single hot correlation key must not buffer unboundedly while
                // waiting for predicate/timeout completion.
                if let Some(max_size) = config.max_bucket_size
                    && let Some(existing) = guard.get(&key_str)
                    && existing.exchanges.len() >= max_size
                {
                    tracing::warn!(
                        max_bucket_size = max_size,
                        correlation_key = %key_str,
                        "Aggregator reached per-bucket size limit, rejecting exchange"
                    );
                    return Err(CamelError::ProcessorError(format!(
                        "Aggregator bucket reached maximum {max_size} exchanges"
                    )));
                }

                let bucket = match guard.get_mut(&key_str) {
                    Some(b) => b,
                    None => guard.entry(key_str.clone()).or_insert_with(Bucket::new),
                };
                bucket.push(exchange, claim);

                let (is_complete, reason) = check_sync_completion(
                    &config.completion,
                    &bucket.exchanges,
                    predicate_satisfied,
                );

                let bucket_parts = if is_complete {
                    guard.remove(&key_str).map(Bucket::into_parts)
                } else {
                    None
                };
                (bucket_parts, reason)
            };

            if completed_bucket.0.is_none() && has_timeout_condition(&config.completion) {
                let timeout_dur = extract_timeout_duration(&config.completion);
                if let Some(timeout) = timeout_dur {
                    // R3-M3: bound the number of concurrently-live per-bucket
                    // timeout tasks. When the cap is reached, skip the dedicated
                    // spawn — the bucket relies on bucket_ttl eviction (graceful
                    // degradation). max_buckets already caps total buckets, so
                    // memory stays bounded regardless.
                    let live_count = timeout_handles
                        .lock()
                        .unwrap_or_else(|e| e.into_inner())
                        .len();
                    if live_count >= config.max_timeout_tasks {
                        tracing::warn!(
                            live_timeout_tasks = live_count,
                            max_timeout_tasks = config.max_timeout_tasks,
                            correlation_key = %key_str,
                            "Aggregator timeout-task cap reached; bucket will rely on \
                             bucket_ttl eviction instead of a dedicated timeout task"
                        );
                    } else {
                        // Cancel old token for this key (if any).
                        {
                            let tt_guard = timeout_tasks.lock().unwrap_or_else(|e| e.into_inner());
                            if let Some(existing) = tt_guard.get(&key_str) {
                                existing.cancel();
                            }
                        }
                        // Remove old handle if present.
                        {
                            let mut hh = timeout_handles.lock().unwrap_or_else(|e| e.into_inner());
                            if let Some(old) = hh.remove(&key_str) {
                                old.abort();
                            }
                        }
                        let cancel = CancellationToken::new();
                        timeout_tasks
                            .lock()
                            .unwrap_or_else(|e| e.into_inner())
                            .insert(key_str.clone(), cancel.clone());
                        let handle = spawn_timeout_task(
                            key_str.clone(),
                            timeout,
                            cancel,
                            buckets.clone(),
                            timeout_tasks.clone(),
                            timeout_handles.clone(),
                            late_tx,
                            config.strategy.clone(),
                            config.discard_on_timeout,
                        );
                        timeout_handles
                            .lock()
                            .unwrap_or_else(|e| e.into_inner())
                            .insert(key_str.clone(), handle);
                    }
                }
            }

            if let Some((exchanges, claims)) = completed_bucket.0 {
                cancel_timeout_task_with_handle(&key_str, &timeout_tasks, &timeout_handles);
                let reason = completed_bucket.1;
                let size = exchanges.len();
                let mut result = aggregate(exchanges, &config.strategy)?;
                result.set_property(CAMEL_AGGREGATED_SIZE, serde_json::json!(size as u64));
                result.set_property(CAMEL_AGGREGATED_KEY, key_value);
                result.set_property(
                    CAMEL_AGGREGATED_COMPLETION_REASON,
                    serde_json::json!(reason.as_str()),
                );
                Ok((result, claims))
            } else {
                let mut pending = Exchange::new(Message {
                    headers: Default::default(),
                    body: Body::Empty,
                });
                pending.set_property(CAMEL_AGGREGATOR_PENDING, serde_json::json!(true));
                // Claims stay stashed in the bucket.
                Ok((pending, Vec::new()))
            }
        }
        .await;

        match inner {
            Ok((exchange, claims)) => AggregationReceipt {
                reply: Ok(exchange),
                claims,
            },
            // The submitted claim was dropped inside the body — rejected
            // means released, so nothing to hand back.
            Err(e) => AggregationReceipt {
                reply: Err(e),
                claims: Vec::new(),
            },
        }
    }
}

async fn extract_correlation_key(
    exchange: &Exchange,
    strategy: &CorrelationStrategy,
    registry: &SharedLanguageRegistry,
) -> Result<serde_json::Value, CamelError> {
    match strategy {
        CorrelationStrategy::HeaderName(h) => {
            exchange.input.headers.get(h).cloned().ok_or_else(|| {
                CamelError::ProcessorError(format!(
                    "Aggregator: missing correlation key header '{}'",
                    h
                ))
            })
        }
        CorrelationStrategy::Expression { expr, language } => {
            let expression = {
                let reg = registry.lock().unwrap_or_else(|e| e.into_inner());
                let lang = reg.get(language).ok_or_else(|| {
                    CamelError::ProcessorError(format!(
                        "Aggregator: language '{}' not found in registry",
                        language
                    ))
                })?;
                lang.create_expression(expr)
                    .map_err(|e| CamelError::ProcessorError(e.to_string()))?
            };
            let value = expression
                .evaluate(exchange)
                .await
                .map_err(|e| CamelError::ProcessorError(e.to_string()))?;
            if value.is_null() {
                return Err(CamelError::ProcessorError(format!(
                    "Aggregator: correlation expression '{}' evaluated to null",
                    expr
                )));
            }
            Ok(value)
        }
        CorrelationStrategy::Fn(f) => f(exchange).map(serde_json::Value::String).ok_or_else(|| {
            CamelError::ProcessorError("Aggregator: correlation function returned None".to_string())
        }),
        // Future correlation strategies are unsupported here.
        _ => Err(CamelError::ProcessorError(
            "Aggregator: unsupported correlation strategy".to_string(),
        )),
    }
}

/// Yield `(&expr, &language)` for every `PredicateExpr` condition in `mode`.
fn iter_predicate_exprs(mode: &CompletionMode) -> Vec<(&String, &String)> {
    let mut out = Vec::new();
    match mode {
        CompletionMode::Single(CompletionCondition::PredicateExpr { expr, language }) => {
            out.push((expr, language));
        }
        CompletionMode::Any(conds) => {
            for c in conds {
                if let CompletionCondition::PredicateExpr { expr, language } = c {
                    out.push((expr, language));
                }
            }
        }
        _ => {}
    }
    out
}

/// Pre-evaluate ALL `PredicateExpr` conditions against the incoming exchange,
/// OR-combining (matches `CompletionMode::Any` semantics). MUST be called BEFORE
/// the `buckets` lock — it awaits the language registry and cannot run under the
/// `!Send` `buckets` MutexGuard. Mirrors `extract_correlation_key`.
async fn evaluate_completion_predicate(
    mode: &CompletionMode,
    incoming: &Exchange,
    registry: &SharedLanguageRegistry,
) -> Result<bool, CamelError> {
    let mut satisfied = false;
    for (expr, language) in iter_predicate_exprs(mode) {
        let predicate = {
            let reg = registry.lock().unwrap_or_else(|e| e.into_inner());
            let lang = reg.get(language).ok_or_else(|| {
                CamelError::ProcessorError(format!(
                    "Aggregator: language '{}' not found in registry",
                    language
                ))
            })?;
            // Predicate grammar, not template mode: completion conditions are
            // strict booleans. A non-bool result is a `LanguageError`
            // (type mismatch) surfaced as `ProcessorError`.
            lang.create_predicate(expr)
                .map_err(|e| CamelError::ProcessorError(e.to_string()))?
        }; // registry guard dropped here — safe to await below
        let matched = predicate
            .matches(incoming)
            .await
            .map_err(|e| CamelError::ProcessorError(e.to_string()))?;
        if matched {
            satisfied = true;
        }
    }
    Ok(satisfied)
}

fn check_sync_completion(
    mode: &CompletionMode,
    exchanges: &[Exchange],
    predicate_satisfied: bool,
) -> (bool, CompletionReason) {
    match mode {
        CompletionMode::Single(cond) => check_single(cond, exchanges, predicate_satisfied),
        CompletionMode::Any(conditions) => {
            for cond in conditions {
                if let CompletionCondition::Timeout(_) = cond {
                    continue;
                }
                let (done, reason) = check_single(cond, exchanges, predicate_satisfied);
                if done {
                    return (true, reason);
                }
            }
            (false, CompletionReason::Size)
        }
        // Future completion modes are not synchronously complete.
        _ => (false, CompletionReason::Size),
    }
}

fn check_single(
    cond: &CompletionCondition,
    exchanges: &[Exchange],
    predicate_satisfied: bool,
) -> (bool, CompletionReason) {
    match cond {
        CompletionCondition::Size(n) => (exchanges.len() >= *n, CompletionReason::Size),
        CompletionCondition::Predicate(pred) => (pred(exchanges), CompletionReason::Predicate),
        CompletionCondition::PredicateExpr { .. } => {
            (predicate_satisfied, CompletionReason::Predicate)
        }
        CompletionCondition::Timeout(_) => (false, CompletionReason::Timeout),
        // Future condition types cannot be evaluated synchronously.
        _ => (false, CompletionReason::Size),
    }
}

fn extract_timeout_duration(mode: &CompletionMode) -> Option<Duration> {
    match mode {
        CompletionMode::Single(CompletionCondition::Timeout(d)) => Some(*d),
        CompletionMode::Any(conditions) => conditions.iter().find_map(|c| {
            if let CompletionCondition::Timeout(d) = c {
                Some(*d)
            } else {
                None
            }
        }),
        _ => None,
    }
}

fn cancel_timeout_task(key: &str, timeout_tasks: &Arc<Mutex<HashMap<String, CancellationToken>>>) {
    let mut guard = timeout_tasks.lock().unwrap_or_else(|e| e.into_inner());
    if let Some(token) = guard.remove(key) {
        token.cancel();
    }
}

/// Also removes the stored JoinHandle for a cancelled/completed timeout task.
fn cancel_timeout_task_with_handle(
    key: &str,
    timeout_tasks: &Arc<Mutex<HashMap<String, CancellationToken>>>,
    timeout_handles: &Arc<Mutex<HashMap<String, JoinHandle<()>>>>,
) {
    cancel_timeout_task(key, timeout_tasks);
    let mut guard = timeout_handles.lock().unwrap_or_else(|e| e.into_inner());
    guard.remove(key);
}

#[allow(clippy::too_many_arguments)]
fn spawn_timeout_task(
    key: String,
    timeout: Duration,
    cancel: CancellationToken,
    buckets: Arc<Mutex<HashMap<String, Bucket>>>,
    timeout_tasks: Arc<Mutex<HashMap<String, CancellationToken>>>,
    timeout_handles: Arc<Mutex<HashMap<String, JoinHandle<()>>>>,
    late_tx: mpsc::Sender<AggregateEmission>,
    strategy: AggregationStrategy,
    discard: bool,
) -> JoinHandle<()> {
    let cancel_clone = cancel.clone();
    tokio::spawn(async move {
        tokio::select! {
            _ = tokio::time::sleep(timeout) => {
                let should_proceed = {
                    let mut tt_guard = timeout_tasks.lock().unwrap_or_else(|e| e.into_inner());
                    if cancel_clone.is_cancelled() {
                        false
                    } else {
                        tt_guard.remove(&key);
                        true
                    }
                };
                if !should_proceed {
                    return;
                }
                // Clean up our own JoinHandle from the map on natural completion.
                // Without this, the handle entry leaks until route shutdown.
                {
                    let mut hh = timeout_handles.lock().unwrap_or_else(|e| e.into_inner());
                    hh.remove(&key);
                }
                let bucket_parts = {
                    let mut guard = buckets.lock().unwrap_or_else(|e| e.into_inner());
                    guard.remove(&key).map(Bucket::into_parts)
                };
                if let Some((exchanges, claims)) = bucket_parts
                    && !discard
                {
                    match aggregate(exchanges, &strategy) {
                        Ok(mut result) => {
                            result.set_property(
                                CAMEL_AGGREGATED_COMPLETION_REASON,
                                serde_json::json!(CompletionReason::Timeout.as_str()),
                            );
                            if late_tx
                                .try_send(AggregateEmission {
                                    exchange: result,
                                    claims,
                                })
                                .is_err()
                            {
                                tracing::warn!(
                                    key = %key,
                                    "aggregator timeout emit dropped: late channel full"
                                );
                            }
                        }
                        Err(e) => {
                            // log-policy: handler-owned
                            tracing::warn!(
                                key = %key,
                                error = %e,
                                "aggregation failed in timeout task"
                            );
                        }
                    }
                }
            }
            _ = cancel_clone.cancelled() => {}
        }
    })
}

fn aggregate(
    exchanges: Vec<Exchange>,
    strategy: &AggregationStrategy,
) -> Result<Exchange, CamelError> {
    match strategy {
        AggregationStrategy::CollectAll => {
            let bodies: Vec<serde_json::Value> = exchanges
                .into_iter()
                .map(|e| match e.input.body {
                    Body::Json(v) => v,
                    Body::Text(s) => serde_json::Value::String(s),
                    Body::Xml(s) => serde_json::Value::String(s),
                    Body::Bytes(b) => {
                        serde_json::Value::String(String::from_utf8_lossy(&b).into_owned())
                    }
                    Body::Stream(s) => serde_json::json!({
                        "_stream": {
                            "origin": s.metadata.origin,
                            "placeholder": true,
                            "hint": "Materialize exchange body with .into_bytes() before aggregation if content needed"
                        }
                    }),
                    // Empty and future variants contribute no extractable value.
                    _ => serde_json::Value::Null,
                })
                .collect();
            Ok(Exchange::new(Message {
                headers: Default::default(),
                body: Body::Json(serde_json::Value::Array(bodies)),
            }))
        }
        AggregationStrategy::Custom(f) => {
            let mut iter = exchanges.into_iter();
            let first = iter.next().ok_or_else(|| {
                CamelError::ProcessorError("Aggregator: empty bucket".to_string())
            })?;
            Ok(iter.fold(first, |acc, next| f(acc, next)))
        }
        // Future aggregation strategies are unsupported here.
        _ => Err(CamelError::ProcessorError(
            "Aggregator: unsupported aggregation strategy".to_string(),
        )),
    }
}

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

    use camel_api::{
        StepLifecycle, StepShutdownReason,
        aggregator::{AggregationStrategy, AggregatorConfig},
        body::Body,
        exchange::Exchange,
        message::Message,
    };
    use tokio::sync::mpsc;
    use tokio_util::sync::CancellationToken;
    use tower::ServiceExt;

    fn make_exchange(header: &str, value: &str, body: &str) -> Exchange {
        let mut msg = Message {
            headers: Default::default(),
            body: Body::Text(body.to_string()),
        };
        msg.headers
            .insert(header.to_string(), serde_json::json!(value));
        Exchange::new(msg)
    }

    fn config_size(n: usize) -> AggregatorConfig {
        AggregatorConfig::correlate_by("orderId")
            .complete_when_size(n)
            .build()
            .unwrap()
    }

    fn new_test_svc(config: AggregatorConfig) -> AggregatorService {
        let (tx, _rx) = mpsc::channel(256);
        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        let cancel = CancellationToken::new();
        AggregatorService::new(config, tx, registry, cancel)
    }

    /// Same as `new_test_svc` but lets the caller install a pre-populated
    /// language registry. Used by `CompletionCondition::PredicateExpr` tests
    /// where the simple language must be resolvable at evaluation time.
    fn new_test_svc_with_registry(
        config: AggregatorConfig,
        registry: SharedLanguageRegistry,
    ) -> AggregatorService {
        let (tx, _rx) = mpsc::channel(256);
        let cancel = CancellationToken::new();
        AggregatorService::new(config, tx, registry, cancel)
    }

    #[tokio::test]
    async fn test_pending_exchange_not_yet_complete() {
        let mut svc = new_test_svc(config_size(3));
        let ex = make_exchange("orderId", "A", "first");
        let result = svc.ready().await.unwrap().call(ex).await.unwrap();
        assert!(matches!(result.input.body, Body::Empty));
        assert_eq!(
            result.property(CAMEL_AGGREGATOR_PENDING),
            Some(&serde_json::json!(true))
        );
    }

    // ── claimfamily (rc-qbigm): embedded (size-only) stash claims ──

    #[tokio::test]
    async fn embedded_stash_keeps_carried_claim_until_completion() {
        let counter = Arc::new(camel_api::InFlightGauge::new());
        let svc = new_test_svc(config_size(2));

        // First exchange: pending stash — the carried claim moves into
        // the bucket and keeps the exchange counted after `call`
        // (the embedding pipeline's ack) resolves.
        let mut first = make_exchange("orderId", "A", "first");
        first.in_flight_claim = Some(InFlightClaim::attach(&counter));
        let pending = svc.clone().oneshot(first).await.unwrap();
        assert_eq!(
            pending.property(CAMEL_AGGREGATOR_PENDING),
            Some(&serde_json::json!(true)),
            "first exchange must be stashed"
        );
        assert!(
            pending.in_flight_claim.is_none(),
            "the pending marker carries no claim (it completes immediately)"
        );
        assert_eq!(counter.total(), 1, "stashed exchange must stay counted");

        // Second exchange completes the bucket: the aggregated output
        // carries ONE claim for its downstream continuation; the other
        // claim dropped (that input was consumed into the aggregate).
        let mut second = make_exchange("orderId", "A", "second");
        second.in_flight_claim = Some(InFlightClaim::attach(&counter));
        let aggregated = svc.clone().oneshot(second).await.unwrap();
        assert!(
            aggregated.property(CAMEL_AGGREGATOR_PENDING).is_none(),
            "second exchange must complete the bucket"
        );
        assert!(
            aggregated.in_flight_claim.is_some(),
            "aggregated output carries one claim downstream"
        );
        assert_eq!(
            counter.total(),
            1,
            "consumed input's claim released; the output's stays held"
        );
        drop(aggregated);
        assert_eq!(
            counter.total(),
            0,
            "dropping the completed output releases the last claim"
        );
    }

    #[tokio::test]
    async fn embedded_rejection_releases_carried_claim() {
        let counter = Arc::new(camel_api::InFlightGauge::new());
        let svc = new_test_svc(config_size(2));

        // Missing correlation header: the submission is rejected and the
        // carried claim drops inside — rejected = released.
        let mut bad = Exchange::new(Message::new("no-header"));
        bad.in_flight_claim = Some(InFlightClaim::attach(&counter));
        let result = svc.clone().oneshot(bad).await;
        assert!(result.is_err(), "missing correlation key must reject");
        assert_eq!(counter.total(), 0, "rejected submission releases its claim");
    }

    #[tokio::test]
    async fn test_completes_on_size() {
        let mut svc = new_test_svc(config_size(3));
        for _ in 0..2 {
            let ex = make_exchange("orderId", "A", "item");
            let r = svc.ready().await.unwrap().call(ex).await.unwrap();
            assert!(matches!(r.input.body, Body::Empty));
        }
        let ex = make_exchange("orderId", "A", "last");
        let result = svc.ready().await.unwrap().call(ex).await.unwrap();
        assert!(result.property(CAMEL_AGGREGATOR_PENDING).is_none());
        assert_eq!(
            result.property(CAMEL_AGGREGATED_SIZE),
            Some(&serde_json::json!(3u64))
        );
    }

    #[tokio::test]
    async fn test_collect_all_produces_json_array() {
        let mut svc = new_test_svc(config_size(2));
        svc.ready()
            .await
            .unwrap()
            .call(make_exchange("orderId", "A", "alpha"))
            .await
            .unwrap();
        let result = svc
            .ready()
            .await
            .unwrap()
            .call(make_exchange("orderId", "A", "beta"))
            .await
            .unwrap();
        let Body::Json(v) = &result.input.body else {
            panic!("expected Body::Json")
        };
        let arr = v.as_array().unwrap();
        assert_eq!(arr.len(), 2);
        assert_eq!(arr[0], serde_json::json!("alpha"));
        assert_eq!(arr[1], serde_json::json!("beta"));
    }

    #[tokio::test]
    async fn test_two_keys_independent_buckets() {
        // completionSize=3 so we can test that A and B accumulate independently.
        let mut svc = new_test_svc(config_size(3));
        svc.ready()
            .await
            .unwrap()
            .call(make_exchange("orderId", "A", "a1"))
            .await
            .unwrap();
        svc.ready()
            .await
            .unwrap()
            .call(make_exchange("orderId", "B", "b1"))
            .await
            .unwrap();
        svc.ready()
            .await
            .unwrap()
            .call(make_exchange("orderId", "A", "a2"))
            .await
            .unwrap();
        // A has 2 items, B has 1 item — neither complete yet
        let ra = svc
            .ready()
            .await
            .unwrap()
            .call(make_exchange("orderId", "A", "a3"))
            .await
            .unwrap();
        // A now has 3 → completes
        assert!(matches!(ra.input.body, Body::Json(_)));
        // B only has 1 → still pending
        let rb = svc
            .ready()
            .await
            .unwrap()
            .call(make_exchange("orderId", "B", "b_check"))
            .await
            .unwrap();
        assert!(matches!(rb.input.body, Body::Empty));
    }

    #[tokio::test]
    async fn test_bucket_resets_after_completion() {
        let mut svc = new_test_svc(config_size(2));
        svc.ready()
            .await
            .unwrap()
            .call(make_exchange("orderId", "A", "x"))
            .await
            .unwrap();
        svc.ready()
            .await
            .unwrap()
            .call(make_exchange("orderId", "A", "x"))
            .await
            .unwrap(); // completes
        // New bucket starts
        let r = svc
            .ready()
            .await
            .unwrap()
            .call(make_exchange("orderId", "A", "new"))
            .await
            .unwrap();
        assert!(matches!(r.input.body, Body::Empty)); // pending again
    }

    #[tokio::test]
    async fn test_completion_size_1_emits_immediately() {
        let mut svc = new_test_svc(config_size(1));
        let ex = make_exchange("orderId", "A", "solo");
        let result = svc.ready().await.unwrap().call(ex).await.unwrap();
        assert!(result.property(CAMEL_AGGREGATOR_PENDING).is_none());
    }

    #[tokio::test]
    async fn test_custom_aggregation_strategy() {
        use camel_api::aggregator::AggregationFn;
        use std::sync::Arc;

        let f: AggregationFn = Arc::new(|mut acc: Exchange, next: Exchange| {
            let combined = format!(
                "{}+{}",
                acc.input.body.as_text().unwrap_or(""),
                next.input.body.as_text().unwrap_or("")
            );
            acc.input.body = Body::Text(combined);
            acc
        });
        let config = AggregatorConfig::correlate_by("key")
            .complete_when_size(2)
            .strategy(AggregationStrategy::Custom(f))
            .build()
            .unwrap();
        let mut svc = new_test_svc(config);
        svc.ready()
            .await
            .unwrap()
            .call(make_exchange("key", "X", "hello"))
            .await
            .unwrap();
        let result = svc
            .ready()
            .await
            .unwrap()
            .call(make_exchange("key", "X", "world"))
            .await
            .unwrap();
        assert_eq!(result.input.body.as_text(), Some("hello+world"));
    }

    #[tokio::test]
    async fn test_completion_predicate() {
        let config = AggregatorConfig::correlate_by("key")
            .complete_when(|bucket| {
                bucket
                    .iter()
                    .any(|e| e.input.body.as_text() == Some("DONE"))
            })
            .build()
            .unwrap();
        let mut svc = new_test_svc(config);
        svc.ready()
            .await
            .unwrap()
            .call(make_exchange("key", "K", "first"))
            .await
            .unwrap();
        svc.ready()
            .await
            .unwrap()
            .call(make_exchange("key", "K", "second"))
            .await
            .unwrap();
        let result = svc
            .ready()
            .await
            .unwrap()
            .call(make_exchange("key", "K", "DONE"))
            .await
            .unwrap();
        assert!(result.property(CAMEL_AGGREGATOR_PENDING).is_none());
    }

    #[tokio::test]
    async fn test_missing_header_returns_error() {
        let mut svc = new_test_svc(config_size(2));
        let msg = Message {
            headers: Default::default(),
            body: Body::Text("no key".into()),
        };
        let ex = Exchange::new(msg);
        let result = svc.ready().await.unwrap().call(ex).await;
        assert!(result.is_err());
        assert!(matches!(
            result.unwrap_err(),
            camel_api::CamelError::ProcessorError(_)
        ));
    }

    #[tokio::test]
    async fn test_cloned_service_shares_state() {
        let svc1 = new_test_svc(config_size(2));
        let mut svc2 = svc1.clone();
        // send first exchange via svc1
        svc1.clone()
            .ready()
            .await
            .unwrap()
            .call(make_exchange("orderId", "A", "from-svc1"))
            .await
            .unwrap();
        // send second exchange via svc2 — should complete because same Arc<Mutex>
        let result = svc2
            .ready()
            .await
            .unwrap()
            .call(make_exchange("orderId", "A", "from-svc2"))
            .await
            .unwrap();
        assert!(result.property(CAMEL_AGGREGATOR_PENDING).is_none());
    }

    #[tokio::test]
    async fn test_camel_aggregated_key_property_set() {
        let mut svc = new_test_svc(config_size(1));
        let ex = make_exchange("orderId", "ORDER-42", "body");
        let result = svc.ready().await.unwrap().call(ex).await.unwrap();
        assert_eq!(
            result.property(CAMEL_AGGREGATED_KEY),
            Some(&serde_json::json!("ORDER-42"))
        );
    }

    #[tokio::test]
    async fn test_aggregator_enforces_max_buckets() {
        let config = AggregatorConfig::correlate_by("orderId")
            .complete_when_size(2)
            .max_buckets(3)
            .build()
            .unwrap();

        let mut svc = new_test_svc(config);

        // Create 3 different correlation keys (fills limit)
        for i in 0..3 {
            let ex = make_exchange("orderId", &format!("key-{}", i), "body");
            let _ = svc.ready().await.unwrap().call(ex).await.unwrap();
        }

        // 4th key should be rejected
        let ex = make_exchange("orderId", "key-4", "body");
        let result = svc.ready().await.unwrap().call(ex).await;

        assert!(result.is_err(), "Should reject when max buckets reached");
        let err = result.unwrap_err().to_string();
        assert!(
            err.contains("maximum"),
            "Error message should contain 'maximum': {}",
            err
        );
    }

    #[tokio::test]
    async fn test_max_buckets_allows_existing_key() {
        let config = AggregatorConfig::correlate_by("orderId")
            .complete_when_size(5) // Large size so bucket doesn't complete
            .max_buckets(2)
            .build()
            .unwrap();

        let mut svc = new_test_svc(config);

        // Create 2 different correlation keys (fills limit)
        let ex1 = make_exchange("orderId", "key-A", "body1");
        let _ = svc.ready().await.unwrap().call(ex1).await.unwrap();
        let ex2 = make_exchange("orderId", "key-B", "body2");
        let _ = svc.ready().await.unwrap().call(ex2).await.unwrap();

        // Should still allow adding to existing key
        let ex3 = make_exchange("orderId", "key-A", "body3");
        let result = svc.ready().await.unwrap().call(ex3).await;
        assert!(
            result.is_ok(),
            "Should allow adding to existing bucket even at max limit"
        );
    }

    /// Audit 2026-08-31, F6-2: one hot correlation key must not buffer
    /// unboundedly. With a predicate-only completion and a tiny
    /// max_bucket_size, the Nth+1 exchange on the same key is rejected.
    #[tokio::test]
    async fn test_aggregator_enforces_max_bucket_size() {
        let config = AggregatorConfig::correlate_by("orderId")
            // Predicate that never fires — completion would otherwise end the test.
            .complete_when(|_| false)
            .max_bucket_size(3)
            .build()
            .unwrap();

        let mut svc = new_test_svc(config);

        for _ in 0..3 {
            let ex = make_exchange("orderId", "hot-key", "body");
            let r = svc.ready().await.unwrap().call(ex).await;
            assert!(r.is_ok(), "first 3 exchanges accepted: {r:?}");
        }

        let ex = make_exchange("orderId", "hot-key", "body");
        let result = svc.ready().await.unwrap().call(ex).await;
        assert!(result.is_err(), "4th exchange on hot key must be rejected");
        let err = result.unwrap_err().to_string();
        assert!(
            err.contains("maximum"),
            "error should mention the per-bucket maximum: {err}"
        );

        // A DIFFERENT key is unaffected (bounded per bucket, not global).
        let ex = make_exchange("orderId", "other-key", "body");
        let result = svc.ready().await.unwrap().call(ex).await;
        assert!(result.is_ok(), "other keys still accepted: {result:?}");
    }

    #[tokio::test]
    async fn test_bucket_ttl_eviction() {
        let config = AggregatorConfig::correlate_by("orderId")
            .complete_when_size(10) // Large size so bucket doesn't complete normally
            .bucket_ttl(Duration::from_millis(50))
            .build()
            .unwrap();

        let mut svc = new_test_svc(config);

        // Create a bucket
        let ex1 = make_exchange("orderId", "key-A", "body1");
        let _ = svc.ready().await.unwrap().call(ex1).await.unwrap();

        // Wait for TTL to expire
        tokio::time::sleep(Duration::from_millis(100)).await;

        // Create a new bucket - this should trigger eviction of the old one
        let ex2 = make_exchange("orderId", "key-B", "body2");
        let _ = svc.ready().await.unwrap().call(ex2).await.unwrap();

        // The expired bucket should have been evicted, so we should be able to
        // add a new key-A bucket again
        let ex3 = make_exchange("orderId", "key-A", "body3");
        let result = svc.ready().await.unwrap().call(ex3).await;
        assert!(result.is_ok(), "Should be able to recreate evicted bucket");
    }

    #[tokio::test(start_paused = true)]
    async fn test_timeout_completes_bucket() {
        let config = AggregatorConfig::correlate_by("key")
            .complete_on_timeout(Duration::from_millis(100))
            .build()
            .unwrap();
        let mut svc = new_test_svc(config);
        let ex = make_exchange("key", "A", "data");
        let result = svc.ready().await.unwrap().call(ex).await.unwrap();
        assert!(result.property(CAMEL_AGGREGATOR_PENDING).is_some());

        tokio::time::sleep(Duration::from_millis(200)).await;

        assert_eq!(
            svc.buckets.lock().unwrap().len(),
            0,
            "bucket should be removed after timeout"
        );
    }

    #[tokio::test(start_paused = true)]
    async fn test_timeout_resets_on_new_exchange() {
        let config = AggregatorConfig::correlate_by("key")
            .complete_on_timeout(Duration::from_millis(150))
            .build()
            .unwrap();
        let mut svc = new_test_svc(config);

        let ex1 = make_exchange("key", "A", "first");
        let _ = svc.ready().await.unwrap().call(ex1).await.unwrap();

        tokio::time::sleep(Duration::from_millis(100)).await;

        let ex2 = make_exchange("key", "A", "second");
        let _ = svc.ready().await.unwrap().call(ex2).await.unwrap();

        tokio::time::sleep(Duration::from_millis(100)).await;

        assert_eq!(
            svc.buckets.lock().unwrap().len(),
            1,
            "bucket should still exist — timeout was reset"
        );

        tokio::time::sleep(Duration::from_millis(100)).await;

        assert_eq!(
            svc.buckets.lock().unwrap().len(),
            0,
            "bucket should be gone after timeout fires"
        );
    }

    #[tokio::test]
    async fn test_composable_size_and_timeout() {
        let config = AggregatorConfig::correlate_by("key")
            .complete_on_size_or_timeout(2, Duration::from_millis(200))
            .build()
            .unwrap();
        let mut svc = new_test_svc(config);

        let ex1 = make_exchange("key", "A", "first");
        let _ = svc.ready().await.unwrap().call(ex1).await.unwrap();
        assert!(svc.buckets.lock().unwrap().contains_key("\"A\""));

        let ex2 = make_exchange("key", "A", "second");
        let result = svc.ready().await.unwrap().call(ex2).await.unwrap();
        assert!(result.property(CAMEL_AGGREGATOR_PENDING).is_none());
        assert_eq!(
            result.property(CAMEL_AGGREGATED_COMPLETION_REASON),
            Some(&serde_json::json!("size"))
        );
    }

    #[tokio::test(start_paused = true)]
    async fn test_discard_on_timeout() {
        let config = AggregatorConfig::correlate_by("key")
            .complete_on_timeout(Duration::from_millis(50))
            .discard_on_timeout(true)
            .build()
            .unwrap();
        let (tx, mut rx) = mpsc::channel(256);
        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        let cancel = CancellationToken::new();
        let mut svc = AggregatorService::new(config, tx, registry, cancel);

        let ex = make_exchange("key", "A", "data");
        let _ = svc.ready().await.unwrap().call(ex).await.unwrap();

        tokio::time::sleep(Duration::from_millis(100)).await;

        assert!(
            rx.try_recv().is_err(),
            "no emit expected with discard_on_timeout"
        );
        assert_eq!(svc.buckets.lock().unwrap().len(), 0);
        assert!(
            svc.timeout_tasks.lock().unwrap().is_empty(),
            "timeout task should be cleaned up"
        );
    }

    #[tokio::test]
    async fn test_force_completion_on_stop() {
        let config = AggregatorConfig::correlate_by("key")
            .complete_when_size(10)
            .force_completion_on_stop(true)
            .build()
            .unwrap();
        let (tx, mut rx) = mpsc::channel(256);
        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        let cancel = CancellationToken::new();
        let svc = AggregatorService::new(config, tx, registry, cancel);

        let mut call_svc = svc.clone();
        let ex = make_exchange("key", "A", "data");
        let _ = call_svc.ready().await.unwrap().call(ex).await.unwrap();

        svc.force_complete_all();

        let result = rx.try_recv().expect("should emit on force-complete");
        let result = result.exchange;
        assert!(
            result.input.body.as_text().is_some() || matches!(result.input.body, Body::Json(_))
        );
        assert_eq!(
            result.property(CAMEL_AGGREGATED_COMPLETION_REASON),
            Some(&serde_json::json!("stop"))
        );
    }

    #[tokio::test]
    async fn test_completion_reason_property_size() {
        let config = AggregatorConfig::correlate_by("key")
            .complete_when_size(1)
            .build()
            .unwrap();
        let mut svc = new_test_svc(config);
        let ex = make_exchange("key", "X", "body");
        let result = svc.ready().await.unwrap().call(ex).await.unwrap();
        assert_eq!(
            result.property(CAMEL_AGGREGATED_COMPLETION_REASON),
            Some(&serde_json::json!("size"))
        );
    }

    #[tokio::test]
    async fn test_completion_reason_property_predicate() {
        let config = AggregatorConfig::correlate_by("key")
            .complete_when(|_| true)
            .build()
            .unwrap();
        let mut svc = new_test_svc(config);
        let ex = make_exchange("key", "X", "body");
        let result = svc.ready().await.unwrap().call(ex).await.unwrap();
        assert_eq!(
            result.property(CAMEL_AGGREGATED_COMPLETION_REASON),
            Some(&serde_json::json!("predicate"))
        );
    }

    #[tokio::test(start_paused = true)]
    async fn test_size_completes_before_timeout() {
        let config = AggregatorConfig::correlate_by("key")
            .complete_on_size_or_timeout(2, Duration::from_millis(200))
            .build()
            .unwrap();
        let mut svc = new_test_svc(config);

        let ex1 = make_exchange("key", "A", "first");
        let _ = svc.ready().await.unwrap().call(ex1).await.unwrap();

        let ex2 = make_exchange("key", "A", "second");
        let result = svc.ready().await.unwrap().call(ex2).await.unwrap();

        assert!(result.property(CAMEL_AGGREGATOR_PENDING).is_none());
        assert_eq!(
            result.property(CAMEL_AGGREGATED_COMPLETION_REASON),
            Some(&serde_json::json!("size"))
        );
        assert_eq!(svc.buckets.lock().unwrap().len(), 0);

        tokio::time::sleep(Duration::from_millis(300)).await;
        assert_eq!(
            svc.buckets.lock().unwrap().len(),
            0,
            "no re-fire after timeout"
        );
    }

    #[tokio::test(start_paused = true)]
    async fn test_concurrent_timeout_fire_and_new_exchange() {
        let config = AggregatorConfig::correlate_by("key")
            .complete_on_size_or_timeout(2, Duration::from_millis(100))
            .build()
            .unwrap();
        let (tx, mut rx) = mpsc::channel(256);
        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        let cancel = CancellationToken::new();
        let mut svc = AggregatorService::new(config, tx, registry, cancel);

        let ex = make_exchange("key", "A", "data");
        let _ = svc.ready().await.unwrap().call(ex).await.unwrap();

        // Advance time past timeout — timeout task fires and removes bucket
        tokio::time::sleep(Duration::from_millis(150)).await;

        // New exchange arrives after timeout — starts a fresh bucket
        let ex2 = make_exchange("key", "A", "data2");
        let result = svc.ready().await.unwrap().call(ex2).await.unwrap();
        assert!(
            result.property(CAMEL_AGGREGATOR_PENDING).is_some(),
            "should be pending in new bucket"
        );

        // Drain late emits from timeout
        let mut late_count = 0;
        while rx.try_recv().is_ok() {
            late_count += 1;
        }
        assert_eq!(
            late_count, 1,
            "exactly 1 late emit from the timed-out bucket"
        );
    }

    #[tokio::test(start_paused = true)]
    async fn test_late_channel_full_drops_with_warning() {
        let config = AggregatorConfig::correlate_by("key")
            .complete_on_timeout(Duration::from_millis(50))
            .build()
            .unwrap();
        let (tx, mut rx) = mpsc::channel(1);
        rx.close();
        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        let cancel = CancellationToken::new();
        let mut svc = AggregatorService::new(config, tx, registry, cancel);

        let ex = make_exchange("key", "A", "data");
        let _ = svc.ready().await.unwrap().call(ex).await.unwrap();

        tokio::time::sleep(Duration::from_millis(100)).await;
        assert_eq!(
            svc.buckets.lock().unwrap().len(),
            0,
            "bucket removed despite channel closed"
        );
    }

    // D-A3 semantic pin: exercises the `force_complete_all()` -> `late_tx.try_send`
    // path specifically (not the timeout-task path covered by
    // `test_late_channel_full_drops_with_warning` above). Pre-saturates the
    // single-slot `late_tx` before constructing the service so every
    // force-completed exchange must drop on the `try_send` failure branch.
    #[tokio::test]
    async fn test_da3_force_complete_all_drops_on_saturated_channel() {
        let config = AggregatorConfig::correlate_by("k")
            .complete_when_size(10)
            .force_completion_on_stop(true)
            .build()
            .unwrap();
        // capacity 1 — deliberately tiny so the pre-fill fully saturates the slot.
        let (late_tx, mut late_rx) = mpsc::channel::<AggregateEmission>(1);
        // Pre-saturate the 1-slot mpsc BEFORE the Sender is moved into
        // AggregatorService::new. The Sender is taken by value into the
        // service, so this `try_send` is the only chance to occupy the slot.
        late_tx
            .try_send(AggregateEmission {
                exchange: make_exchange("k", "99", "dummy"),
                claims: Vec::new(),
            })
            .expect("pre-fill succeeds");

        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        let cancel = CancellationToken::new();
        let mut svc = AggregatorService::new(config, late_tx, registry, cancel);

        // Three distinct correlation keys -> three buckets, each holding one
        // exchange. size=10 means no bucket auto-completes.
        for v in ["1", "2", "3"] {
            let ex = make_exchange("k", v, "body");
            let _ = svc.ready().await.unwrap().call(ex).await.unwrap();
        }
        assert_eq!(svc.buckets.lock().unwrap().len(), 3);

        // Action: drive the force-complete path.
        svc.force_complete_all();

        // (a) the manually-sent pre-fill item is still in the channel.
        let pre_fill = late_rx
            .try_recv()
            .expect("pre-fill should still be in channel");
        assert_eq!(
            pre_fill.exchange.input.headers.get("k"),
            Some(&serde_json::json!("99"))
        );

        // (b) channel drained -- no force-completed exchange got through
        // (all three try_send calls hit the full channel and dropped).
        assert!(matches!(
            late_rx.try_recv(),
            Err(mpsc::error::TryRecvError::Empty)
        ));

        // (c) all three buckets were removed during force_complete_all.
        assert!(svc.buckets.lock().unwrap().is_empty());
    }

    #[tokio::test]
    async fn test_aggregate_stream_bodies_creates_valid_json() {
        use bytes::Bytes;
        use camel_api::{Body, StreamBody, StreamMetadata};
        use futures::stream;
        use tokio::sync::Mutex;

        let chunks = vec![Ok(Bytes::from("test"))];
        let stream_body = StreamBody {
            stream: Arc::new(Mutex::new(Some(Box::pin(stream::iter(chunks))))),
            metadata: StreamMetadata {
                origin: Some("file:///test.txt".to_string()),
                ..Default::default()
            },
        };

        let ex1 = Exchange::new(Message {
            headers: Default::default(),
            body: Body::Stream(stream_body),
        });

        let exchanges = vec![ex1];
        let result = aggregate(exchanges, &AggregationStrategy::CollectAll);

        let exchange = result.expect("Expected Ok result");
        assert!(
            matches!(exchange.input.body, Body::Json(_)),
            "Expected Json body"
        );

        if let Body::Json(value) = exchange.input.body {
            let json_str = serde_json::to_string(&value).unwrap();
            let parsed: serde_json::Value = serde_json::from_str(&json_str).unwrap();

            assert!(parsed.is_array(), "Result should be an array");
            let arr = parsed.as_array().unwrap();
            assert!(arr[0].is_object(), "First element should be an object");
            assert!(
                arr[0]["_stream"].is_object(),
                "Should contain _stream object"
            );
            assert_eq!(arr[0]["_stream"]["origin"], "file:///test.txt");
            assert_eq!(
                arr[0]["_stream"]["placeholder"], true,
                "placeholder flag should be true"
            );
        }
    }

    #[tokio::test]
    async fn test_aggregate_stream_bodies_with_none_origin() {
        use bytes::Bytes;
        use camel_api::{Body, StreamBody, StreamMetadata};
        use futures::stream;
        use tokio::sync::Mutex;

        let chunks = vec![Ok(Bytes::from("test"))];
        let stream_body = StreamBody {
            stream: Arc::new(Mutex::new(Some(Box::pin(stream::iter(chunks))))),
            metadata: StreamMetadata {
                origin: None,
                ..Default::default()
            },
        };

        let ex1 = Exchange::new(Message {
            headers: Default::default(),
            body: Body::Stream(stream_body),
        });

        let exchanges = vec![ex1];
        let result = aggregate(exchanges, &AggregationStrategy::CollectAll);

        let exchange = result.expect("Expected Ok result");
        assert!(
            matches!(exchange.input.body, Body::Json(_)),
            "Expected Json body"
        );

        if let Body::Json(value) = exchange.input.body {
            let json_str = serde_json::to_string(&value).unwrap();
            let parsed: serde_json::Value = serde_json::from_str(&json_str).unwrap();

            assert!(parsed.is_array(), "Result should be an array");
            let arr = parsed.as_array().unwrap();
            assert!(arr[0].is_object(), "First element should be an object");
            assert!(
                arr[0]["_stream"].is_object(),
                "Should contain _stream object"
            );
            assert_eq!(
                arr[0]["_stream"]["origin"],
                serde_json::Value::Null,
                "origin should be null when None"
            );
            assert_eq!(
                arr[0]["_stream"]["placeholder"], true,
                "placeholder flag should be true"
            );
        }
    }

    #[tokio::test]
    async fn timeout_completion_clears_handle_from_map() {
        // Regression: Oracle audit found that natural timeout completion removed
        // the bucket but left the JoinHandle in `timeout_handles`, leaking the
        // entry until route shutdown. After fix, the timeout task itself cleans
        // its handle from the map on natural completion.
        let config = AggregatorConfig::correlate_by("key")
            .complete_on_timeout(Duration::from_millis(50))
            .build()
            .unwrap();
        let (tx, _rx) = mpsc::channel(256);
        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        let cancel = CancellationToken::new();
        let svc = AggregatorService::new(config, tx, registry, cancel);

        // Send an exchange to create a pending bucket with a timeout task.
        let mut call_svc = svc.clone();
        let ex = make_exchange("key", "A", "data");
        let _ = call_svc.ready().await.unwrap().call(ex).await.unwrap();
        assert!(
            !svc.timeout_handles.lock().unwrap().is_empty(),
            "handle should exist while timeout pending"
        );

        // Wait real time for the 50ms timeout to fire + spawned task to complete.
        tokio::time::sleep(Duration::from_millis(200)).await;

        assert!(
            svc.timeout_handles.lock().unwrap().is_empty(),
            "handle should be cleared from map after natural timeout completion (was leak)"
        );
    }

    #[tokio::test]
    async fn aggregator_shutdown_via_trait_dispatch() {
        // RED: builds an AggregatorService, dispatches through Arc<dyn StepLifecycle>,
        // and asserts idempotent shutdown works.
        let config = AggregatorConfig::correlate_by("key")
            .complete_when_size(10)
            .build()
            .unwrap();
        let (tx, _rx) = mpsc::channel(256);
        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        let cancel = CancellationToken::new();
        let svc = AggregatorService::new(config, tx, registry, cancel);

        let step: Arc<dyn StepLifecycle> = Arc::new(svc);
        step.shutdown(StepShutdownReason::RouteStop)
            .await
            .expect("first shutdown should succeed");
        step.shutdown(StepShutdownReason::RouteStop)
            .await
            .expect("second shutdown (idempotent) should succeed");
    }

    #[tokio::test(start_paused = true)]
    async fn test_shutdown_awaits_timeout_handles() {
        let config = AggregatorConfig::correlate_by("key")
            .complete_on_timeout(Duration::from_millis(100))
            .build()
            .unwrap();
        let (tx, _rx) = mpsc::channel(256);
        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        let cancel = CancellationToken::new();
        let svc = AggregatorService::new(config, tx, registry, cancel);

        // Send an exchange to create a pending bucket with a timeout task.
        let mut call_svc = svc.clone();
        let ex = make_exchange("key", "A", "data");
        let _ = call_svc.ready().await.unwrap().call(ex).await.unwrap();

        // Verify timeout handle exists.
        assert!(
            !svc.timeout_handles.lock().unwrap().is_empty(),
            "should have a timeout handle"
        );

        // Shutdown should complete within the 5s deadline (the timeout task
        // gets cancelled so it won't wait for the full 100ms sleep).
        svc.shutdown_inner().await;

        assert!(
            svc.timeout_handles.lock().unwrap().is_empty(),
            "all handles should be cleaned up after shutdown"
        );
    }

    // ── R3-C1 Batch 1: DoS cap + background sweep ───────────────────

    /// R3-C1: a flood of unique correlation keys must stay bounded.
    /// Default `max_buckets` is 10_000; the 10_001st unique key is rejected
    /// with `Aggregator reached maximum N buckets` (or its updated equivalent
    /// after the fix). The unique-key flood does NOT OOM the process.
    #[tokio::test]
    async fn test_unique_key_flood_stays_bounded_by_default() {
        // Builder defaults to max_buckets = 10_000, bucket_ttl = 300s.
        let config = AggregatorConfig::correlate_by("orderId")
            .complete_when_size(1_000_000) // never completes normally
            .build()
            .unwrap();
        let mut svc = new_test_svc(config);

        // Send 10_001 unique keys. The first 10_000 should be accepted
        // (pending in their buckets); the 10_001st MUST be rejected.
        for i in 0..10_000usize {
            let ex = make_exchange("orderId", &format!("key-{i}"), "body");
            let result = svc.ready().await.unwrap().call(ex).await;
            assert!(result.is_ok(), "key {i} should be accepted under the cap");
        }
        let ex = make_exchange("orderId", "key-10001", "body");
        let result = svc.ready().await.unwrap().call(ex).await;
        assert!(
            result.is_err(),
            "10_001st unique key must be rejected by the max_buckets cap"
        );
        let err = result.unwrap_err().to_string();
        assert!(
            err.contains("maximum") || err.contains("max"),
            "error should mention cap: {err}"
        );
    }

    /// The `AggregatorService` exposes a `sweep_handle` for the background
    /// sweep task. When `config.bucket_ttl` is `Some`, `AggregatorService::new`
    /// automatically spawns the sweep task and stores the handle, so the
    /// caller never sees `None` for a TTL-configured service. Cancelling the
    /// route token (via `shutdown`) aborts the sweep.
    #[tokio::test]
    async fn test_background_sweep_spawns_on_first_poll_not_construction() {
        let config = AggregatorConfig::correlate_by("key")
            .complete_when_size(10_000)
            .bucket_ttl(Duration::from_millis(50))
            .build()
            .unwrap();
        let cancel = CancellationToken::new();
        let (tx, _rx) = mpsc::channel(8);
        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        let mut svc = AggregatorService::new(config, tx, registry, cancel.clone());

        // Construction is runtime-free: no sweep spawned yet.
        assert!(
            svc.sweep_handle
                .lock()
                .unwrap_or_else(|e| e.into_inner())
                .is_none(),
            "sweep must NOT be spawned at construction (runtime-free new)"
        );

        // First readiness poll lazily spawns the sweep (a runtime is present
        // here). `ready()` drives `poll_ready` until Ready.
        let _ = svc.ready().await.unwrap();
        let sweep_present = svc
            .sweep_handle
            .lock()
            .unwrap_or_else(|e| e.into_inner())
            .is_some();
        assert!(
            sweep_present,
            "sweep handle should be Some after first poll when bucket_ttl is set"
        );

        // Cancel the route token; the sweep task observes it and exits.
        cancel.cancel();
        // Give the task a moment to observe the cancel.
        tokio::time::sleep(Duration::from_millis(50)).await;
    }

    /// Shared `MetricsCollector` double that logs every `set_queue_depth`
    /// call (queue label, depth). Used by the queue-depth sampling tests.
    struct QueueDepthRecorder(Mutex<Vec<(String, usize)>>);
    impl MetricsCollector for QueueDepthRecorder {
        fn record_exchange_duration(&self, _: &str, _: std::time::Duration) {}
        fn increment_errors(&self, _: &str, _: &str) {}
        fn increment_exchanges(&self, _: &str) {}
        fn set_queue_depth(&self, queue: &str, depth: usize) {
            self.0
                .lock()
                .unwrap_or_else(|e| e.into_inner())
                .push((queue.to_string(), depth));
        }
        fn record_circuit_breaker_change(&self, _: &str, _: &str, _: &str) {}
    }

    /// Poll the recorder until a sample for `label` matches `pred`, or fail
    /// after `deadline`. Keeps the sampling tests scheduling-tolerant.
    async fn await_depth_sample(
        recorder: &QueueDepthRecorder,
        label: &str,
        pred: impl Fn(usize) -> bool,
    ) {
        let deadline = std::time::Instant::now() + Duration::from_secs(2);
        loop {
            let matched = recorder
                .0
                .lock()
                .unwrap_or_else(|e| e.into_inner())
                .iter()
                .any(|(q, d)| q == label && pred(*d));
            if matched {
                return;
            }
            assert!(
                std::time::Instant::now() < deadline,
                "no queue-depth sample for '{label}' matched within 2s"
            );
            tokio::time::sleep(Duration::from_millis(20)).await;
        }
    }

    /// Queue-depth sampling (dashboard-observability T3.3): when
    /// `with_queue_metrics` is set, the TTL-sweep maintenance pass publishes
    /// the buffered group count, and the count returns to zero once the
    /// group completes.
    #[tokio::test]
    async fn test_sweep_reports_queue_depth() {
        let config = AggregatorConfig::correlate_by("orderId")
            .complete_when_size(3)
            // Short TTL keeps the sweep interval at its 50ms floor so the
            // test observes samples quickly; the bucket is completed well
            // before eviction.
            .bucket_ttl(Duration::from_millis(100))
            .build()
            .unwrap();
        let cancel = CancellationToken::new();
        let (tx, _rx) = mpsc::channel(8);
        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        let recorder = Arc::new(QueueDepthRecorder(Mutex::new(Vec::new())));
        let mut svc = AggregatorService::new(config, tx, registry, cancel).with_queue_metrics(
            Arc::clone(&recorder) as Arc<dyn MetricsCollector>,
            "aggregator:t",
        );

        // Spawn the sweep via the first readiness poll.
        let _ = svc.ready().await.unwrap();

        // Partial group: 1 of 3 messages.
        let ex = make_exchange("orderId", "g1", "partial");
        let _ = svc.ready().await.unwrap().call(ex).await;

        await_depth_sample(&recorder, "aggregator:t", |d| d > 0).await;

        // Complete the group; the sweep must report zero again.
        let ex = make_exchange("orderId", "g1", "b");
        let _ = svc.ready().await.unwrap().call(ex).await;
        let ex = make_exchange("orderId", "g1", "c");
        let _ = svc.ready().await.unwrap().call(ex).await;

        await_depth_sample(&recorder, "aggregator:t", |d| d == 0).await;

        svc.shutdown(StepShutdownReason::RouteStop).await.unwrap();
    }

    /// F2 (T3.3 review): a no-TTL aggregator (size-only completion,
    /// `bucket_ttl = None`) with `queue_metrics` set must still sample —
    /// the sweep used to live entirely inside the TTL branch and never ran.
    /// Eviction stays TTL-gated; only sampling runs.
    #[tokio::test]
    async fn test_no_ttl_aggregator_reports_queue_depth() {
        use camel_api::aggregator::CorrelationStrategy;

        let config = AggregatorConfig {
            header_name: "orderId".into(),
            completion: CompletionMode::Single(CompletionCondition::Size(3)),
            correlation: CorrelationStrategy::HeaderName("orderId".into()),
            strategy: AggregationStrategy::CollectAll,
            max_buckets: Some(100),
            max_bucket_size: Some(100),
            bucket_ttl: None,
            force_completion_on_stop: false,
            discard_on_timeout: false,
            max_timeout_tasks: 64,
        };
        let cancel = CancellationToken::new();
        let (tx, _rx) = mpsc::channel(8);
        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        let recorder = Arc::new(QueueDepthRecorder(Mutex::new(Vec::new())));
        let mut svc = AggregatorService::new(config, tx, registry, cancel).with_queue_metrics(
            Arc::clone(&recorder) as Arc<dyn MetricsCollector>,
            "aggregator:nottl",
        );

        let _ = svc.ready().await.unwrap();
        assert!(
            svc.sweep_handle
                .lock()
                .unwrap_or_else(|e| e.into_inner())
                .is_some(),
            "metrics-only config (bucket_ttl = None) must still spawn the sweep"
        );

        // Partial group: 1 of 3 messages — depth must publish > 0.
        let ex = make_exchange("orderId", "g1", "partial");
        let _ = svc.ready().await.unwrap().call(ex).await;
        await_depth_sample(&recorder, "aggregator:nottl", |d| d > 0).await;

        // Complete the group; depth must publish 0 again.
        let ex = make_exchange("orderId", "g1", "b");
        let _ = svc.ready().await.unwrap().call(ex).await;
        let ex = make_exchange("orderId", "g1", "c");
        let _ = svc.ready().await.unwrap().call(ex).await;
        await_depth_sample(&recorder, "aggregator:nottl", |d| d == 0).await;

        svc.shutdown(StepShutdownReason::RouteStop).await.unwrap();
    }

    /// F3 (T3.3 review): the last-owner guard — the pipeline clones the
    /// service on every `poll_ready`; a transient clone dropping must NOT
    /// abort the shared sweep. The sweep keeps ticking (depth still
    /// published) until the final owner drops.
    #[tokio::test]
    async fn test_transient_clone_drop_keeps_sweep_sampling() {
        let config = AggregatorConfig::correlate_by("orderId")
            .complete_when_size(3)
            .bucket_ttl(Duration::from_millis(100))
            .build()
            .unwrap();
        let cancel = CancellationToken::new();
        let (tx, _rx) = mpsc::channel(8);
        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        let recorder = Arc::new(QueueDepthRecorder(Mutex::new(Vec::new())));
        let mut svc = AggregatorService::new(config, tx, registry, cancel).with_queue_metrics(
            Arc::clone(&recorder) as Arc<dyn MetricsCollector>,
            "aggregator:clone",
        );

        let _ = svc.ready().await.unwrap();

        // Transient clone drops (as the pipeline does per poll_ready).
        drop(svc.clone());
        assert!(
            svc.sweep_handle
                .lock()
                .unwrap_or_else(|e| e.into_inner())
                .is_some(),
            "transient clone drop must not abort the shared sweep"
        );

        // The sweep still ticks: a partial group produces depth samples.
        let ex = make_exchange("orderId", "g1", "partial");
        let _ = svc.ready().await.unwrap().call(ex).await;
        await_depth_sample(&recorder, "aggregator:clone", |d| d > 0).await;

        svc.shutdown(StepShutdownReason::RouteStop).await.unwrap();
    }

    // ── R3-M3: bounded timeout-task spawn ─────────────────────────────

    #[tokio::test]
    async fn test_aggregator_timeout_task_cap_no_panic_under_flood() {
        // R3-M3: a flood of unique keys with a tiny max_timeout_tasks must not
        // spawn unbounded tasks, panic, or deadlock. Each call returns Ok(pending).
        use camel_api::aggregator::CorrelationStrategy;

        let config = AggregatorConfig {
            header_name: "k".into(),
            completion: CompletionMode::Any(vec![
                CompletionCondition::Size(999),
                CompletionCondition::Timeout(Duration::from_secs(30)),
            ]),
            correlation: CorrelationStrategy::HeaderName("k".into()),
            strategy: AggregationStrategy::CollectAll,
            max_buckets: Some(50),
            max_bucket_size: Some(50),
            bucket_ttl: Some(Duration::from_secs(30)),
            force_completion_on_stop: false,
            discard_on_timeout: false,
            max_timeout_tasks: 2,
        };
        let (late_tx, mut late_rx) = mpsc::channel(64);
        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        let cancel = CancellationToken::new();
        let svc = AggregatorService::new(config, late_tx, registry, cancel);

        // Drive 20 unique-key exchanges — far exceeding max_timeout_tasks=2.
        for i in 0..20u64 {
            let mut ex = Exchange::new(Message {
                headers: HashMap::from([("k".to_string(), serde_json::json!(i))]),
                body: Body::Text(i.to_string()),
            });
            ex.input
                .headers
                .insert("k".to_string(), serde_json::json!(i));
            let outcome = tokio::time::timeout(Duration::from_secs(2), async {
                let mut s = svc.clone();
                use tower::ServiceExt;
                s.ready().await.unwrap().call(ex).await
            })
            .await;
            assert!(outcome.is_ok(), "call {} hung/panicked under task cap", i);
            // Each returns Ok(pending) since Size(999) is never reached.
            let res = outcome.unwrap().unwrap();
            assert_eq!(
                res.properties
                    .get(CAMEL_AGGREGATOR_PENDING)
                    .and_then(|v| v.as_bool()),
                Some(true),
                "exchange {} should be pending",
                i
            );
        }
        // Drain any late emissions to avoid blocking the channel.
        let _ = late_rx.try_recv();
    }

    /// bd rc-iioeq: consumer-exit release is decided PER BUCKET, not from
    /// config-global `has_timeout()`. With the timeout-task cap reached, a
    /// timeout-configured bucket stays unarmed; after the consumer exits
    /// it can never complete on its own (no task owns it, the bucket_ttl
    /// sweep only runs inside the next `call`, which never arrives).
    /// `release_unarmed_buckets` must discard exactly those buckets while
    /// leaving armed ones to their timeout tasks.
    #[tokio::test]
    async fn test_release_unarmed_buckets_discards_cap_exceeded_keeps_armed() {
        let config = AggregatorConfig::correlate_by("k")
            .complete_on_timeout(Duration::from_millis(400))
            .max_timeout_tasks(1)
            .build()
            .unwrap();
        let (late_tx, mut late_rx) = mpsc::channel(8);
        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        let cancel = CancellationToken::new();
        let svc = AggregatorService::new(config, late_tx, registry, cancel);

        // Key "a" arms the only timeout slot; key "b" hits the cap and
        // stays unarmed.
        let a = make_exchange("k", "a", "body-a");
        let mut sa = svc.clone();
        let _ = sa.ready().await.unwrap().call(a).await.unwrap();
        let b = make_exchange("k", "b", "body-b");
        let mut sb = svc.clone();
        let _ = sb.ready().await.unwrap().call(b).await.unwrap();

        {
            let armed = svc.timeout_tasks.lock().unwrap_or_else(|e| e.into_inner());
            assert_eq!(armed.len(), 1, "cap=1 must arm exactly one timeout task");
            assert!(
                armed.contains_key("\"a\""),
                "first key must hold the armed slot, got {armed:?}"
            );
        }

        // Consumer-exit release: unarmed bucket discarded, armed kept.
        svc.release_unarmed_buckets();
        {
            let buckets = svc.buckets.lock().unwrap_or_else(|e| e.into_inner());
            assert!(
                buckets.contains_key("\"a\""),
                "armed bucket must survive the release"
            );
            assert!(
                !buckets.contains_key("\"b\""),
                "cap-exceeded (unarmed) bucket must be released, not orphaned"
            );
        }

        // The armed bucket's timeout still owns completion: it emits
        // reason=timeout through the late channel.
        let emitted = tokio::time::timeout(Duration::from_secs(2), late_rx.recv())
            .await
            .expect("armed bucket must emit on its timeout")
            .expect("late channel must stay open");
        assert_eq!(
            emitted
                .exchange
                .properties
                .get(CAMEL_AGGREGATED_COMPLETION_REASON),
            Some(&serde_json::json!("timeout"))
        );

        // The released bucket must never emit: both keys were sent within
        // milliseconds of each other, so 600 ms after "a"'s emission is
        // past "b"'s would-be deadline (400 ms) with margin.
        assert!(
            tokio::time::timeout(Duration::from_millis(600), late_rx.recv())
                .await
                .is_err(),
            "released unarmed bucket must not emit"
        );

        svc.shutdown(StepShutdownReason::RouteStop).await.unwrap();
    }

    #[tokio::test]
    async fn evaluate_completion_predicate_or_combines_any() {
        use camel_api::aggregator::{CompletionCondition, CompletionMode};
        use camel_language_api::Language;
        use std::collections::HashMap;

        // Register the simple language in an otherwise-empty registry.
        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        registry.lock().unwrap().insert(
            "simple".to_string(),
            Arc::new(camel_language_simple::SimpleLanguage::new()) as Arc<dyn Language>,
        );

        let incoming = make_exchange("k", "X", "hello");

        // Any with two PredicateExpr conditions; second matches body == "hello".
        let mode = CompletionMode::Any(vec![
            CompletionCondition::PredicateExpr {
                expr: "${body} == 'NOPE'".to_string(),
                language: "simple".to_string(),
            },
            CompletionCondition::PredicateExpr {
                expr: "${body} == 'hello'".to_string(),
                language: "simple".to_string(),
            },
        ]);

        let satisfied = evaluate_completion_predicate(&mode, &incoming, &registry)
            .await
            .expect("eval must succeed");
        assert!(satisfied, "second predicate matches → OR true");
    }

    #[tokio::test]
    async fn evaluate_completion_predicate_no_predicate_skips_registry() {
        use camel_api::aggregator::{CompletionCondition, CompletionMode};
        use std::collections::HashMap;

        // Size-only completion: iter_predicate_exprs returns empty vec.
        // The registry is EMPTY — any accidental lock-and-get would fail loudly,
        // proving the fast path does not touch the registry.
        let mode = CompletionMode::Single(CompletionCondition::Size(3));
        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        let incoming = make_exchange("k", "X", "hello");
        let result = evaluate_completion_predicate(&mode, &incoming, &registry).await;
        assert!(result.is_ok(), "fast path must not error: {:?}", result);
        assert!(!result.unwrap(), "no predicate → not satisfied");
    }

    #[tokio::test]
    async fn evaluate_completion_predicate_unregistered_language_errors() {
        use camel_api::aggregator::{CompletionCondition, CompletionMode};

        let mode = CompletionMode::Single(CompletionCondition::PredicateExpr {
            expr: "${body} == 'DONE'".to_string(),
            language: "nonexistent-lang".to_string(),
        });
        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        let incoming = make_exchange("k", "X", "hello");
        let result = evaluate_completion_predicate(&mode, &incoming, &registry).await;
        assert!(result.is_err(), "unregistered language must error");
    }

    #[tokio::test]
    async fn evaluate_completion_predicate_all_miss_returns_false() {
        use camel_api::aggregator::{CompletionCondition, CompletionMode};
        use camel_language_api::Language;
        use std::collections::HashMap;

        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        registry.lock().unwrap().insert(
            "simple".to_string(),
            Arc::new(camel_language_simple::SimpleLanguage::new()) as Arc<dyn Language>,
        );

        let mode = CompletionMode::Single(CompletionCondition::PredicateExpr {
            expr: "${body} == 'NOPE'".to_string(),
            language: "simple".to_string(),
        });
        let incoming = make_exchange("k", "X", "hello");
        let result = evaluate_completion_predicate(&mode, &incoming, &registry).await;
        assert!(result.is_ok(), "eval must succeed: {:?}", result);
        assert!(!result.unwrap(), "predicate does not match");
    }

    #[tokio::test]
    async fn completion_predicate_non_bool_is_error() {
        use camel_api::aggregator::{CompletionCondition, CompletionMode};
        use camel_language_api::Language;
        use std::collections::HashMap;

        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        registry.lock().unwrap().insert(
            "simple".to_string(),
            Arc::new(camel_language_simple::SimpleLanguage::new()) as Arc<dyn Language>,
        );

        // A bare slot is a value, not a boolean. Completion predicates are
        // strict: the Simple predicate reports a type mismatch (expected bool)
        // instead of coercing the string to false.
        let mode = CompletionMode::Single(CompletionCondition::PredicateExpr {
            expr: "${body}".to_string(),
            language: "simple".to_string(),
        });
        let incoming = make_exchange("k", "X", "x");
        let err = evaluate_completion_predicate(&mode, &incoming, &registry)
            .await
            .expect_err("non-bool predicate must error, not coerce to false");
        // Redacted diagnostic: type names only, asserted via the "bool" text,
        // never the runtime body value. The aggregator has no `EvalMeta`, so
        // there is no `ExpressionFailed` class to assert.
        assert!(
            matches!(&err, CamelError::ProcessorError(msg) if msg.contains("bool")),
            "error must carry the type-mismatch text naming bool: {err:?}"
        );
    }

    #[tokio::test]
    async fn completion_predicate_eval_error_propagates() {
        use camel_api::aggregator::{CompletionCondition, CompletionMode};
        use camel_language_api::Language;
        use std::collections::HashMap;

        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        registry.lock().unwrap().insert(
            "simple".to_string(),
            Arc::new(camel_language_simple::SimpleLanguage::new()) as Arc<dyn Language>,
        );

        // `${body} > 1` over a non-numeric body is an evaluation error. The
        // aggregator has no `EvalMeta`, so it surfaces as `ProcessorError`.
        let mode = CompletionMode::Single(CompletionCondition::PredicateExpr {
            expr: "${body} > 1".to_string(),
            language: "simple".to_string(),
        });
        let incoming = make_exchange("k", "X", "hello");
        let err = evaluate_completion_predicate(&mode, &incoming, &registry)
            .await
            .expect_err("evaluation error must propagate, not coerce to false");
        assert!(
            matches!(err, CamelError::ProcessorError(_)),
            "eval error must be a ProcessorError: {err:?}"
        );
    }

    #[tokio::test]
    async fn completion_predicate_expr_completes_on_match() {
        use camel_api::aggregator::{CompletionCondition, CompletionMode};
        use camel_language_api::Language;
        use std::collections::HashMap;

        // Build a registry with the simple language registered.
        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        registry.lock().unwrap().insert(
            "simple".to_string(),
            Arc::new(camel_language_simple::SimpleLanguage::new()) as Arc<dyn Language>,
        );

        // Build the config via the builder for correlation + size defaults
        // (required by `AggregatorService::new`'s validate()), then OVERRIDE
        // the completion field directly — there is no builder setter for
        // `PredicateExpr`.
        let mut config = AggregatorConfig::correlate_by("key")
            .complete_when_size(999) // placeholder; replaced below
            .build()
            .expect("config builds");
        config.completion = CompletionMode::Single(CompletionCondition::PredicateExpr {
            expr: "${body} == 'DONE'".to_string(),
            language: "simple".to_string(),
        });

        let mut svc = new_test_svc_with_registry(config, registry);

        // First exchange — predicate false → still pending.
        let r = svc
            .ready()
            .await
            .unwrap()
            .call(make_exchange("key", "K", "first"))
            .await
            .unwrap();
        assert!(r.property(CAMEL_AGGREGATOR_PENDING).is_some());

        // Matching exchange — predicate true → bucket completes.
        let r = svc
            .ready()
            .await
            .unwrap()
            .call(make_exchange("key", "K", "DONE"))
            .await
            .unwrap();
        assert!(r.property(CAMEL_AGGREGATOR_PENDING).is_none());
        assert_eq!(
            r.property(CAMEL_AGGREGATED_COMPLETION_REASON),
            Some(&serde_json::json!("predicate"))
        );
    }

    // ── ADR-0046 D-A1: binary-fold strategy contract (no null oldExchange) ──
    //
    // Apache Camel's AggregationStrategy receives a null `oldExchange` on the
    // FIRST message of a bucket so the strategy can initialize. rust-camel's
    // AggregationFn has a binary signature `(Exchange, Exchange) -> Exchange`
    // and the bucket model holds the first message untouched; the strategy is
    // only first invoked on the SECOND message with both exchanges present.
    //
    // This test pins that contract: a strategy needing initialize-on-first
    // logic cannot rely on a null oldExchange — it must branch on a sentinel
    // in the accumulated body or on a property.
    #[tokio::test]
    async fn test_da1_strategy_receives_two_exchanges_first_message_preserved() {
        use camel_api::aggregator::{AggregationFn, AggregationStrategy};
        use std::sync::Arc;

        let recorded: Arc<std::sync::Mutex<Vec<(String, String)>>> =
            Arc::new(std::sync::Mutex::new(Vec::new()));
        let recorded_for_closure = Arc::clone(&recorded);

        let f: AggregationFn = Arc::new(move |old: Exchange, new: Exchange| {
            let old_body = old.input.body.as_text().unwrap_or("").to_string();
            let new_body = new.input.body.as_text().unwrap_or("").to_string();
            recorded_for_closure
                .lock()
                .expect("recorded mutex poisoned")
                .push((old_body, new_body));
            new
        });

        let config = AggregatorConfig::correlate_by("k")
            .complete_when_size(2)
            .strategy(AggregationStrategy::Custom(f))
            .build()
            .unwrap();
        let mut svc = new_test_svc(config);

        // First message: bucket goes 0→1, strategy NOT invoked, return pending.
        let first = svc
            .ready()
            .await
            .unwrap()
            .call(make_exchange("k", "1", "A"))
            .await
            .unwrap();
        assert!(
            first.property(CAMEL_AGGREGATOR_PENDING).is_some(),
            "first message must leave the bucket pending (size < 2)"
        );
        assert!(
            recorded.lock().expect("recorded mutex poisoned").is_empty(),
            "strategy must NOT be invoked on the first message of a bucket"
        );

        // Second message: bucket goes 1→2, strategy IS invoked as f(ex1, ex2),
        // and the result is emitted.
        let _completed = svc
            .ready()
            .await
            .unwrap()
            .call(make_exchange("k", "1", "B"))
            .await
            .unwrap();

        let recorded = recorded.lock().expect("recorded mutex poisoned");
        assert_eq!(
            recorded.len(),
            1,
            "strategy must be invoked exactly once across the two-message bucket, got {recorded:?}"
        );
        assert_eq!(
            recorded[0],
            ("A".to_string(), "B".to_string()),
            "strategy must observe the first message as `old` and the second as `new`, \
             with both bodies preserved unchanged"
        );
    }

    // ── Sweep lifecycle tests (aggregate-route-cancel-threading) ──────

    #[tokio::test]
    async fn sweep_shutdown_cancels_task() {
        let config = AggregatorConfig::correlate_by("key")
            .complete_when_size(10)
            .bucket_ttl(Duration::from_millis(100))
            .build()
            .unwrap();
        let (tx, _rx) = mpsc::channel(256);
        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        let cancel = CancellationToken::new();
        let mut svc = AggregatorService::new(config, tx, registry, cancel);

        let _ = svc.ready().await.unwrap();

        assert!(
            svc.sweep_handle
                .lock()
                .unwrap_or_else(|e| e.into_inner())
                .is_some(),
            "sweep handle should be Some after poll_ready"
        );

        svc.shutdown(StepShutdownReason::RouteStop).await.unwrap();

        assert!(
            svc.sweep_handle
                .lock()
                .unwrap_or_else(|e| e.into_inner())
                .is_none(),
            "sweep handle should be None after shutdown (taken + aborted)"
        );
        assert!(
            svc.sweep_cancel
                .lock()
                .unwrap_or_else(|e| e.into_inner())
                .is_cancelled(),
            "sweep_cancel token should be cancelled after shutdown"
        );

        tokio::time::sleep(Duration::from_millis(50)).await;
    }

    #[tokio::test]
    async fn sweep_start_respawns_after_shutdown() {
        let config = AggregatorConfig::correlate_by("key")
            .complete_when_size(10)
            .bucket_ttl(Duration::from_millis(100))
            .build()
            .unwrap();
        let (tx, _rx) = mpsc::channel(256);
        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        let cancel = CancellationToken::new();
        let mut svc = AggregatorService::new(config, tx, registry, cancel);

        let _ = svc.ready().await.unwrap();
        svc.shutdown(StepShutdownReason::RouteStop).await.unwrap();

        svc.start().await.unwrap();
        let _ = svc.ready().await.unwrap();

        assert!(
            svc.sweep_handle
                .lock()
                .unwrap_or_else(|e| e.into_inner())
                .is_some(),
            "sweep handle should be Some after start + poll_ready"
        );
        assert!(
            !svc.sweep_cancel
                .lock()
                .unwrap_or_else(|e| e.into_inner())
                .is_cancelled(),
            "sweep_cancel should be a fresh uncancelled token after start"
        );
    }

    #[tokio::test]
    async fn sweep_shutdown_hotswap_cancels_task() {
        let config = AggregatorConfig::correlate_by("key")
            .complete_when_size(10)
            .bucket_ttl(Duration::from_millis(100))
            .build()
            .unwrap();
        let (tx, _rx) = mpsc::channel(256);
        let registry: SharedLanguageRegistry = Arc::new(std::sync::Mutex::new(HashMap::new()));
        let cancel = CancellationToken::new();
        let mut svc = AggregatorService::new(config, tx, registry, cancel);

        let _ = svc.ready().await.unwrap();

        assert!(
            svc.sweep_handle
                .lock()
                .unwrap_or_else(|e| e.into_inner())
                .is_some(),
            "sweep handle should be Some after poll_ready"
        );

        svc.shutdown(StepShutdownReason::HotSwap).await.unwrap();

        assert!(
            svc.sweep_handle
                .lock()
                .unwrap_or_else(|e| e.into_inner())
                .is_none(),
            "sweep handle should be None after HotSwap shutdown"
        );
        assert!(
            svc.sweep_cancel
                .lock()
                .unwrap_or_else(|e| e.into_inner())
                .is_cancelled(),
            "sweep_cancel token should be cancelled after HotSwap shutdown"
        );

        tokio::time::sleep(Duration::from_millis(50)).await;
    }

    // ── Task 4.1 (direct-inline-dispatch): constant-key memo fast path ──

    /// A constant scalar correlation key must serialize exactly ONCE: the
    /// first fragment fills the memo, later fragments reuse the memoized
    /// string via a scalar equality check. All fragments still land in ONE
    /// bucket (the size-3 completion with `CamelAggregatedSize == 3` proves
    /// a single bucket held all three).
    #[tokio::test]
    async fn constant_key_skips_reserialization() {
        let mut svc = new_test_svc(config_size(3));
        let mut result = None;
        for body in ["first", "second", "third"] {
            result = Some(
                svc.ready()
                    .await
                    .unwrap()
                    .call(make_exchange("orderId", "A", body))
                    .await
                    .unwrap(),
            );
        }
        let result = result.unwrap();
        assert_eq!(
            result.property(CAMEL_AGGREGATED_SIZE),
            Some(&serde_json::json!(3u64)),
            "all 3 fragments must aggregate out of a single bucket"
        );
        assert!(svc.buckets.lock().unwrap().is_empty());
        assert_eq!(
            svc.key_serializations
                .load(std::sync::atomic::Ordering::Relaxed),
            1,
            "constant scalar key must serialize exactly once"
        );
    }

    /// Alternating scalar keys keep byte-identical bucket names to direct
    /// `serde_json::to_string` calls: k1, k2, k1 → exactly two buckets named
    /// `"k1"` and `"k2"` (including the JSON quotes).
    #[tokio::test]
    async fn divergent_keys_keep_serde_semantics() {
        let mut svc = new_test_svc(config_size(10)); // never completes
        for key in ["k1", "k2", "k1"] {
            svc.ready()
                .await
                .unwrap()
                .call(make_exchange("orderId", key, "body"))
                .await
                .unwrap();
        }
        let guard = svc.buckets.lock().unwrap();
        assert_eq!(guard.len(), 2, "k1/k2/k1 → exactly two buckets");
        for key in ["k1", "k2"] {
            let expected = serde_json::to_string(&serde_json::json!(key)).unwrap();
            assert!(
                guard.contains_key(expected.as_str()),
                "bucket name must be byte-identical to serde_json::to_string: \
                 expected {expected}, have {:?}",
                guard.keys().collect::<Vec<_>>()
            );
        }
    }

    /// Object correlation keys bypass the memo entirely — every fragment
    /// serializes, and bucketing stays byte-identical to direct
    /// `serde_json::to_string` of the objects. The two fixtures share an
    /// equal key-set but are built in different insertion orders; the
    /// expected bucket-key set is DERIVED from serde_json itself so the pin
    /// holds under both order-canonicalizing (BTreeMap) and
    /// order-preserving (`preserve_order`) serde_json builds.
    #[tokio::test]
    async fn object_keys_bypass_cache() {
        let mut svc = new_test_svc(config_size(10)); // never completes
        let obj_a = serde_json::json!({"a": 1, "b": 2});
        let obj_b = serde_json::json!({"b": 2, "a": 1});
        for obj in [&obj_a, &obj_b, &obj_a] {
            let mut msg = Message {
                headers: Default::default(),
                body: Body::Text("body".into()),
            };
            msg.headers.insert("orderId".to_string(), obj.clone());
            svc.ready()
                .await
                .unwrap()
                .call(Exchange::new(msg))
                .await
                .unwrap();
        }
        let expected_keys: std::collections::HashSet<String> = [&obj_a, &obj_b]
            .into_iter()
            .map(|o| serde_json::to_string(o).unwrap())
            .collect();
        let guard = svc.buckets.lock().unwrap();
        let actual_keys: std::collections::HashSet<String> = guard.keys().cloned().collect();
        assert_eq!(
            actual_keys, expected_keys,
            "object keys must bucket exactly per serde_json::to_string (cache bypassed)"
        );
        assert_eq!(
            svc.key_serializations
                .load(std::sync::atomic::Ordering::Relaxed),
            3,
            "each object fragment must serialize (memo never consulted for objects)"
        );
    }

    /// Float ±0.0 regression: `json!(0.0) == json!(-0.0)` under `Value`
    /// equality, but serde serializes them as "0.0" and "-0.0". Numbers are
    /// NOT memoizable — every fragment serializes, and the two zero signs
    /// stay in distinct buckets named byte-identically to
    /// `serde_json::to_string` of each value.
    #[tokio::test]
    async fn float_zero_sign_keys_stay_distinct() {
        let mut svc = new_test_svc(config_size(10)); // never completes
        for key in [0.0_f64, -0.0, 0.0] {
            let mut msg = Message {
                headers: Default::default(),
                body: Body::Text("body".into()),
            };
            msg.headers
                .insert("orderId".to_string(), serde_json::json!(key));
            svc.ready()
                .await
                .unwrap()
                .call(Exchange::new(msg))
                .await
                .unwrap();
        }
        let guard = svc.buckets.lock().unwrap();
        assert_eq!(guard.len(), 2, "±0.0 must stay in distinct buckets");
        for key in [0.0_f64, -0.0] {
            let expected = serde_json::to_string(&serde_json::json!(key)).unwrap();
            assert!(
                guard.contains_key(expected.as_str()),
                "bucket name must be byte-identical to serde_json::to_string: \
                 expected {expected}, have {:?}",
                guard.keys().collect::<Vec<_>>()
            );
        }
        assert_eq!(
            svc.key_serializations
                .load(std::sync::atomic::Ordering::Relaxed),
            3,
            "number keys serialize per fragment (never memoized)"
        );
    }
}