awa_worker/

client.rs

use crate::completion::CompletionBatcher;
use crate::dispatcher::{
    shared_rate_limiter, ConcurrencyMode, Dispatcher, OverflowPool, QueueConfig,
};
use crate::events::{BoxedUntypedEventHandler, JobEvent, UntypedJobEvent};
use crate::executor::{BoxedWorker, DlqPolicy, JobError, JobExecutor, JobResult, Worker};
use crate::heartbeat::HeartbeatService;
use crate::maintenance::{MaintenanceService, RetentionPolicy};
use crate::runtime::{InFlightMap, InFlightRegistry};
use crate::storage::{QueueStorageRuntime, RuntimeStorage};
use awa_model::admin::{
    self, JobKindDescriptor, NamedJobKindDescriptor, NamedQueueDescriptor, QueueDescriptor,
    QueueRuntimeConfigSnapshot, QueueRuntimeMode, QueueRuntimeSnapshot, RateLimitSnapshot,
    RuntimeSnapshotInput, StorageCapability, TransitionRole,
};
use awa_model::{
    storage as transition, JobArgs, PartitionedQueue, PeriodicJob, QueueStorageConfig,
};
use chrono::{DateTime, Utc};
use serde::de::DeserializeOwned;
use sqlx::PgPool;
use std::any::{Any, TypeId};
use std::collections::{HashMap, HashSet};
use std::sync::atomic::{AtomicBool, AtomicU32, Ordering};
use std::sync::Arc;
use std::time::Duration;
use tokio::sync::{Mutex, RwLock};
use tokio::task::JoinSet;
use tokio_util::sync::CancellationToken;
use tracing::{info, warn};
use uuid::Uuid;

/// Errors returned when building a worker client.
#[derive(Debug, Clone, PartialEq, Eq, thiserror::Error)]
pub enum BuildError {
    #[error("at least one queue must be configured")]
    NoQueuesConfigured,
    #[error("queue descriptor declared for unknown queue '{queue}'")]
    QueueDescriptorWithoutQueue { queue: String },
    #[error("queue '{queue}' configured more than once")]
    DuplicateQueue { queue: String },
    #[error("sum of min_workers ({total_min}) exceeds global_max_workers ({global_max})")]
    MinWorkersExceedGlobal { total_min: u32, global_max: u32 },
    #[error("rate_limit max_rate must be > 0.0")]
    InvalidRateLimit,
    #[error("queue weight must be > 0")]
    InvalidWeight,
    #[error("queue claimers must be > 0")]
    InvalidClaimers,
    #[error("queue claim_batch_size must be > 0")]
    InvalidClaimBatchSize,
    #[error("cleanup_batch_size must be > 0")]
    InvalidBatchSize,
    #[error("dlq_cleanup_batch_size must be > 0")]
    InvalidDlqBatchSize,
    #[error("terminal_count_rollup_interval must be > 0")]
    InvalidTerminalCountRollupInterval,
    #[error("invalid queue storage config: {0}")]
    InvalidQueueStorage(String),
}

/// Health check result.
#[derive(Debug, Clone)]
pub struct HealthCheck {
    pub healthy: bool,
    pub postgres_connected: bool,
    pub poll_loop_alive: bool,
    pub heartbeat_alive: bool,
    pub maintenance_alive: bool,
    pub shutting_down: bool,
    pub leader: bool,
    pub queues: HashMap<String, QueueHealth>,
}

/// Per-queue health.
#[derive(Debug, Clone)]
pub struct QueueHealth {
    pub in_flight: u32,
    pub available: u64,
    /// Capacity interpretation depends on mode.
    pub capacity: QueueCapacity,
}

/// Capacity information for a queue, mode-dependent.
#[derive(Debug, Clone)]
pub enum QueueCapacity {
    /// Hard-reserved: fixed max.
    HardReserved { max_workers: u32 },
    /// Weighted: min guaranteed + current overflow.
    Weighted {
        min_workers: u32,
        weight: u32,
        overflow_held: u32,
    },
}

/// Temporary execution role used during a storage transition.
///
/// This is not intended as a long-term “run either backend forever” feature.
/// It exists so a `0.6` rollout can keep some runtimes draining canonical
/// backlog while other runtimes are already prepared to execute queue-storage
/// work as soon as routing flips.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
pub enum TransitionWorkerRole {
    /// Follow `awa.storage_status()`:
    /// canonical in `canonical` / `prepared`,
    /// queue storage in `mixed_transition` / `active`.
    #[default]
    Auto,
    /// Stay on canonical execution even after routing flips.
    CanonicalDrain,
    /// Run queue storage immediately, even before routing flips.
    QueueStorageTarget,
}

impl From<TransitionWorkerRole> for TransitionRole {
    fn from(role: TransitionWorkerRole) -> Self {
        match role {
            TransitionWorkerRole::Auto => Self::Auto,
            TransitionWorkerRole::CanonicalDrain => Self::CanonicalDrain,
            TransitionWorkerRole::QueueStorageTarget => Self::QueueStorageTarget,
        }
    }
}

/// Builder for the Awa worker client.
pub struct ClientBuilder {
    pool: PgPool,
    queues: Vec<(String, QueueConfig)>,
    queue_descriptors: HashMap<String, QueueDescriptor>,
    job_kind_descriptors: HashMap<String, JobKindDescriptor>,
    workers: HashMap<String, BoxedWorker>,
    lifecycle_handlers: HashMap<String, Vec<BoxedUntypedEventHandler>>,
    enqueue_specs: HashMap<
        crate::enqueue_specs::Outcome,
        HashMap<String, Vec<crate::enqueue_specs::BoxedEnqueueSpec>>,
    >,
    state: HashMap<TypeId, Box<dyn Any + Send + Sync>>,
    heartbeat_interval: Duration,
    promote_interval: Duration,
    heartbeat_rescue_interval: Option<Duration>,
    heartbeat_staleness: Option<Duration>,
    deadline_rescue_interval: Option<Duration>,
    callback_rescue_interval: Option<Duration>,
    periodic_jobs: Vec<PeriodicJob>,
    global_max_workers: Option<u32>,
    leader_election_interval: Option<Duration>,
    leader_check_interval: Option<Duration>,
    priority_aging_interval: Option<Duration>,
    terminal_count_rollup_interval: Option<Duration>,
    completed_retention: Option<Duration>,
    failed_retention: Option<Duration>,
    descriptor_retention: Option<Duration>,
    cleanup_batch_size: Option<i64>,
    cleanup_interval: Option<Duration>,
    queue_retention_overrides: HashMap<String, RetentionPolicy>,
    runtime_snapshot_interval: Duration,
    queue_stats_interval: Option<Duration>,
    dlq_enabled_by_default: bool,
    dlq_retention: Option<Duration>,
    dlq_cleanup_batch_size: Option<i64>,
    dlq_overrides: HashMap<String, bool>,
    storage: RuntimeStorage,
    transition_role: TransitionWorkerRole,
    storage_error: Option<BuildError>,
}

impl ClientBuilder {
    pub fn new(pool: PgPool) -> Self {
        // #169: keep lease rotation on the same conservative cadence as
        // queue / claim rotation. Faster lease rotation only advances a
        // one-row metadata cursor more often; under a pinned MVCC horizon
        // that metadata churn can dominate the otherwise append-only queue
        // path without improving job safety.
        let (storage, storage_error) = match QueueStorageRuntime::new(
            QueueStorageConfig::default(),
            Duration::from_millis(1_000),
            Duration::from_millis(1_000),
        ) {
            Ok(runtime) => (RuntimeStorage::QueueStorage(runtime), None),
            Err(err) => (
                RuntimeStorage::Canonical,
                Some(BuildError::InvalidQueueStorage(err.to_string())),
            ),
        };

        Self {
            pool,
            queues: Vec::new(),
            queue_descriptors: HashMap::new(),
            job_kind_descriptors: HashMap::new(),
            workers: HashMap::new(),
            lifecycle_handlers: HashMap::new(),
            enqueue_specs: HashMap::new(),
            state: HashMap::new(),
            heartbeat_interval: Duration::from_secs(30),
            promote_interval: Duration::from_millis(250),
            heartbeat_rescue_interval: None,
            heartbeat_staleness: None,
            deadline_rescue_interval: None,
            callback_rescue_interval: None,
            periodic_jobs: Vec::new(),
            global_max_workers: None,
            leader_election_interval: None,
            leader_check_interval: None,
            priority_aging_interval: None,
            terminal_count_rollup_interval: None,
            completed_retention: None,
            failed_retention: None,
            descriptor_retention: None,
            cleanup_batch_size: None,
            cleanup_interval: None,
            queue_retention_overrides: HashMap::new(),
            runtime_snapshot_interval: Duration::from_secs(10),
            queue_stats_interval: None,
            dlq_enabled_by_default: false,
            dlq_retention: None,
            dlq_cleanup_batch_size: None,
            dlq_overrides: HashMap::new(),
            storage,
            transition_role: TransitionWorkerRole::Auto,
            storage_error,
        }
    }

    /// Add a queue with its configuration.
    pub fn queue(mut self, name: impl Into<String>, config: QueueConfig) -> Self {
        self.queues.push((name.into(), config));
        self
    }

    /// Add every physical queue in a logical partitioned queue.
    ///
    /// This is equivalent to calling [`queue`] once for each physical queue in
    /// the partitioned queue. Producers should route inserts through the same
    /// [`PartitionedQueue`] so workers and producers agree on the physical queue
    /// names.
    ///
    /// The `config` is applied to each physical queue. In hard-reserved mode,
    /// total logical capacity is therefore roughly
    /// `partitioned_queue.partitions() * config.max_workers`; per-queue rate
    /// limits also apply per physical
    /// queue. Divide those knobs yourself, or use weighted mode with a
    /// `global_max_workers` cap, when you need a logical total.
    ///
    /// [`queue`]: ClientBuilder::queue
    pub fn partitioned_queue(
        mut self,
        partitioned_queue: &PartitionedQueue,
        config: QueueConfig,
    ) -> Self {
        for queue in partitioned_queue.physical_queues() {
            self.queues.push((queue.clone(), config.clone()));
        }
        self
    }

    /// Attach descriptive metadata (display name, description, owner,
    /// docs URL, tags, extra JSON) to a queue so it appears labelled in
    /// the admin API and UI. The queue must also be declared via
    /// [`queue`]; otherwise [`build`] fails with
    /// [`BuildError::QueueDescriptorWithoutQueue`].
    ///
    /// [`queue`]: ClientBuilder::queue
    /// [`build`]: ClientBuilder::build
    pub fn queue_descriptor(
        mut self,
        name: impl Into<String>,
        descriptor: QueueDescriptor,
    ) -> Self {
        self.queue_descriptors.insert(name.into(), descriptor);
        self
    }

    /// Attach descriptive metadata to a typed job kind. The kind string is
    /// taken from [`JobArgs::kind`] on `T`.
    pub fn job_kind_descriptor<T: JobArgs>(mut self, descriptor: JobKindDescriptor) -> Self {
        self.job_kind_descriptors
            .insert(T::kind().to_string(), descriptor);
        self
    }

    /// Attach descriptive metadata to a job kind by string name. Useful
    /// when the kind is known dynamically (e.g. from language bridges).
    pub fn job_kind_descriptor_kind(
        mut self,
        kind: impl Into<String>,
        descriptor: JobKindDescriptor,
    ) -> Self {
        self.job_kind_descriptors.insert(kind.into(), descriptor);
        self
    }

    /// Register a typed worker.
    ///
    /// The worker handles jobs of type `T` where `T: JobArgs + DeserializeOwned`.
    /// The handler function receives the deserialized args and job context.
    pub fn register<T, F, Fut>(mut self, handler: F) -> Self
    where
        T: JobArgs + DeserializeOwned + Send + Sync + 'static,
        F: Fn(T, &crate::context::JobContext) -> Fut + Send + Sync + 'static,
        Fut: std::future::Future<Output = Result<JobResult, JobError>> + Send + 'static,
    {
        let kind = T::kind().to_string();
        let worker = TypedWorker {
            kind: T::kind(),
            handler: Arc::new(handler),
            _phantom: std::marker::PhantomData,
        };
        self.workers.insert(kind, Box::new(worker));
        self
    }

    /// Register a typed lifecycle event handler for a job kind.
    ///
    /// `Started` dispatch is scheduled after the claim commits and before the
    /// worker handler is invoked. Outcome handlers run only after the
    /// corresponding DB state transition commits. Hooks are best-effort
    /// in-process notifications, not a durable workflow mechanism; because
    /// they run in detached tasks, a very short job can complete before a
    /// `Started` handler finishes. Capture any shared dependencies you need in
    /// the closure environment.
    pub fn on_event<T, F, Fut>(mut self, handler: F) -> Self
    where
        T: JobArgs + DeserializeOwned + Send + Sync + 'static,
        F: Fn(JobEvent<T>) -> Fut + Send + Sync + 'static,
        Fut: std::future::Future<Output = ()> + Send + 'static,
    {
        let kind = T::kind().to_string();
        let handler = Arc::new(handler);
        let erased: BoxedUntypedEventHandler = Arc::new(move |event: UntypedJobEvent| {
            let handler = handler.clone();
            Box::pin(async move {
                let args: T = match serde_json::from_value(event.job().args.clone()) {
                    Ok(args) => args,
                    Err(err) => {
                        warn!(
                            job_id = event.job().id,
                            kind = %event.job().kind,
                            error = %err,
                            "Failed to deserialize args for lifecycle event handler"
                        );
                        return;
                    }
                };

                (handler)(event.into_typed(args)).await;
            })
        });
        self.lifecycle_handlers
            .entry(kind)
            .or_default()
            .push(erased);
        self
    }

    /// Register an untyped lifecycle event handler for a specific job kind.
    ///
    /// Use this with `register_worker(...)` or for cross-cutting logic that
    /// doesn't need typed args. Timing matches `on_event(...)`.
    pub fn on_event_kind<F, Fut>(mut self, kind: impl Into<String>, handler: F) -> Self
    where
        F: Fn(UntypedJobEvent) -> Fut + Send + Sync + 'static,
        Fut: std::future::Future<Output = ()> + Send + 'static,
    {
        let kind = kind.into();
        let handler = Arc::new(handler);
        let erased: BoxedUntypedEventHandler = Arc::new(move |event: UntypedJobEvent| {
            let handler = handler.clone();
            Box::pin(async move {
                (handler)(event).await;
            })
        });
        self.lifecycle_handlers
            .entry(kind)
            .or_default()
            .push(erased);
        self
    }

    /// Register a durable follow-up Awa job to enqueue when a job of type `T`
    /// completes successfully.
    ///
    /// `make` receives the trigger's deserialised args plus its post-completion
    /// [`awa_model::JobRow`] and returns the follow-up's `JobArgs` value. The
    /// follow-up is enqueued with default [`awa_model::InsertOpts`] — use
    /// [`Self::on_completed_enqueue_with`] to override queue, priority, or
    /// other insert options.
    ///
    /// # Atomicity
    ///
    /// Whether the trigger completes from the worker handler
    /// (`Ok(Completed)`) or via callback resolution
    /// ([`Client::complete_external`], [`Client::resolve_callback`] with
    /// a `Complete` action), the follow-up `INSERT`s in the *same
    /// database transaction* as the completion's state change. The
    /// trigger and the follow-up commit or roll back together (ADR-029,
    /// ADR-013 run-lease guard). A spec INSERT failure or a panic in
    /// this closure rolls the completion back as well, so a failed
    /// callback can be redelivered by the external sender rather than
    /// leaving the job half-applied.
    ///
    /// This is the durable counterpart to [`Self::on_event`]: hooks are
    /// best-effort, in-process; this enqueue is at-least-once and rides
    /// Awa's existing retry/DLQ machinery.
    ///
    /// Multiple registrations for the same kind stack and all run in the
    /// same dispatch.
    pub fn on_completed_enqueue<T, F, MakeFn>(self, make: MakeFn) -> Self
    where
        T: JobArgs + DeserializeOwned + Send + Sync + 'static,
        F: JobArgs + Send + Sync + 'static,
        MakeFn: Fn(T, &awa_model::JobRow) -> F + Send + Sync + 'static,
    {
        self.on_completed_enqueue_with::<T, F, _>(move |args, job| {
            crate::enqueue_specs::EnqueueRequest::new(make(args, job))
        })
    }

    /// Like [`Self::on_completed_enqueue`] but lets the closure return an
    /// [`EnqueueRequest`](crate::EnqueueRequest) with explicit queue,
    /// priority, or other [`awa_model::InsertOpts`] overrides.
    pub fn on_completed_enqueue_with<T, F, MakeFn>(mut self, make: MakeFn) -> Self
    where
        T: JobArgs + DeserializeOwned + Send + Sync + 'static,
        F: JobArgs + Send + Sync + 'static,
        MakeFn: Fn(T, &awa_model::JobRow) -> crate::enqueue_specs::EnqueueRequest<F>
            + Send
            + Sync
            + 'static,
    {
        let kind = T::kind().to_string();
        let spec: crate::enqueue_specs::BoxedEnqueueSpec =
            Arc::new(crate::enqueue_specs::CompletedFollowUp::<T, F, _> {
                make,
                _phantom: std::marker::PhantomData,
            });
        self.enqueue_specs
            .entry(crate::enqueue_specs::Outcome::Completed)
            .or_default()
            .entry(kind)
            .or_default()
            .push(spec);
        self
    }

    /// Register a durable follow-up Awa job to enqueue when a job of type `T`
    /// is retried (whether via `Ok(RetryAfter)` or a retryable `Err`).
    /// `make` receives the trigger's args, its post-retry `JobRow`, the error
    /// string, the previous attempt number, and the next-run-at timestamp.
    ///
    /// Atomicity matches [`Self::on_completed_enqueue`]: worker-driven
    /// retries dispatch in the transition's transaction, and a retry
    /// driven by [`Client::retry_external`] dispatches in the same
    /// transaction as the resolution. A spec failure rolls the retry
    /// transition back so the external sender can redeliver.
    pub fn on_retried_enqueue<T, F, MakeFn>(self, make: MakeFn) -> Self
    where
        T: JobArgs + DeserializeOwned + Send + Sync + 'static,
        F: JobArgs + Send + Sync + 'static,
        MakeFn: Fn(T, &awa_model::JobRow, &str, i16, chrono::DateTime<chrono::Utc>) -> F
            + Send
            + Sync
            + 'static,
    {
        self.on_retried_enqueue_with::<T, F, _>(move |args, job, error, attempt, next_run_at| {
            crate::enqueue_specs::EnqueueRequest::new(make(args, job, error, attempt, next_run_at))
        })
    }

    /// Like [`Self::on_retried_enqueue`] but the closure returns an
    /// [`EnqueueRequest`](crate::EnqueueRequest) with `InsertOpts` overrides.
    pub fn on_retried_enqueue_with<T, F, MakeFn>(mut self, make: MakeFn) -> Self
    where
        T: JobArgs + DeserializeOwned + Send + Sync + 'static,
        F: JobArgs + Send + Sync + 'static,
        MakeFn: Fn(
                T,
                &awa_model::JobRow,
                &str,
                i16,
                chrono::DateTime<chrono::Utc>,
            ) -> crate::enqueue_specs::EnqueueRequest<F>
            + Send
            + Sync
            + 'static,
    {
        let kind = T::kind().to_string();
        let spec: crate::enqueue_specs::BoxedEnqueueSpec =
            Arc::new(crate::enqueue_specs::RetriedFollowUp::<T, F, _> {
                make,
                _phantom: std::marker::PhantomData,
            });
        self.enqueue_specs
            .entry(crate::enqueue_specs::Outcome::Retried)
            .or_default()
            .entry(kind)
            .or_default()
            .push(spec);
        self
    }

    /// Register a durable follow-up Awa job to enqueue when a job of type `T`
    /// is exhausted (retries used up *or* `Err(JobError::Terminal)`).
    /// `make` receives the trigger's args, its post-failure `JobRow`, the
    /// error string, and the attempt number.
    ///
    /// Atomicity matches [`Self::on_completed_enqueue`]: worker-driven
    /// exhaustion dispatches in the transition's transaction, and an
    /// exhaustion driven by [`Client::fail_external`] or
    /// [`Client::resolve_callback`] with a `Fail` action dispatches in
    /// the same transaction as the resolution. A spec failure rolls the
    /// failure transition back so the external sender can redeliver.
    pub fn on_exhausted_enqueue<T, F, MakeFn>(self, make: MakeFn) -> Self
    where
        T: JobArgs + DeserializeOwned + Send + Sync + 'static,
        F: JobArgs + Send + Sync + 'static,
        MakeFn: Fn(T, &awa_model::JobRow, &str, i16) -> F + Send + Sync + 'static,
    {
        self.on_exhausted_enqueue_with::<T, F, _>(move |args, job, error, attempt| {
            crate::enqueue_specs::EnqueueRequest::new(make(args, job, error, attempt))
        })
    }

    /// Like [`Self::on_exhausted_enqueue`] but the closure returns an
    /// [`EnqueueRequest`](crate::EnqueueRequest) with `InsertOpts` overrides.
    pub fn on_exhausted_enqueue_with<T, F, MakeFn>(mut self, make: MakeFn) -> Self
    where
        T: JobArgs + DeserializeOwned + Send + Sync + 'static,
        F: JobArgs + Send + Sync + 'static,
        MakeFn: Fn(T, &awa_model::JobRow, &str, i16) -> crate::enqueue_specs::EnqueueRequest<F>
            + Send
            + Sync
            + 'static,
    {
        let kind = T::kind().to_string();
        let spec: crate::enqueue_specs::BoxedEnqueueSpec =
            Arc::new(crate::enqueue_specs::ExhaustedFollowUp::<T, F, _> {
                make,
                _phantom: std::marker::PhantomData,
            });
        self.enqueue_specs
            .entry(crate::enqueue_specs::Outcome::Exhausted)
            .or_default()
            .entry(kind)
            .or_default()
            .push(spec);
        self
    }

    /// Register a durable follow-up Awa job to enqueue when a job of type `T`
    /// is cancelled via `Ok(JobResult::Cancel(reason))`. `make` receives the
    /// trigger's args, its post-cancellation `JobRow`, and the reason string.
    ///
    /// See [`Self::on_completed_enqueue`] for the same-transaction semantics.
    pub fn on_cancelled_enqueue<T, F, MakeFn>(self, make: MakeFn) -> Self
    where
        T: JobArgs + DeserializeOwned + Send + Sync + 'static,
        F: JobArgs + Send + Sync + 'static,
        MakeFn: Fn(T, &awa_model::JobRow, &str) -> F + Send + Sync + 'static,
    {
        self.on_cancelled_enqueue_with::<T, F, _>(move |args, job, reason| {
            crate::enqueue_specs::EnqueueRequest::new(make(args, job, reason))
        })
    }

    /// Like [`Self::on_cancelled_enqueue`] but the closure returns an
    /// [`EnqueueRequest`](crate::EnqueueRequest) with `InsertOpts` overrides.
    pub fn on_cancelled_enqueue_with<T, F, MakeFn>(mut self, make: MakeFn) -> Self
    where
        T: JobArgs + DeserializeOwned + Send + Sync + 'static,
        F: JobArgs + Send + Sync + 'static,
        MakeFn: Fn(T, &awa_model::JobRow, &str) -> crate::enqueue_specs::EnqueueRequest<F>
            + Send
            + Sync
            + 'static,
    {
        let kind = T::kind().to_string();
        let spec: crate::enqueue_specs::BoxedEnqueueSpec =
            Arc::new(crate::enqueue_specs::CancelledFollowUp::<T, F, _> {
                make,
                _phantom: std::marker::PhantomData,
            });
        self.enqueue_specs
            .entry(crate::enqueue_specs::Outcome::Cancelled)
            .or_default()
            .entry(kind)
            .or_default()
            .push(spec);
        self
    }

    /// Register a durable follow-up Awa job to enqueue when a job of type `T`
    /// parks on an external callback (`Ok(JobResult::WaitForCallback)`).
    /// `make` receives the trigger's args plus the parked `JobRow`
    /// (`job.callback_id` and `job.callback_timeout_at` identify the
    /// pending callback).
    ///
    /// See [`Self::on_completed_enqueue`] for the same-transaction semantics.
    pub fn on_waiting_for_callback_enqueue<T, F, MakeFn>(self, make: MakeFn) -> Self
    where
        T: JobArgs + DeserializeOwned + Send + Sync + 'static,
        F: JobArgs + Send + Sync + 'static,
        MakeFn: Fn(T, &awa_model::JobRow) -> F + Send + Sync + 'static,
    {
        self.on_waiting_for_callback_enqueue_with::<T, F, _>(move |args, job| {
            crate::enqueue_specs::EnqueueRequest::new(make(args, job))
        })
    }

    /// Like [`Self::on_waiting_for_callback_enqueue`] but the closure returns
    /// an [`EnqueueRequest`](crate::EnqueueRequest) with `InsertOpts` overrides.
    pub fn on_waiting_for_callback_enqueue_with<T, F, MakeFn>(mut self, make: MakeFn) -> Self
    where
        T: JobArgs + DeserializeOwned + Send + Sync + 'static,
        F: JobArgs + Send + Sync + 'static,
        MakeFn: Fn(T, &awa_model::JobRow) -> crate::enqueue_specs::EnqueueRequest<F>
            + Send
            + Sync
            + 'static,
    {
        let kind = T::kind().to_string();
        let spec: crate::enqueue_specs::BoxedEnqueueSpec = Arc::new(
            crate::enqueue_specs::WaitingForCallbackFollowUp::<T, F, _> {
                make,
                _phantom: std::marker::PhantomData,
            },
        );
        self.enqueue_specs
            .entry(crate::enqueue_specs::Outcome::WaitingForCallback)
            .or_default()
            .entry(kind)
            .or_default()
            .push(spec);
        self
    }

    /// Register a durable follow-up Awa job to enqueue when a job of type `T`
    /// is rescued by maintenance (expired callback, stale heartbeat, or
    /// deadline exceeded). `make` receives the trigger's args, its post-rescue
    /// `JobRow`, and the [`RescueReason`](crate::events::RescueReason).
    ///
    /// **Atomicity: best-effort, separate transaction.** Maintenance rescues
    /// commit the rescue UPDATE before this dispatcher runs; the follow-up
    /// dispatch opens its own transaction afterwards. If the spec INSERT
    /// fails the rescue stands and the failure is logged. Don't predicate
    /// a workflow on a rescue-driven follow-up landing — for compensation
    /// after a stuck attempt, build the follow-up handler to be safely
    /// re-runnable. The atomic variant is tracked as an open extension in
    /// ADR-029.
    pub fn on_rescued_enqueue<T, F, MakeFn>(self, make: MakeFn) -> Self
    where
        T: JobArgs + DeserializeOwned + Send + Sync + 'static,
        F: JobArgs + Send + Sync + 'static,
        MakeFn: Fn(T, &awa_model::JobRow, crate::events::RescueReason) -> F + Send + Sync + 'static,
    {
        self.on_rescued_enqueue_with::<T, F, _>(move |args, job, reason| {
            crate::enqueue_specs::EnqueueRequest::new(make(args, job, reason))
        })
    }

    /// Like [`Self::on_rescued_enqueue`] but the closure returns an
    /// [`EnqueueRequest`](crate::EnqueueRequest) with `InsertOpts` overrides.
    pub fn on_rescued_enqueue_with<T, F, MakeFn>(mut self, make: MakeFn) -> Self
    where
        T: JobArgs + DeserializeOwned + Send + Sync + 'static,
        F: JobArgs + Send + Sync + 'static,
        MakeFn: Fn(
                T,
                &awa_model::JobRow,
                crate::events::RescueReason,
            ) -> crate::enqueue_specs::EnqueueRequest<F>
            + Send
            + Sync
            + 'static,
    {
        let kind = T::kind().to_string();
        let spec: crate::enqueue_specs::BoxedEnqueueSpec =
            Arc::new(crate::enqueue_specs::RescuedFollowUp::<T, F, _> {
                make,
                _phantom: std::marker::PhantomData,
            });
        self.enqueue_specs
            .entry(crate::enqueue_specs::Outcome::Rescued)
            .or_default()
            .entry(kind)
            .or_default()
            .push(spec);
        self
    }

    /// Register a raw worker implementation.
    pub fn register_worker(mut self, worker: impl Worker + 'static) -> Self {
        let kind = worker.kind().to_string();
        self.workers.insert(kind, Box::new(worker));
        self
    }

    /// Register an HTTP worker that dispatches jobs to a remote endpoint.
    ///
    /// In async mode the worker POSTs the job and parks in `waiting_external`.
    /// In sync mode the worker awaits the HTTP response directly.
    ///
    /// Requires the `http-worker` feature.
    #[cfg(feature = "http-worker")]
    pub fn http_worker(
        self,
        kind: impl Into<String>,
        config: crate::http_worker::HttpWorkerConfig,
    ) -> Self {
        let worker = crate::http_worker::HttpWorker::new(kind.into(), config);
        self.register_worker(worker)
    }

    /// Add shared state accessible via `ctx.extract::<T>()`.
    pub fn state<T: Any + Send + Sync + Clone>(mut self, value: T) -> Self {
        self.state.insert(TypeId::of::<T>(), Box::new(value));
        self
    }

    /// Set the heartbeat interval (default: 30s).
    pub fn heartbeat_interval(mut self, interval: Duration) -> Self {
        self.heartbeat_interval = interval;
        self
    }

    /// Set the scheduled/retryable promotion interval (default: 250ms).
    pub fn promote_interval(mut self, interval: Duration) -> Self {
        self.promote_interval = interval;
        self
    }

    /// Set the stale-heartbeat rescue interval (default: 30s).
    pub fn heartbeat_rescue_interval(mut self, interval: Duration) -> Self {
        self.heartbeat_rescue_interval = Some(interval);
        self
    }

    /// Set how long a heartbeat must be stale before the job is rescued (default: 90s).
    ///
    /// Should be at least 3× the heartbeat interval to avoid false rescues.
    pub fn heartbeat_staleness(mut self, staleness: Duration) -> Self {
        self.heartbeat_staleness = Some(staleness);
        self
    }

    /// Set the deadline rescue interval (default: 30s).
    pub fn deadline_rescue_interval(mut self, interval: Duration) -> Self {
        self.deadline_rescue_interval = Some(interval);
        self
    }

    /// Set the callback-timeout rescue interval (default: 30s).
    pub fn callback_rescue_interval(mut self, interval: Duration) -> Self {
        self.callback_rescue_interval = Some(interval);
        self
    }

    /// Set the leader election retry interval (default: 10s).
    ///
    /// Controls how often a non-leader instance retries acquiring the maintenance
    /// advisory lock. Lower values are useful in tests.
    pub fn leader_election_interval(mut self, interval: Duration) -> Self {
        self.leader_election_interval = Some(interval);
        self
    }

    /// Set the leader connection health-check interval (default: 30s).
    pub fn leader_check_interval(mut self, interval: Duration) -> Self {
        self.leader_check_interval = Some(interval);
        self
    }

    /// Set a global maximum worker count across all queues (enables weighted mode).
    ///
    /// When set, each queue gets `min_workers` guaranteed permits plus a share
    /// of the remaining overflow capacity based on `weight`.
    pub fn global_max_workers(mut self, max: u32) -> Self {
        self.global_max_workers = Some(max);
        self
    }

    /// Set retention for completed jobs (default: 24h).
    pub fn completed_retention(mut self, retention: Duration) -> Self {
        self.completed_retention = Some(retention);
        self
    }

    /// Set retention for failed/cancelled jobs (default: 72h).
    pub fn failed_retention(mut self, retention: Duration) -> Self {
        self.failed_retention = Some(retention);
        self
    }

    /// How long a descriptor catalog row can go un-refreshed before the
    /// maintenance leader deletes it (default: 30 days). Pass
    /// `Duration::ZERO` to disable — the catalog will then accumulate
    /// rows indefinitely. See [`MaintenanceService::descriptor_retention`].
    pub fn descriptor_retention(mut self, retention: Duration) -> Self {
        self.descriptor_retention = Some(retention);
        self
    }

    /// Set the maximum number of jobs to delete per cleanup pass (default: 1000).
    pub fn cleanup_batch_size(mut self, batch_size: i64) -> Self {
        self.cleanup_batch_size = Some(batch_size);
        self
    }

    /// Set the cleanup interval (default: 60s).
    pub fn cleanup_interval(mut self, interval: Duration) -> Self {
        self.cleanup_interval = Some(interval);
        self
    }

    /// Set a per-queue retention override.
    pub fn queue_retention(mut self, queue: impl Into<String>, policy: RetentionPolicy) -> Self {
        self.queue_retention_overrides.insert(queue.into(), policy);
        self
    }

    /// Set how often runtime observability snapshots are published (default: 10s).
    pub fn runtime_snapshot_interval(mut self, interval: Duration) -> Self {
        self.runtime_snapshot_interval = interval;
        self
    }

    /// Set the maintenance priority aging interval.
    ///
    /// This controls how often waiting available jobs are promoted toward
    /// higher priority to prevent starvation. It is a global maintenance
    /// setting for this worker runtime.
    pub fn priority_aging_interval(mut self, interval: Duration) -> Self {
        self.priority_aging_interval = Some(interval);
        self
    }

    /// Set how often queue-storage terminal-count deltas are rolled into
    /// folded live counters (default: 30s).
    ///
    /// Exact queue-depth reads include both folded counters and pending
    /// deltas, so increasing this interval trades a larger append-only delta
    /// ledger for fewer maintenance writes to the mutable counter table.
    pub fn terminal_count_rollup_interval(mut self, interval: Duration) -> Self {
        self.terminal_count_rollup_interval = Some(interval);
        self
    }

    /// Set how often queue depth/lag metrics are published (default: 30s).
    pub fn queue_stats_interval(mut self, interval: Duration) -> Self {
        self.queue_stats_interval = Some(interval);
        self
    }

    /// Enable or disable DLQ routing by default.
    pub fn dlq_enabled_by_default(mut self, enabled: bool) -> Self {
        self.dlq_enabled_by_default = enabled;
        self
    }

    /// Override DLQ routing for a single queue.
    pub fn queue_dlq_enabled(mut self, queue: impl Into<String>, enabled: bool) -> Self {
        self.dlq_overrides.insert(queue.into(), enabled);
        self
    }

    /// Set retention for DLQ rows.
    pub fn dlq_retention(mut self, retention: Duration) -> Self {
        self.dlq_retention = Some(retention);
        self
    }

    /// Set the maximum number of DLQ rows deleted per cleanup pass.
    pub fn dlq_cleanup_batch_size(mut self, batch_size: i64) -> Self {
        self.dlq_cleanup_batch_size = Some(batch_size);
        self
    }

    /// Override the segmented queue storage configuration for this runtime.
    ///
    /// Queue storage is the default worker engine. Canonical tables remain
    /// migration compatibility, not a second supported worker runtime.
    /// Use this to change the schema name or rotation sizing/timing.
    pub fn queue_storage(
        mut self,
        config: QueueStorageConfig,
        queue_rotate_interval: Duration,
        lease_rotate_interval: Duration,
    ) -> Self {
        match QueueStorageRuntime::new(config, queue_rotate_interval, lease_rotate_interval) {
            Ok(runtime) => {
                self.storage = RuntimeStorage::QueueStorage(runtime);
                self.storage_error = None;
            }
            Err(err) => {
                self.storage = RuntimeStorage::Canonical;
                self.storage_error = Some(BuildError::InvalidQueueStorage(err.to_string()));
            }
        }
        self
    }

    /// Override the ADR-023 claim-ring rotation cadence.
    ///
    /// Defaults to `queue_rotate_interval` so claim partitions age out in
    /// step with the ready / done partitions they reference. Only takes
    /// effect when queue storage is active; no-op on the canonical engine.
    pub fn claim_rotate_interval(mut self, claim_rotate_interval: Duration) -> Self {
        if let RuntimeStorage::QueueStorage(runtime) = self.storage {
            self.storage = RuntimeStorage::QueueStorage(
                runtime.with_claim_rotate_interval(claim_rotate_interval),
            );
        }
        self
    }

    /// Force the worker runtime onto canonical storage.
    ///
    /// This is primarily useful for migration/testing and benchmark
    /// comparisons against the pre-0.6 engine. Production 0.6 runtimes should
    /// normally use queue storage.
    pub fn canonical_storage(mut self) -> Self {
        self.storage = RuntimeStorage::Canonical;
        self.storage_error = None;
        self
    }

    /// Choose how this runtime participates in a storage transition.
    pub fn transition_role(mut self, role: TransitionWorkerRole) -> Self {
        self.transition_role = role;
        self
    }

    /// Register a periodic (cron) job schedule.
    ///
    /// The schedule is synced to the database by the leader and evaluated
    /// every second. When a fire is due, a job is atomically enqueued.
    pub fn periodic(mut self, job: PeriodicJob) -> Self {
        self.periodic_jobs.push(job);
        self
    }

    /// Build the client.
    pub fn build(self) -> Result<Client, BuildError> {
        if self.queues.is_empty() {
            return Err(BuildError::NoQueuesConfigured);
        }

        if let Some(err) = self.storage_error.clone() {
            return Err(err);
        }

        let mut queue_names = HashSet::with_capacity(self.queues.len());
        for (queue, _) in &self.queues {
            if !queue_names.insert(queue.as_str()) {
                return Err(BuildError::DuplicateQueue {
                    queue: queue.clone(),
                });
            }
        }

        for queue in self.queue_descriptors.keys() {
            if !self.queues.iter().any(|(name, _)| name == queue) {
                return Err(BuildError::QueueDescriptorWithoutQueue {
                    queue: queue.clone(),
                });
            }
        }

        // Validate rate limits and weights
        for (_, config) in &self.queues {
            if let Some(rl) = &config.rate_limit {
                if rl.max_rate <= 0.0 {
                    return Err(BuildError::InvalidRateLimit);
                }
            }
            if config.weight == 0 {
                return Err(BuildError::InvalidWeight);
            }
            if config.claimers == 0 {
                return Err(BuildError::InvalidClaimers);
            }
            if config.claim_batch_size == 0 {
                return Err(BuildError::InvalidClaimBatchSize);
            }
        }

        // Validate batch size
        if let Some(bs) = self.cleanup_batch_size {
            if bs <= 0 {
                return Err(BuildError::InvalidBatchSize);
            }
        }
        if let Some(bs) = self.dlq_cleanup_batch_size {
            if bs <= 0 {
                return Err(BuildError::InvalidDlqBatchSize);
            }
        }
        if self
            .terminal_count_rollup_interval
            .is_some_and(|interval| interval.is_zero())
        {
            return Err(BuildError::InvalidTerminalCountRollupInterval);
        }

        // Validate weighted mode constraints
        let overflow_pool = if let Some(global_max) = self.global_max_workers {
            let total_min: u32 = self.queues.iter().map(|(_, c)| c.min_workers).sum();
            if total_min > global_max {
                return Err(BuildError::MinWorkersExceedGlobal {
                    total_min,
                    global_max,
                });
            }
            let overflow_capacity = global_max - total_min;
            let weights: HashMap<String, u32> = self
                .queues
                .iter()
                .map(|(name, c)| (name.clone(), c.weight.max(1)))
                .collect();
            Some(Arc::new(OverflowPool::new(overflow_capacity, weights)))
        } else {
            None
        };

        // Warn if heartbeat_staleness is less than 3× heartbeat_interval
        if let Some(staleness) = self.heartbeat_staleness {
            let min_safe = self.heartbeat_interval * 3;
            if staleness < min_safe {
                tracing::warn!(
                    heartbeat_staleness_ms = staleness.as_millis() as u64,
                    heartbeat_interval_ms = self.heartbeat_interval.as_millis() as u64,
                    recommended_min_ms = min_safe.as_millis() as u64,
                    "heartbeat_staleness ({:?}) is less than 3× heartbeat_interval ({:?}); \
                     this may cause false rescues of jobs that are still running",
                    staleness,
                    self.heartbeat_interval,
                );
            }
        }

        let metrics = crate::metrics::AwaMetrics::from_global();
        let queue_in_flight = Arc::new(
            self.queues
                .iter()
                .map(|(name, _)| (name.clone(), Arc::new(AtomicU32::new(0))))
                .collect(),
        );
        let dispatcher_alive = Arc::new(
            self.queues
                .iter()
                .map(|(name, _)| (name.clone(), Arc::new(AtomicBool::new(false))))
                .collect(),
        );
        let dlq_policy = DlqPolicy::new(self.dlq_enabled_by_default, self.dlq_overrides);

        Ok(Client {
            pool: self.pool,
            queues: self.queues,
            queue_descriptors: self.queue_descriptors,
            job_kind_descriptors: self.job_kind_descriptors,
            workers: Arc::new(self.workers),
            lifecycle_handlers: Arc::new(self.lifecycle_handlers),
            enqueue_specs: Arc::new(self.enqueue_specs),
            state: Arc::new(self.state),
            heartbeat_interval: self.heartbeat_interval,
            promote_interval: self.promote_interval,
            heartbeat_rescue_interval: self.heartbeat_rescue_interval,
            heartbeat_staleness: self.heartbeat_staleness,
            deadline_rescue_interval: self.deadline_rescue_interval,
            callback_rescue_interval: self.callback_rescue_interval,
            periodic_jobs: Arc::new(self.periodic_jobs),
            dispatch_cancel: CancellationToken::new(),
            service_cancel: CancellationToken::new(),
            dispatcher_handles: RwLock::new(Vec::new()),
            service_handles: RwLock::new(Vec::new()),
            job_set: Arc::new(Mutex::new(JoinSet::new())),
            in_flight: Arc::new(InFlightRegistry::default()),
            queue_in_flight,
            dispatcher_alive,
            heartbeat_alive: Arc::new(AtomicBool::new(false)),
            maintenance_alive: Arc::new(AtomicBool::new(false)),
            leader: Arc::new(AtomicBool::new(false)),
            overflow_pool,
            metrics,
            leader_election_interval: self.leader_election_interval,
            leader_check_interval: self.leader_check_interval,
            priority_aging_interval: self.priority_aging_interval,
            terminal_count_rollup_interval: self.terminal_count_rollup_interval,
            completed_retention: self.completed_retention,
            failed_retention: self.failed_retention,
            descriptor_retention: self.descriptor_retention,
            cleanup_batch_size: self.cleanup_batch_size,
            cleanup_interval: self.cleanup_interval,
            queue_retention_overrides: self.queue_retention_overrides,
            queue_stats_interval: self.queue_stats_interval,
            dlq_policy,
            dlq_retention: self.dlq_retention,
            dlq_cleanup_batch_size: self.dlq_cleanup_batch_size,
            effective_storage: Arc::new(RwLock::new(self.storage.clone())),
            storage: self.storage,
            transition_role: self.transition_role,
            global_max_workers: self.global_max_workers,
            runtime_snapshot_interval: self.runtime_snapshot_interval,
            runtime_instance_id: Uuid::new_v4(),
            runtime_started_at: Utc::now(),
            runtime_hostname: std::env::var("HOSTNAME").ok(),
            runtime_pid: std::process::id() as i32,
            runtime_version: env!("CARGO_PKG_VERSION"),
        })
    }
}

/// A typed worker that deserializes args and calls a handler function.
struct TypedWorker<T, F, Fut>
where
    T: JobArgs + DeserializeOwned + Send + Sync + 'static,
    F: Fn(T, &crate::context::JobContext) -> Fut + Send + Sync + 'static,
    Fut: std::future::Future<Output = Result<JobResult, JobError>> + Send + 'static,
{
    kind: &'static str,
    handler: Arc<F>,
    _phantom: std::marker::PhantomData<fn() -> (T, Fut)>,
}

#[async_trait::async_trait]
impl<T, F, Fut> Worker for TypedWorker<T, F, Fut>
where
    T: JobArgs + DeserializeOwned + Send + Sync + 'static,
    F: Fn(T, &crate::context::JobContext) -> Fut + Send + Sync + 'static,
    Fut: std::future::Future<Output = Result<JobResult, JobError>> + Send + 'static,
{
    fn kind(&self) -> &'static str {
        self.kind
    }

    async fn perform(&self, ctx: &crate::context::JobContext) -> Result<JobResult, JobError> {
        let args: T = serde_json::from_value(ctx.job.args.clone())
            .map_err(|err| JobError::Terminal(format!("failed to deserialize args: {}", err)))?;

        (self.handler)(args, ctx).await
    }
}

/// The Awa worker client — manages dispatchers, heartbeat, and maintenance.
pub struct Client {
    pool: PgPool,
    queues: Vec<(String, QueueConfig)>,
    queue_descriptors: HashMap<String, QueueDescriptor>,
    job_kind_descriptors: HashMap<String, JobKindDescriptor>,
    workers: Arc<HashMap<String, BoxedWorker>>,
    lifecycle_handlers: Arc<HashMap<String, Vec<BoxedUntypedEventHandler>>>,
    enqueue_specs: Arc<
        HashMap<
            crate::enqueue_specs::Outcome,
            HashMap<String, Vec<crate::enqueue_specs::BoxedEnqueueSpec>>,
        >,
    >,
    state: Arc<HashMap<TypeId, Box<dyn Any + Send + Sync>>>,
    heartbeat_interval: Duration,
    promote_interval: Duration,
    heartbeat_rescue_interval: Option<Duration>,
    heartbeat_staleness: Option<Duration>,
    deadline_rescue_interval: Option<Duration>,
    callback_rescue_interval: Option<Duration>,
    periodic_jobs: Arc<Vec<PeriodicJob>>,
    /// Cancellation token for dispatchers only — stops claiming new jobs.
    dispatch_cancel: CancellationToken,
    /// Cancellation token for heartbeat + maintenance — kept alive during drain.
    service_cancel: CancellationToken,
    /// Handles for dispatcher tasks.
    dispatcher_handles: RwLock<Vec<tokio::task::JoinHandle<()>>>,
    /// Handles for service tasks (heartbeat + maintenance).
    service_handles: RwLock<Vec<tokio::task::JoinHandle<()>>>,
    /// JoinSet tracking in-flight job tasks for graceful drain.
    job_set: Arc<Mutex<JoinSet<()>>>,
    in_flight: InFlightMap,
    queue_in_flight: Arc<HashMap<String, Arc<AtomicU32>>>,
    dispatcher_alive: Arc<HashMap<String, Arc<AtomicBool>>>,
    heartbeat_alive: Arc<AtomicBool>,
    maintenance_alive: Arc<AtomicBool>,
    leader: Arc<AtomicBool>,
    /// Shared overflow pool for weighted mode (None in hard-reserved mode).
    overflow_pool: Option<Arc<OverflowPool>>,
    metrics: crate::metrics::AwaMetrics,
    leader_election_interval: Option<Duration>,
    leader_check_interval: Option<Duration>,
    priority_aging_interval: Option<Duration>,
    terminal_count_rollup_interval: Option<Duration>,
    completed_retention: Option<Duration>,
    failed_retention: Option<Duration>,
    descriptor_retention: Option<Duration>,
    cleanup_batch_size: Option<i64>,
    cleanup_interval: Option<Duration>,
    queue_retention_overrides: HashMap<String, RetentionPolicy>,
    queue_stats_interval: Option<Duration>,
    dlq_policy: DlqPolicy,
    dlq_retention: Option<Duration>,
    dlq_cleanup_batch_size: Option<i64>,
    storage: RuntimeStorage,
    transition_role: TransitionWorkerRole,
    effective_storage: Arc<RwLock<RuntimeStorage>>,
    global_max_workers: Option<u32>,
    runtime_snapshot_interval: Duration,
    runtime_instance_id: Uuid,
    runtime_started_at: DateTime<Utc>,
    runtime_hostname: Option<String>,
    runtime_pid: i32,
    runtime_version: &'static str,
}

#[derive(Clone)]
struct RuntimeReporterState {
    pool: PgPool,
    queues: Vec<(String, QueueConfig)>,
    queue_descriptors: HashMap<String, QueueDescriptor>,
    job_kind_descriptors: HashMap<String, JobKindDescriptor>,
    worker_kinds: Vec<String>,
    queue_in_flight: Arc<HashMap<String, Arc<AtomicU32>>>,
    dispatcher_alive: Arc<HashMap<String, Arc<AtomicBool>>>,
    heartbeat_alive: Arc<AtomicBool>,
    maintenance_alive: Arc<AtomicBool>,
    leader: Arc<AtomicBool>,
    dispatch_cancel: CancellationToken,
    overflow_pool: Option<Arc<OverflowPool>>,
    global_max_workers: Option<u32>,
    dlq_policy: DlqPolicy,
    instance_id: Uuid,
    started_at: DateTime<Utc>,
    hostname: Option<String>,
    pid: i32,
    version: &'static str,
    snapshot_interval: Duration,
    effective_storage: Arc<RwLock<RuntimeStorage>>,
    queue_storage_capable: bool,
    transition_role: TransitionWorkerRole,
    metrics: crate::metrics::AwaMetrics,
}

/// Best-effort extraction of the most recent error message from a job's
/// `errors` history, for populating callback-driven `Exhausted`/`Retried`
/// events. Returns an empty string when no structured error is present.
fn latest_error_message(job: &awa_model::JobRow) -> String {
    job.errors
        .as_ref()
        .and_then(|errors| errors.last())
        .and_then(|entry| entry.get("error"))
        .and_then(|value| value.as_str())
        .map(str::to_string)
        .unwrap_or_default()
}

impl Client {
    /// Create a new builder.
    pub fn builder(pool: PgPool) -> ClientBuilder {
        ClientBuilder::new(pool)
    }

    fn expected_queue_storage_schema(
        status: &transition::StorageStatus,
    ) -> Result<Option<String>, awa_model::AwaError> {
        let prepared_schema = || {
            status
                .details
                .get("schema")
                .and_then(serde_json::Value::as_str)
                .unwrap_or("awa")
                .to_string()
        };

        match status.state.as_str() {
            "prepared" if status.prepared_engine.as_deref() == Some("queue_storage") => {
                Ok(Some(prepared_schema()))
            }
            "mixed_transition" | "active" if status.active_engine == "queue_storage" => {
                Ok(Some(prepared_schema()))
            }
            "canonical" if status.prepared_engine.as_deref() == Some("queue_storage") => {
                Ok(Some(prepared_schema()))
            }
            "mixed_transition" | "active" => Err(awa_model::AwaError::Validation(format!(
                "unsupported active storage engine '{}'",
                status.active_engine
            ))),
            _ => Ok(None),
        }
    }

    async fn resolve_effective_storage(&self) -> Result<RuntimeStorage, awa_model::AwaError> {
        let Some(runtime) = self.storage.queue_storage() else {
            return Ok(RuntimeStorage::Canonical);
        };

        let status = transition::status(&self.pool).await?;
        let expected_schema = Self::expected_queue_storage_schema(&status)?;
        let prepared_schema_ready = if let Some(schema) = expected_schema.as_deref() {
            if runtime.store.schema() != schema {
                return Err(awa_model::AwaError::Validation(format!(
                    "queue storage runtime configured for schema '{}' but transition state requires '{}'",
                    runtime.store.schema(),
                    schema
                )));
            }
            transition::queue_storage_schema_ready(&self.pool, schema).await?
        } else {
            false
        };

        match self.transition_role {
            TransitionWorkerRole::CanonicalDrain => Ok(RuntimeStorage::Canonical),
            TransitionWorkerRole::QueueStorageTarget => {
                let schema = expected_schema.ok_or_else(|| {
                    awa_model::AwaError::Validation(
                        "queue_storage_target requires a prepared queue-storage schema".into(),
                    )
                })?;
                if !prepared_schema_ready {
                    return Err(awa_model::AwaError::Validation(format!(
                        "queue storage schema '{schema}' is not prepared; run schema preparation before starting queue-storage-target runtimes"
                    )));
                }
                Ok(RuntimeStorage::QueueStorage(runtime.clone()))
            }
            TransitionWorkerRole::Auto => {
                if let Some(schema) = expected_schema.as_deref() {
                    if !prepared_schema_ready {
                        return Err(awa_model::AwaError::Validation(format!(
                            "queue storage schema '{schema}' is not prepared; run schema preparation before starting 0.6 runtimes"
                        )));
                    }
                }

                // Fresh-install auto-finalize: state=canonical with no
                // operator commands run yet means a brand-new cluster
                // shouldn't have to step through prepare → enter-mixed-
                // transition → finalize manually. Install the queue-
                // storage schema (idempotent — concurrent workers are
                // safe), then ask the SQL gate to advance state directly
                // to `active` if the fresh-install conditions hold (no
                // canonical jobs, no live workers, prepared_engine
                // still NULL). Returns FALSE on any non-fresh DB; the
                // caller then falls back to the canonical-storage path
                // and the staged transition is unaffected.
                if status.state == "canonical" && status.prepared_engine.is_none() {
                    let configured_schema = runtime.store.schema().to_string();
                    if !transition::queue_storage_schema_ready(&self.pool, &configured_schema)
                        .await?
                    {
                        runtime.store.prepare_schema(&self.pool).await?;
                    }
                    let promoted: bool =
                        sqlx::query_scalar("SELECT awa.storage_auto_finalize_if_fresh($1)")
                            .bind(&configured_schema)
                            .fetch_one(&self.pool)
                            .await?;
                    if promoted {
                        return Ok(RuntimeStorage::QueueStorage(runtime.clone()));
                    }
                    // Another worker promoted concurrently while we
                    // were installing the schema — re-fetch and use
                    // the now-current state.
                    let refetched = transition::status(&self.pool).await?;
                    if matches!(refetched.state.as_str(), "mixed_transition" | "active")
                        && refetched.active_engine == "queue_storage"
                    {
                        return Ok(RuntimeStorage::QueueStorage(runtime.clone()));
                    }
                    // Function returned FALSE because conditions weren't
                    // met (canonical jobs exist, or another runtime is
                    // live but in canonical-only mode). Fall through to
                    // the canonical path; operators will need the
                    // staged transition.
                }

                if matches!(status.state.as_str(), "mixed_transition" | "active")
                    && status.active_engine == "queue_storage"
                {
                    Ok(RuntimeStorage::QueueStorage(runtime.clone()))
                } else {
                    Ok(RuntimeStorage::Canonical)
                }
            }
        }
    }

    fn declared_queue_descriptors(&self) -> Vec<NamedQueueDescriptor> {
        self.queues
            .iter()
            .map(|(queue, _)| NamedQueueDescriptor {
                queue: queue.clone(),
                descriptor: self
                    .queue_descriptors
                    .get(queue)
                    .cloned()
                    .unwrap_or_default(),
            })
            .collect()
    }

    fn declared_job_kind_descriptors(&self) -> Vec<NamedJobKindDescriptor> {
        let mut kinds: Vec<String> = self.workers.keys().cloned().collect();
        for kind in self.job_kind_descriptors.keys() {
            if !kinds.iter().any(|existing| existing == kind) {
                kinds.push(kind.clone());
            }
        }
        kinds.sort();

        kinds
            .into_iter()
            .map(|kind| NamedJobKindDescriptor {
                descriptor: self
                    .job_kind_descriptors
                    .get(&kind)
                    .cloned()
                    .unwrap_or_default(),
                kind,
            })
            .collect()
    }

    fn runtime_reporter_state(&self) -> RuntimeReporterState {
        RuntimeReporterState {
            pool: self.pool.clone(),
            queues: self.queues.clone(),
            queue_descriptors: self.queue_descriptors.clone(),
            job_kind_descriptors: self.job_kind_descriptors.clone(),
            worker_kinds: self.workers.keys().cloned().collect(),
            queue_in_flight: self.queue_in_flight.clone(),
            dispatcher_alive: self.dispatcher_alive.clone(),
            heartbeat_alive: self.heartbeat_alive.clone(),
            maintenance_alive: self.maintenance_alive.clone(),
            leader: self.leader.clone(),
            dispatch_cancel: self.dispatch_cancel.clone(),
            overflow_pool: self.overflow_pool.clone(),
            global_max_workers: self.global_max_workers,
            dlq_policy: self.dlq_policy.clone(),
            instance_id: self.runtime_instance_id,
            started_at: self.runtime_started_at,
            hostname: self.runtime_hostname.clone(),
            pid: self.runtime_pid,
            version: self.runtime_version,
            snapshot_interval: self.runtime_snapshot_interval,
            effective_storage: self.effective_storage.clone(),
            queue_storage_capable: self.storage.queue_storage().is_some(),
            transition_role: self.transition_role,
            metrics: self.metrics.clone(),
        }
    }

    async fn publish_runtime_snapshot(&self) {
        let reporter = self.runtime_reporter_state();
        reporter.publish_snapshot().await;
    }

    async fn log_transition_startup_status(
        &self,
        effective_storage: &RuntimeStorage,
    ) -> Result<(), awa_model::AwaError> {
        if self.storage.queue_storage().is_none() {
            return Ok(());
        }

        let report = transition::status_report(&self.pool).await?;
        let effective_engine = match effective_storage {
            RuntimeStorage::Canonical => "canonical",
            RuntimeStorage::QueueStorage(_) => "queue_storage",
        };

        info!(
            transition_role = ?self.transition_role,
            state = %report.status.state,
            current_engine = %report.status.current_engine,
            active_engine = %report.status.active_engine,
            prepared_engine = ?report.status.prepared_engine,
            effective_engine,
            canonical_live_backlog = report.canonical_live_backlog,
            "Resolved storage transition state for worker startup"
        );

        if report.status.state == "prepared" && !report.can_enter_mixed_transition {
            warn!(
                blockers = %report.enter_mixed_transition_blockers.join("; "),
                "Storage transition is prepared but cannot yet enter mixed transition"
            );
        }

        if report.status.state == "mixed_transition" && !report.can_finalize {
            warn!(
                blockers = %report.finalize_blockers.join("; "),
                "Storage transition is in mixed_transition but cannot yet finalize"
            );
        }

        Ok(())
    }

    /// Start the worker runtime. Spawns dispatchers, heartbeat, and maintenance.
    pub async fn start(&self) -> Result<(), awa_model::AwaError> {
        info!(
            queues = self.queues.len(),
            workers = self.workers.len(),
            "Starting Awa worker runtime"
        );

        let effective_storage = self.resolve_effective_storage().await?;
        {
            let mut guard = self.effective_storage.write().await;
            *guard = effective_storage.clone();
        }

        self.log_transition_startup_status(&effective_storage)
            .await?;

        admin::sync_queue_descriptors(
            &self.pool,
            &self.declared_queue_descriptors(),
            self.runtime_snapshot_interval,
        )
        .await?;
        admin::sync_job_kind_descriptors(
            &self.pool,
            &self.declared_job_kind_descriptors(),
            self.runtime_snapshot_interval,
        )
        .await?;

        // Completion batcher stays alive during drain so tasks can release
        // only after their completion has been acknowledged.
        let runtime_worker_capacity = self.global_max_workers.unwrap_or_else(|| {
            self.queues
                .iter()
                .map(|(_, config)| config.max_workers)
                .sum()
        });
        let (completion_batcher, completion_handle) = CompletionBatcher::new(
            self.pool.clone(),
            self.service_cancel.clone(),
            self.metrics.clone(),
            effective_storage.clone(),
            runtime_worker_capacity,
        );

        // Create executor with metrics
        let executor = Arc::new(JobExecutor::new(
            self.pool.clone(),
            self.workers.clone(),
            self.lifecycle_handlers.clone(),
            self.enqueue_specs.clone(),
            self.in_flight.clone(),
            self.queue_in_flight.clone(),
            self.state.clone(),
            self.metrics.clone(),
            completion_handle,
            effective_storage.clone(),
            self.dlq_policy.clone(),
        ));

        // Admin cancellation listener: fires the in-flight cancel flag
        // for any locally-running attempt when an admin issues
        // `cancel(job_id)` on the DB. Listen before dispatchers start
        // claiming so an early admin cancel cannot race listener setup.
        let cancel_listener = crate::cancel_listener::CancelListener::new(
            self.pool.clone(),
            self.in_flight.clone(),
            self.service_cancel.clone(),
        );
        let cancel_listener_handle = cancel_listener.spawn().await;

        let mut service_handles = self.service_handles.write().await;

        service_handles.extend(completion_batcher.spawn());
        if let Some(handle) = cancel_listener_handle {
            service_handles.push(handle);
        }

        // Start heartbeat service (uses service_cancel — stays alive during drain)
        let heartbeat = HeartbeatService::new(
            self.pool.clone(),
            self.storage.clone(),
            self.in_flight.clone(),
            self.heartbeat_interval,
            self.heartbeat_alive.clone(),
            self.service_cancel.clone(),
            self.metrics.clone(),
        );
        service_handles.push(tokio::spawn(async move {
            heartbeat.run().await;
        }));

        // Start maintenance service (uses service_cancel — stays alive during drain)
        let mut maintenance = MaintenanceService::new(
            self.pool.clone(),
            self.metrics.clone(),
            self.leader.clone(),
            self.maintenance_alive.clone(),
            self.service_cancel.clone(),
            self.periodic_jobs.clone(),
            self.in_flight.clone(),
            effective_storage.clone(),
            self.enqueue_specs.clone(),
            self.lifecycle_handlers.clone(),
        )
        .standby_queue_storage(
            // #456: a canonical-resolved worker keeps its configured
            // queue-storage runtime available so a pre-flip maintenance
            // leader can promote the queue-storage deferred backlog once
            // routing flips.
            match (&effective_storage, &self.storage) {
                (RuntimeStorage::Canonical, RuntimeStorage::QueueStorage(runtime)) => {
                    Some(runtime.clone())
                }
                _ => None,
            },
        )
        .promote_interval(self.promote_interval);
        if let Some(interval) = self.heartbeat_rescue_interval {
            maintenance = maintenance.heartbeat_rescue_interval(interval);
        }
        if let Some(staleness) = self.heartbeat_staleness {
            maintenance = maintenance.heartbeat_staleness(staleness);
        }
        if let Some(interval) = self.deadline_rescue_interval {
            maintenance = maintenance.deadline_rescue_interval(interval);
        }
        if let Some(interval) = self.callback_rescue_interval {
            maintenance = maintenance.callback_rescue_interval(interval);
        }
        if let Some(interval) = self.leader_election_interval {
            maintenance = maintenance.leader_election_interval(interval);
        }
        if let Some(interval) = self.leader_check_interval {
            maintenance = maintenance.leader_check_interval(interval);
        }
        if let Some(interval) = self.priority_aging_interval {
            maintenance = maintenance.priority_aging_interval(interval);
        }
        if let Some(interval) = self.terminal_count_rollup_interval {
            maintenance = maintenance.terminal_count_rollup_interval(interval);
        }
        if let Some(retention) = self.completed_retention {
            maintenance = maintenance.completed_retention(retention);
        }
        if let Some(retention) = self.failed_retention {
            maintenance = maintenance.failed_retention(retention);
        }
        if let Some(retention) = self.descriptor_retention {
            maintenance = maintenance.descriptor_retention(retention);
        }
        if let Some(batch_size) = self.cleanup_batch_size {
            maintenance = maintenance.cleanup_batch_size(batch_size);
        }
        if let Some(interval) = self.cleanup_interval {
            maintenance = maintenance.cleanup_interval(interval);
        }
        if !self.queue_retention_overrides.is_empty() {
            maintenance =
                maintenance.queue_retention_overrides(self.queue_retention_overrides.clone());
        }
        if let Some(interval) = self.queue_stats_interval {
            maintenance = maintenance.queue_stats_interval(interval);
        }
        if let Some(retention) = self.dlq_retention {
            maintenance = maintenance.dlq_retention(retention);
        }
        if let Some(batch_size) = self.dlq_cleanup_batch_size {
            maintenance = maintenance.dlq_cleanup_batch_size(batch_size);
        }
        maintenance = maintenance.dlq_policy(self.dlq_policy.clone());
        service_handles.push(tokio::spawn(async move {
            maintenance.run().await;
        }));

        // Start dispatcher/claimer loops per queue (uses dispatch_cancel — stops claiming first).
        let mut dispatcher_handles = self.dispatcher_handles.write().await;
        for (queue_name, config) in &self.queues {
            let alive = self
                .dispatcher_alive
                .get(queue_name)
                .cloned()
                .unwrap_or_else(|| Arc::new(AtomicBool::new(false)));
            let claimers = usize::from(config.claimers.max(1));
            let capacity_wake = Arc::new(tokio::sync::Notify::new());
            let rate_limiter = shared_rate_limiter(config);

            let hard_reserved = self
                .overflow_pool
                .is_none()
                .then(|| Arc::new(tokio::sync::Semaphore::new(config.max_workers as usize)));
            let weighted_local = self
                .overflow_pool
                .as_ref()
                .map(|_| Arc::new(tokio::sync::Semaphore::new(config.min_workers as usize)));

            for claimer_idx in 0..claimers {
                let concurrency = if let Some(overflow_pool) = &self.overflow_pool {
                    ConcurrencyMode::Weighted {
                        local_semaphore: weighted_local
                            .as_ref()
                            .expect("weighted local semaphore should exist")
                            .clone(),
                        overflow_pool: overflow_pool.clone(),
                        queue_name: queue_name.clone(),
                    }
                } else {
                    ConcurrencyMode::HardReserved {
                        semaphore: hard_reserved
                            .as_ref()
                            .expect("hard-reserved semaphore should exist")
                            .clone(),
                    }
                };
                let claimer_owner_id = if claimer_idx == 0 {
                    self.runtime_instance_id
                } else {
                    // queue_claimer_leases owner ids are independent lease tokens.
                    // Extra dispatcher loops in the same runtime need distinct tokens
                    // so they can hold separate bounded-claimer slots.
                    Uuid::new_v4()
                };
                let dispatcher = Dispatcher::with_concurrency(
                    queue_name.clone(),
                    self.runtime_instance_id,
                    config.clone(),
                    self.pool.clone(),
                    executor.clone(),
                    self.metrics.clone(),
                    self.in_flight.clone(),
                    alive.clone(),
                    self.dispatch_cancel.clone(),
                    self.job_set.clone(),
                    concurrency,
                    rate_limiter.clone(),
                    capacity_wake.clone(),
                    claimer_owner_id,
                    effective_storage.clone(),
                );
                dispatcher_handles.push(tokio::spawn(async move {
                    dispatcher.run().await;
                }));
            }
        }

        self.publish_runtime_snapshot().await;

        let reporter = self.runtime_reporter_state();
        service_handles.push(tokio::spawn(async move {
            reporter.run().await;
        }));

        info!("Awa worker runtime started");
        Ok(())
    }

    /// Graceful shutdown with drain timeout.
    ///
    /// Phased lifecycle:
    /// 1. Stop dispatchers (no new jobs claimed)
    /// 2. Signal in-flight jobs to cancel
    /// 3. Wait for dispatchers to exit
    /// 4. Drain in-flight jobs (heartbeat + maintenance still alive!)
    /// 5. Stop heartbeat + maintenance
    pub async fn shutdown(&self, timeout: Duration) {
        info!("Initiating graceful shutdown");

        // Phase 1: Stop claiming new jobs
        self.dispatch_cancel.cancel();

        self.publish_runtime_snapshot().await;

        // Phase 2: Signal in-flight cancellation flags
        for flag in self.in_flight.flags() {
            flag.store(true, Ordering::SeqCst);
        }

        // Phase 3: Wait for dispatchers to exit their poll loops
        let dispatcher_handles: Vec<_> = {
            let mut guard = self.dispatcher_handles.write().await;
            std::mem::take(&mut *guard)
        };
        for handle in dispatcher_handles {
            let _ = handle.await;
        }

        // Phase 4: Drain in-flight jobs (heartbeat + maintenance still alive)
        let drain = async {
            let mut set = self.job_set.lock().await;
            while set.join_next().await.is_some() {}
        };
        if tokio::time::timeout(timeout, drain).await.is_err() {
            warn!(
                timeout_secs = timeout.as_secs(),
                "Shutdown drain timeout exceeded, some jobs may not have completed"
            );
        }

        // Phase 5: Stop background services (heartbeat + maintenance)
        self.service_cancel.cancel();
        let service_handles: Vec<_> = {
            let mut guard = self.service_handles.write().await;
            std::mem::take(&mut *guard)
        };
        for handle in service_handles {
            let _ = handle.await;
        }

        info!("Awa worker runtime stopped");
    }

    /// Get the pool reference.
    pub fn pool(&self) -> &PgPool {
        &self.pool
    }

    /// Resolve a pending external callback and dispatch the matching
    /// lifecycle event + ADR-029 follow-up specs.
    ///
    /// The callback transition (Complete / Fail) and any registered
    /// `on_completed_enqueue` / `on_exhausted_enqueue` follow-up `INSERT`s
    /// commit in a single transaction. A follow-up `INSERT` failure (or a
    /// panic in the user-supplied closure) rolls the callback transition
    /// back; the caller sees an `Err` and can surface a retryable failure
    /// to the external sender so the callback can be redelivered. `Ignored`
    /// produces no transition; `Resumed` does not currently fire a
    /// follow-up spec because resume re-enters execution and the executor
    /// emits the eventual outcome event itself.
    ///
    /// Prefer this over [`awa_model::admin::resolve_callback`] when you
    /// want `Completed`/`Exhausted` hooks or follow-up enqueues to fire
    /// for callback-driven outcomes. Hooks fire only in this process and
    /// only after the transaction commits.
    pub async fn resolve_callback(
        &self,
        callback_id: Uuid,
        payload: Option<serde_json::Value>,
        default_action: awa_model::DefaultAction,
        run_lease: Option<i64>,
    ) -> Result<awa_model::ResolveOutcome, awa_model::AwaError> {
        let mut tx = self.pool.begin().await?;
        let outcome =
            admin::resolve_callback_in_tx(&mut tx, callback_id, payload, default_action, run_lease)
                .await?;
        // ADR-029: dispatch follow-up specs in the same transaction. A
        // spec failure rolls the transition back together with the
        // follow-up, so callers can retry the external callback rather
        // than discovering a half-applied result.
        let event = match &outcome {
            awa_model::ResolveOutcome::Completed { job, .. } => {
                self.dispatch_callback_followups_in_tx(
                    &mut tx,
                    job,
                    crate::enqueue_specs::Outcome::Completed,
                    None,
                )
                .await?;
                Some(UntypedJobEvent::Completed {
                    job: job.clone(),
                    duration: Duration::ZERO,
                })
            }
            awa_model::ResolveOutcome::Failed { job } => {
                let outcome_ctx = crate::enqueue_specs::OutcomeContext::Exhausted {
                    error: latest_error_message(job),
                    attempt: job.attempt,
                };
                self.dispatch_callback_followups_in_tx(
                    &mut tx,
                    job,
                    crate::enqueue_specs::Outcome::Exhausted,
                    Some(outcome_ctx),
                )
                .await?;
                Some(UntypedJobEvent::Exhausted {
                    job: job.clone(),
                    error: latest_error_message(job),
                    attempt: job.attempt,
                })
            }
            awa_model::ResolveOutcome::Ignored { .. } => None,
        };
        tx.commit().await?;
        if let Some(event) = event {
            self.dispatch_callback_event(event).await;
        }
        Ok(outcome)
    }

    /// Complete a waiting job via its callback. The callback completion
    /// and any registered `on_completed_enqueue` follow-up `INSERT`s
    /// commit atomically. A spec INSERT failure rolls the completion
    /// back and returns `Err` so the external sender can retry. The
    /// in-process `Completed` hook fires after the transaction commits.
    pub async fn complete_external(
        &self,
        callback_id: Uuid,
        payload: Option<serde_json::Value>,
        run_lease: Option<i64>,
    ) -> Result<awa_model::JobRow, awa_model::AwaError> {
        let mut tx = self.pool.begin().await?;
        let job = admin::complete_external_in_tx(&mut tx, callback_id, payload, run_lease).await?;
        self.dispatch_callback_followups_in_tx(
            &mut tx,
            &job,
            crate::enqueue_specs::Outcome::Completed,
            None,
        )
        .await?;
        tx.commit().await?;
        self.dispatch_callback_event(UntypedJobEvent::Completed {
            job: job.clone(),
            duration: Duration::ZERO,
        })
        .await;
        Ok(job)
    }

    /// Fail a waiting job via its callback. The callback failure and any
    /// registered `on_exhausted_enqueue` follow-up `INSERT`s commit
    /// atomically. A spec INSERT failure rolls the failure back and
    /// returns `Err` so the external sender can retry. The in-process
    /// `Exhausted` hook fires after the transaction commits.
    pub async fn fail_external(
        &self,
        callback_id: Uuid,
        error: &str,
        run_lease: Option<i64>,
    ) -> Result<awa_model::JobRow, awa_model::AwaError> {
        let mut tx = self.pool.begin().await?;
        let job = admin::fail_external_in_tx(&mut tx, callback_id, error, run_lease).await?;
        let outcome_ctx = crate::enqueue_specs::OutcomeContext::Exhausted {
            error: error.to_string(),
            attempt: job.attempt,
        };
        self.dispatch_callback_followups_in_tx(
            &mut tx,
            &job,
            crate::enqueue_specs::Outcome::Exhausted,
            Some(outcome_ctx),
        )
        .await?;
        tx.commit().await?;
        self.dispatch_callback_event(UntypedJobEvent::Exhausted {
            job: job.clone(),
            error: error.to_string(),
            attempt: job.attempt,
        })
        .await;
        Ok(job)
    }

    /// Requeue a waiting job via its callback. The callback retry and
    /// any registered `on_retried_enqueue` follow-up `INSERT`s commit
    /// atomically. A spec INSERT failure rolls the retry back and
    /// returns `Err` so the external sender can retry. The in-process
    /// `Retried` hook fires after the transaction commits.
    ///
    /// `admin::retry_external_in_tx` resets `attempt` to 0 as part of
    /// requeuing the job from scratch, so the returned `JobRow` no
    /// longer reflects the attempt that was being retried. The pre-retry
    /// attempt number is captured under the same row lock before the
    /// transition runs so the `Retried` event and spec context report
    /// the failed-attempt number — matching the inline `Retried`
    /// semantics where `attempt` is "the attempt that was retried", not
    /// the post-transition value.
    pub async fn retry_external(
        &self,
        callback_id: Uuid,
        run_lease: Option<i64>,
    ) -> Result<awa_model::JobRow, awa_model::AwaError> {
        let mut tx = self.pool.begin().await?;
        let parked_attempt: Option<i16> = sqlx::query_scalar(
            "SELECT attempt FROM awa.jobs \
             WHERE callback_id = $1 AND state = 'waiting_external' \
               AND ($2::bigint IS NULL OR run_lease = $2)",
        )
        .bind(callback_id)
        .bind(run_lease)
        .fetch_optional(&mut *tx)
        .await?;

        let job = admin::retry_external_in_tx(&mut tx, callback_id, run_lease).await?;
        let attempt = parked_attempt.unwrap_or(job.attempt);
        let error_msg = latest_error_message(&job);
        let outcome_ctx = crate::enqueue_specs::OutcomeContext::Retried {
            error: error_msg.clone(),
            attempt,
            next_run_at: job.run_at,
        };
        self.dispatch_callback_followups_in_tx(
            &mut tx,
            &job,
            crate::enqueue_specs::Outcome::Retried,
            Some(outcome_ctx),
        )
        .await?;
        tx.commit().await?;
        self.dispatch_callback_event(UntypedJobEvent::Retried {
            job: job.clone(),
            error: error_msg,
            attempt,
            next_run_at: job.run_at,
        })
        .await;
        Ok(job)
    }

    async fn dispatch_callback_event(&self, event: UntypedJobEvent) {
        let kind = event.job().kind.clone();
        crate::executor::dispatch_lifecycle_event(&self.lifecycle_handlers, &kind, event).await;
    }

    /// ADR-029 callback-resolution follow-up dispatch inside the
    /// caller-owned transaction. Errors (including caught panics from
    /// user-supplied `make` closures) propagate out so the caller can
    /// roll the callback transition back together with the failed
    /// follow-up; the external sender then sees a retryable failure
    /// rather than a half-applied result.
    async fn dispatch_callback_followups_in_tx(
        &self,
        tx: &mut sqlx::Transaction<'_, sqlx::Postgres>,
        job: &awa_model::JobRow,
        outcome: crate::enqueue_specs::Outcome,
        outcome_context: Option<crate::enqueue_specs::OutcomeContext>,
    ) -> Result<(), awa_model::AwaError> {
        let Some(specs) = self
            .enqueue_specs
            .get(&outcome)
            .and_then(|by_kind| by_kind.get(&job.kind))
            .cloned()
        else {
            return Ok(());
        };
        if specs.is_empty() {
            return Ok(());
        }
        crate::enqueue_specs::dispatch_specs_in_tx(tx, job, &specs, outcome_context.as_ref()).await
    }

    /// Health check.
    pub async fn health_check(&self) -> HealthCheck {
        let postgres_connected = sqlx::query("SELECT 1").execute(&self.pool).await.is_ok();
        let poll_loop_alive = self
            .dispatcher_alive
            .values()
            .all(|alive| alive.load(Ordering::SeqCst));
        let heartbeat_alive = self.heartbeat_alive.load(Ordering::SeqCst);
        let maintenance_alive = self.maintenance_alive.load(Ordering::SeqCst);
        let shutting_down = self.dispatch_cancel.is_cancelled();
        let leader = self.leader.load(Ordering::SeqCst);
        let effective_storage = self.effective_storage.read().await.clone();
        let available_rows = if let Some(store) = effective_storage.queue_storage_store() {
            sqlx::query_as::<_, (String, i64)>(&format!(
                r#"
                SELECT
                    enqueues.queue,
                    COALESCE(
                        sum(GREATEST(
                            {}.sequence_next_value(enqueues.seq_name)
                                - {}.sequence_next_value(claims.seq_name),
                            0
                        )),
                        0
                    )::bigint AS available
                FROM {}.queue_enqueue_heads AS enqueues
                JOIN {}.queue_claim_heads AS claims
                  ON claims.queue = enqueues.queue
                 AND claims.priority = enqueues.priority
                 AND claims.enqueue_shard = enqueues.enqueue_shard
                GROUP BY enqueues.queue
                "#,
                store.schema(),
                store.schema(),
                store.schema(),
                store.schema()
            ))
            .fetch_all(&self.pool)
            .await
            .unwrap_or_default()
        } else {
            sqlx::query_as::<_, (String, i64)>(
                r#"
                SELECT queue, count(*)::bigint AS available
                FROM awa.jobs_hot
                WHERE state = 'available'
                GROUP BY queue
                "#,
            )
            .fetch_all(&self.pool)
            .await
            .unwrap_or_default()
        };
        let available_by_queue: HashMap<_, _> = available_rows.into_iter().collect();
        let queues = self
            .queues
            .iter()
            .map(|(queue, config)| {
                let in_flight = self
                    .queue_in_flight
                    .get(queue)
                    .map(|counter| counter.load(Ordering::SeqCst))
                    .unwrap_or(0);
                let available = available_by_queue.get(queue).copied().unwrap_or(0).max(0) as u64;
                let capacity = if let Some(overflow_pool) = &self.overflow_pool {
                    QueueCapacity::Weighted {
                        min_workers: config.min_workers,
                        weight: config.weight,
                        overflow_held: overflow_pool.held(queue),
                    }
                } else {
                    QueueCapacity::HardReserved {
                        max_workers: config.max_workers,
                    }
                };
                (
                    queue.clone(),
                    QueueHealth {
                        in_flight,
                        available,
                        capacity,
                    },
                )
            })
            .collect();

        HealthCheck {
            healthy: postgres_connected
                && poll_loop_alive
                && heartbeat_alive
                && maintenance_alive
                && !shutting_down,
            postgres_connected,
            poll_loop_alive,
            heartbeat_alive,
            maintenance_alive,
            shutting_down,
            leader,
            queues,
        }
    }
}

impl RuntimeReporterState {
    async fn storage_capability(&self) -> StorageCapability {
        if !self.queue_storage_capable {
            return StorageCapability::Canonical;
        }

        let effective_storage = self.effective_storage.read().await.clone();
        if matches!(effective_storage, RuntimeStorage::QueueStorage(_)) {
            return StorageCapability::QueueStorage;
        }

        match transition::status(&self.pool).await {
            Ok(status)
                if matches!(status.state.as_str(), "mixed_transition" | "active")
                    && status.active_engine == "queue_storage" =>
            {
                StorageCapability::CanonicalDrainOnly
            }
            Ok(_) => StorageCapability::QueueStorage,
            Err(err) => {
                warn!(
                    error = %err,
                    "Failed to resolve storage transition status for runtime snapshot"
                );
                StorageCapability::QueueStorage
            }
        }
    }

    fn queue_descriptor_hashes(&self) -> HashMap<String, String> {
        self.declared_queue_descriptors()
            .into_iter()
            .map(|named| (named.queue, named.descriptor.descriptor_hash()))
            .collect()
    }

    fn job_kind_descriptor_hashes(&self) -> HashMap<String, String> {
        self.declared_job_kind_descriptors()
            .into_iter()
            .map(|named| (named.kind, named.descriptor.descriptor_hash()))
            .collect()
    }

    fn declared_queue_descriptors(&self) -> Vec<NamedQueueDescriptor> {
        self.queues
            .iter()
            .map(|(queue, _)| NamedQueueDescriptor {
                queue: queue.clone(),
                descriptor: self
                    .queue_descriptors
                    .get(queue)
                    .cloned()
                    .unwrap_or_default(),
            })
            .collect()
    }

    fn declared_job_kind_descriptors(&self) -> Vec<NamedJobKindDescriptor> {
        let mut kinds = self.worker_kinds.clone();
        for kind in self.job_kind_descriptors.keys() {
            if !kinds.iter().any(|existing| existing == kind) {
                kinds.push(kind.clone());
            }
        }
        kinds.sort();
        kinds.dedup();

        kinds
            .into_iter()
            .map(|kind| NamedJobKindDescriptor {
                descriptor: self
                    .job_kind_descriptors
                    .get(&kind)
                    .cloned()
                    .unwrap_or_default(),
                kind,
            })
            .collect()
    }

    fn queue_snapshot(&self, queue: &str, config: &QueueConfig) -> QueueRuntimeSnapshot {
        let in_flight = self
            .queue_in_flight
            .get(queue)
            .map(|counter| counter.load(Ordering::SeqCst))
            .unwrap_or(0);

        let (mode, max_workers, min_workers, weight, overflow_held) =
            if let Some(overflow_pool) = &self.overflow_pool {
                (
                    QueueRuntimeMode::Weighted,
                    None,
                    Some(config.min_workers),
                    Some(config.weight),
                    Some(overflow_pool.held(queue)),
                )
            } else {
                (
                    QueueRuntimeMode::HardReserved,
                    Some(config.max_workers),
                    None,
                    None,
                    None,
                )
            };

        QueueRuntimeSnapshot {
            queue: queue.to_string(),
            in_flight,
            overflow_held,
            config: QueueRuntimeConfigSnapshot {
                mode,
                max_workers,
                min_workers,
                weight,
                global_max_workers: self.global_max_workers,
                poll_interval_ms: config.poll_interval.as_millis() as u64,
                deadline_duration_secs: config.deadline_duration.as_secs(),
                priority_aging_interval_secs: config.priority_aging_interval.as_secs(),
                claimers: Some(config.claimers),
                claim_batch_size: Some(config.claim_batch_size),
                dlq_enabled: Some(self.dlq_policy.enabled_for(queue)),
                rate_limit: config.rate_limit.as_ref().map(|rl| RateLimitSnapshot {
                    max_rate: rl.max_rate,
                    burst: rl.burst,
                }),
            },
        }
    }

    async fn snapshot_input(&self) -> RuntimeSnapshotInput {
        let postgres_connected = sqlx::query("SELECT 1").execute(&self.pool).await.is_ok();
        let poll_loop_alive = self
            .dispatcher_alive
            .values()
            .all(|alive| alive.load(Ordering::SeqCst));
        let heartbeat_alive = self.heartbeat_alive.load(Ordering::SeqCst);
        let maintenance_alive = self.maintenance_alive.load(Ordering::SeqCst);
        let shutting_down = self.dispatch_cancel.is_cancelled();
        let leader = self.leader.load(Ordering::SeqCst);
        let healthy = postgres_connected
            && poll_loop_alive
            && heartbeat_alive
            && maintenance_alive
            && !shutting_down;
        let storage_capability = self.storage_capability().await;
        let queues = self
            .queues
            .iter()
            .map(|(queue, config)| self.queue_snapshot(queue, config))
            .collect();

        RuntimeSnapshotInput {
            instance_id: self.instance_id,
            hostname: self.hostname.clone(),
            pid: self.pid,
            version: self.version.to_string(),
            storage_capability,
            transition_role: TransitionRole::from(self.transition_role),
            started_at: self.started_at,
            snapshot_interval_ms: self.snapshot_interval.as_millis() as i64,
            healthy,
            postgres_connected,
            poll_loop_alive,
            heartbeat_alive,
            maintenance_alive,
            shutting_down,
            leader,
            global_max_workers: self.global_max_workers,
            queues,
            queue_descriptor_hashes: self.queue_descriptor_hashes(),
            job_kind_descriptor_hashes: self.job_kind_descriptor_hashes(),
        }
    }

    async fn publish_snapshot(&self) {
        let queue_descriptors = self.declared_queue_descriptors();
        let kind_descriptors = self.declared_job_kind_descriptors();

        if let Err(err) =
            admin::sync_queue_descriptors(&self.pool, &queue_descriptors, self.snapshot_interval)
                .await
        {
            warn!(error = %err, "Failed to sync queue descriptors");
        }
        if let Err(err) =
            admin::sync_job_kind_descriptors(&self.pool, &kind_descriptors, self.snapshot_interval)
                .await
        {
            warn!(error = %err, "Failed to sync job kind descriptors");
        }

        // Emit OTel info gauges for every declared descriptor. One series per
        // descriptor, value=1, with all descriptor fields as attributes. Panels
        // lift descriptor fields into existing metrics via a Prometheus label
        // join: `awa_job_completed_total * on(awa_job_queue) group_left(awa_queue_display_name) awa_queue_info`.
        for named in &queue_descriptors {
            self.metrics.record_queue_info(
                &named.queue,
                named.descriptor.display_name.as_deref(),
                named.descriptor.description.as_deref(),
                named.descriptor.owner.as_deref(),
                named.descriptor.docs_url.as_deref(),
                &named.descriptor.tags,
            );
        }
        for named in &kind_descriptors {
            self.metrics.record_job_kind_info(
                &named.kind,
                named.descriptor.display_name.as_deref(),
                named.descriptor.description.as_deref(),
                named.descriptor.owner.as_deref(),
                named.descriptor.docs_url.as_deref(),
                &named.descriptor.tags,
            );
        }

        let snapshot = self.snapshot_input().await;
        if let Err(err) = admin::upsert_runtime_snapshot(&self.pool, &snapshot).await {
            warn!(error = %err, "Failed to publish runtime snapshot");
        }

        if self.queue_storage_capable {
            match transition::status_report(&self.pool).await {
                Ok(report) => {
                    self.metrics.record_storage_state(&report.status);
                    self.metrics.record_storage_transition_ready(
                        "enter_mixed_transition",
                        report.can_enter_mixed_transition,
                    );
                    self.metrics
                        .record_storage_transition_ready("finalize", report.can_finalize);
                    self.metrics
                        .record_storage_canonical_live_backlog(report.canonical_live_backlog);

                    for capability in ["canonical", "canonical_drain_only", "queue_storage"] {
                        let count = report
                            .live_runtime_capability_counts
                            .get(capability)
                            .copied()
                            .unwrap_or(0) as i64;
                        self.metrics
                            .record_storage_live_runtime_capability(capability, count);
                    }

                    for (capability, count) in report.live_runtime_capability_counts {
                        if capability != "canonical"
                            && capability != "canonical_drain_only"
                            && capability != "queue_storage"
                        {
                            self.metrics
                                .record_storage_live_runtime_capability(&capability, count as i64);
                        }
                    }
                }
                Err(err) => {
                    warn!(error = %err, "Failed to publish storage transition metrics");
                }
            }
        }
    }

    async fn run(self) {
        let mut interval = tokio::time::interval(self.snapshot_interval);
        interval.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Delay);
        interval.tick().await;
        loop {
            tokio::select! {
                _ = self.dispatch_cancel.cancelled() => {
                    self.publish_snapshot().await;
                    break;
                }
                _ = interval.tick() => {
                    self.publish_snapshot().await;
                }
            }
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use awa_model::{migrations, storage, JobArgs, QueueStorage, QueueStorageConfig};
    use sqlx::postgres::PgPoolOptions;
    use sqlx::PgPool;
    use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
    use std::sync::{Arc, Mutex, OnceLock};
    use std::time::{Duration, Instant};
    use tokio::sync::{oneshot, Notify};

    static TEST_MUTEX: OnceLock<tokio::sync::Mutex<()>> = OnceLock::new();

    fn test_mutex() -> &'static tokio::sync::Mutex<()> {
        TEST_MUTEX.get_or_init(|| tokio::sync::Mutex::new(()))
    }

    fn lazy_pool() -> PgPool {
        PgPoolOptions::new()
            .connect_lazy("postgres://postgres:test@localhost/awa_test")
            .expect("lazy pool should build")
    }

    #[test]
    fn queue_config_defaults_use_throughput_oriented_claim_batch() {
        let config = QueueConfig::default();
        assert_eq!(config.claimers, 1);
        assert_eq!(config.claim_batch_size, 512);
    }

    #[tokio::test]
    async fn register_accepts_send_non_sync_handler_future() {
        #[derive(serde::Serialize, serde::Deserialize, awa_macros::JobArgs)]
        struct NonSyncFutureJob;

        let _client = Client::builder(lazy_pool())
            .queue("non_sync_future", QueueConfig::default())
            .register::<NonSyncFutureJob, _, _>(|_, _| async move {
                let cell = std::cell::Cell::new(0_u8);
                tokio::task::yield_now().await;
                cell.set(1);
                Ok(JobResult::Completed)
            })
            .build()
            .expect("send but non-sync handler future should compile");
    }

    fn base_database_url() -> String {
        std::env::var("DATABASE_URL")
            .unwrap_or_else(|_| "postgres://postgres:test@localhost:15432/awa_test".to_string())
    }

    fn replace_database_name(url: &str, database_name: &str) -> String {
        let (without_query, query_suffix) = match url.split_once('?') {
            Some((prefix, query)) => (prefix, Some(query)),
            None => (url, None),
        };
        let (base, _) = without_query
            .rsplit_once('/')
            .expect("database URL should include a database name");
        let mut out = format!("{base}/{database_name}");
        if let Some(query) = query_suffix {
            out.push('?');
            out.push_str(query);
        }
        out
    }

    fn database_url() -> String {
        std::env::var("DATABASE_URL_WORKER_CLIENT").unwrap_or_else(|_| {
            replace_database_name(&base_database_url(), "awa_test_worker_client")
        })
    }

    async fn ensure_database_exists(url: &str) {
        let database_name = url
            .split_once('?')
            .map(|(prefix, _)| prefix)
            .unwrap_or(url)
            .rsplit_once('/')
            .map(|(_, database_name)| database_name.to_string())
            .expect("database URL should include a database name");
        let admin_url = replace_database_name(url, "postgres");
        let admin_pool = PgPoolOptions::new()
            .max_connections(1)
            .connect(&admin_url)
            .await
            .expect("Failed to connect to admin database for client tests");
        let create_sql = format!("CREATE DATABASE {database_name}");
        match sqlx::query(&create_sql).execute(&admin_pool).await {
            Ok(_) => {}
            Err(sqlx::Error::Database(db_err)) if db_err.code().as_deref() == Some("42P04") => {}
            Err(sqlx::Error::Database(db_err)) if db_err.code().as_deref() == Some("23505") => {}
            Err(err) => panic!("Failed to create client test database {database_name}: {err}"),
        }
    }

    async fn setup_pool(max_connections: u32) -> PgPool {
        let url = database_url();
        ensure_database_exists(&url).await;
        PgPoolOptions::new()
            .max_connections(max_connections)
            .acquire_timeout(Duration::from_secs(5))
            .connect(&url)
            .await
            .expect("Failed to connect to client test database")
    }

    async fn reset_schema(pool: &PgPool) {
        sqlx::raw_sql("DROP SCHEMA IF EXISTS awa CASCADE")
            .execute(pool)
            .await
            .expect("Failed to drop awa schema");
    }

    async fn apply_migrations_through(pool: &PgPool, version: i32) {
        for (_version, _desc, sql) in migrations::migration_sql_range(0, version) {
            sqlx::raw_sql(&sql).execute(pool).await.unwrap();
        }
    }

    async fn drop_queue_storage_schema(pool: &PgPool, schema: &str) {
        let sql = format!("DROP SCHEMA IF EXISTS {schema} CASCADE");
        sqlx::query(&sql)
            .execute(pool)
            .await
            .expect("Failed to drop queue storage schema");
    }

    async fn insert_available_job(pool: &PgPool, kind: &str, queue: &str) -> i64 {
        sqlx::query_scalar(
            r#"
            INSERT INTO awa.jobs (
                kind,
                queue,
                args,
                state,
                priority,
                max_attempts,
                run_at,
                metadata,
                tags
            )
            VALUES (
                $1,
                $2,
                '{}'::jsonb,
                'available'::awa.job_state,
                2,
                25,
                clock_timestamp(),
                '{}'::jsonb,
                '{}'::text[]
            )
            RETURNING id
            "#,
        )
        .bind(kind)
        .bind(queue)
        .fetch_one(pool)
        .await
        .expect("Failed to insert job")
    }

    async fn insert_canonical_available_job(pool: &PgPool, kind: &str, queue: &str) -> i64 {
        sqlx::query_scalar(
            r#"
            INSERT INTO awa.jobs_hot (
                kind,
                queue,
                args,
                state,
                priority,
                max_attempts,
                run_at,
                metadata,
                tags
            )
            VALUES (
                $1,
                $2,
                '{}'::jsonb,
                'available'::awa.job_state,
                2,
                25,
                clock_timestamp(),
                '{}'::jsonb,
                '{}'::text[]
            )
            RETURNING id
            "#,
        )
        .bind(kind)
        .bind(queue)
        .fetch_one(pool)
        .await
        .expect("Failed to insert canonical job")
    }

    async fn active_queue_storage_schema(pool: &PgPool) -> Option<String> {
        sqlx::query_scalar("SELECT awa.active_queue_storage_schema()")
            .fetch_one(pool)
            .await
            .expect("Failed to fetch active queue storage schema")
    }

    /// Insert a synthetic `transition_role=queue_storage_target` runtime
    /// row so the mixed-transition gate is satisfied without needing a
    /// second real client. Used by tests that only exercise the
    /// canonical-drain side of the transition.
    async fn insert_fake_queue_storage_target(pool: &PgPool) {
        sqlx::query(
            r#"
            INSERT INTO awa.runtime_instances (
                instance_id, hostname, pid, version,
                started_at, last_seen_at, snapshot_interval_ms,
                healthy, postgres_connected, poll_loop_alive,
                heartbeat_alive, maintenance_alive, shutting_down,
                leader, global_max_workers, queues,
                storage_capability, transition_role
            )
            VALUES (
                $1, 'fake-target', 7777, '0.6.0-test',
                now() - interval '1 minute', now(), 1000,
                TRUE, TRUE, TRUE,
                TRUE, TRUE, FALSE,
                FALSE, NULL, '[]'::jsonb,
                'queue_storage', 'queue_storage_target'
            )
            "#,
        )
        .bind(Uuid::new_v4())
        .execute(pool)
        .await
        .expect("Failed to insert fake queue_storage_target runtime row");
    }

    async fn wait_for_runtime_capability(
        pool: &PgPool,
        instance_id: Uuid,
        capability: StorageCapability,
        timeout: Duration,
    ) {
        let start = Instant::now();
        loop {
            let current: Option<String> = sqlx::query_scalar(
                "SELECT storage_capability FROM awa.runtime_instances WHERE instance_id = $1",
            )
            .bind(instance_id)
            .fetch_optional(pool)
            .await
            .expect("Failed to fetch runtime storage capability");
            if current.as_deref() == Some(capability.as_str()) {
                return;
            }
            assert!(
                start.elapsed() <= timeout,
                "Timed out waiting for runtime {instance_id} to report capability {}; last={current:?}",
                capability.as_str()
            );
            tokio::time::sleep(Duration::from_millis(25)).await;
        }
    }

    async fn expire_runtime_instance(pool: &PgPool, instance_id: Uuid) {
        sqlx::query(
            "UPDATE awa.runtime_instances SET last_seen_at = now() - interval '1 hour' WHERE instance_id = $1",
        )
        .bind(instance_id)
        .execute(pool)
        .await
        .expect("Failed to expire runtime instance");
    }

    /// Drop and re-create the queue-storage schema using the supplied
    /// config, then advance the storage transition state machine to the
    /// `prepared` engine. The `lease_claim_receipts` flag must already
    /// match what the runtime under test will use, because the
    /// receipts-vs-legacy claim CTE is baked into `claim_ready_runtime`
    /// at `prepare_schema` time — a later store built with a different
    /// flag value would still hit the SQL function compiled here.
    async fn prepare_queue_storage_transition_with_config(
        pool: &PgPool,
        config: QueueStorageConfig,
    ) -> QueueStorage {
        let schema = config.schema.clone();
        let store = QueueStorage::new(config).expect("Failed to build queue storage store");
        drop_queue_storage_schema(pool, &schema).await;
        store
            .prepare_schema(pool)
            .await
            .expect("Failed to prepare queue storage schema");
        storage::prepare(
            pool,
            "queue_storage",
            serde_json::json!({ "schema": schema }),
        )
        .await
        .expect("Failed to prepare queue storage transition");
        store
    }

    async fn wait_for_state(pool: &PgPool, job_id: i64, state: &str, timeout: Duration) {
        let start = Instant::now();
        loop {
            let current: Option<String> = sqlx::query_scalar(
                "SELECT state::text FROM awa.jobs_hot WHERE id = $1 UNION ALL SELECT state::text FROM awa.scheduled_jobs WHERE id = $1 LIMIT 1",
            )
            .bind(job_id)
            .fetch_optional(pool)
            .await
            .expect("Failed to fetch canonical job state");
            if current.as_deref() == Some(state) {
                return;
            }
            assert!(
                start.elapsed() <= timeout,
                "Timed out waiting for job {job_id} to reach state {state}; last_state={current:?}"
            );
            tokio::time::sleep(Duration::from_millis(25)).await;
        }
    }

    async fn wait_for_queue_storage_done(
        pool: &PgPool,
        schema: &str,
        job_id: i64,
        timeout: Duration,
    ) {
        let sql = format!(
            "SELECT EXISTS(SELECT 1 FROM {schema}.terminal_jobs WHERE job_id = $1 AND state = 'completed')"
        );
        let start = Instant::now();
        loop {
            let done: bool = sqlx::query_scalar(&sql)
                .bind(job_id)
                .fetch_one(pool)
                .await
                .expect("Failed to query queue storage terminal rows");
            if done {
                return;
            }
            assert!(
                start.elapsed() <= timeout,
                "Timed out waiting for queue storage job {job_id} to complete"
            );
            tokio::time::sleep(Duration::from_millis(25)).await;
        }
    }

    fn force_canonical(mut builder: ClientBuilder) -> ClientBuilder {
        builder.storage = RuntimeStorage::Canonical;
        builder.storage_error = None;
        builder
    }

    #[tokio::test]
    async fn queue_storage_target_requires_prepared_schema() {
        let _guard = test_mutex().lock().await;
        let pool = setup_pool(4).await;
        let queue_storage_schema = "awa_cutover_target_requires_prepare";
        reset_schema(&pool).await;
        migrations::run(&pool)
            .await
            .expect("fresh 0.6 schema install should succeed");
        drop_queue_storage_schema(&pool, queue_storage_schema).await;

        storage::prepare(
            &pool,
            "queue_storage",
            serde_json::json!({ "schema": queue_storage_schema }),
        )
        .await
        .expect("Failed to prepare queue storage transition without schema");

        let client = Client::builder(pool.clone())
            .queue(
                "cutover",
                QueueConfig {
                    max_workers: 1,
                    poll_interval: Duration::from_millis(25),
                    ..QueueConfig::default()
                },
            )
            .queue_storage(
                QueueStorageConfig {
                    schema: queue_storage_schema.to_string(),
                    queue_slot_count: 4,
                    lease_slot_count: 2,
                    ..Default::default()
                },
                Duration::from_millis(1_000),
                Duration::from_millis(50),
            )
            .transition_role(TransitionWorkerRole::QueueStorageTarget)
            .register::<CutoverShortJob, _, _>(move |_args, _ctx| async move {
                Ok(JobResult::Completed)
            })
            .build()
            .expect("Failed to build queue-storage target client");

        let err = client
            .start()
            .await
            .expect_err("queue-storage target should refuse to start without prepared schema");
        match err {
            awa_model::AwaError::Validation(msg) => {
                assert!(
                    msg.contains("not prepared"),
                    "unexpected validation message: {msg}"
                );
            }
            other => panic!("expected Validation error, got {other:?}"),
        }
    }

    #[tokio::test]
    async fn queue_descriptor_requires_declared_queue() {
        let result = Client::builder(lazy_pool())
            .queue("default", QueueConfig::default())
            .queue_descriptor("billing", QueueDescriptor::new().display_name("Billing"))
            .build();

        assert!(matches!(
            result,
            Err(BuildError::QueueDescriptorWithoutQueue { queue }) if queue == "billing"
        ));
    }

    #[tokio::test]
    async fn queue_descriptor_allows_declared_queue() {
        let result = Client::builder(lazy_pool())
            .queue("billing", QueueConfig::default())
            .queue_descriptor("billing", QueueDescriptor::new().display_name("Billing"))
            .build();

        assert!(result.is_ok(), "descriptor for declared queue should build");
    }

    #[tokio::test]
    async fn partitioned_queue_declares_each_physical_queue() {
        let partitioned_queue =
            PartitionedQueue::new("email", 3).expect("partitioned queue should build");

        let client = Client::builder(lazy_pool())
            .partitioned_queue(
                &partitioned_queue,
                QueueConfig {
                    max_workers: 7,
                    ..QueueConfig::default()
                },
            )
            .build()
            .expect("partitioned queue queues should build");

        let queues: Vec<_> = client
            .queues
            .iter()
            .map(|(queue, config)| (queue.as_str(), config.max_workers))
            .collect();
        assert_eq!(
            queues,
            vec![("email", 7), ("email__p1", 7), ("email__p2", 7)]
        );
    }

    #[tokio::test]
    async fn duplicate_queue_declarations_are_rejected() {
        let partitioned_queue =
            PartitionedQueue::new("email", 2).expect("partitioned queue should build");

        let result = Client::builder(lazy_pool())
            .queue("email", QueueConfig::default())
            .partitioned_queue(&partitioned_queue, QueueConfig::default())
            .build();

        assert!(matches!(
            result,
            Err(BuildError::DuplicateQueue { queue }) if queue == "email"
        ));
    }

    #[tokio::test]
    async fn job_kind_descriptor_allows_registered_kind() {
        #[derive(serde::Serialize, serde::Deserialize, awa_macros::JobArgs)]
        struct TestJob;

        let result = Client::builder(lazy_pool())
            .queue("default", QueueConfig::default())
            .register::<TestJob, _, _>(|_args, _ctx| async { Ok(JobResult::Completed) })
            .job_kind_descriptor::<TestJob>(JobKindDescriptor::new().display_name("Test job"))
            .build();

        assert!(
            result.is_ok(),
            "descriptor for registered kind should build"
        );
    }

    #[tokio::test]
    async fn dlq_cleanup_batch_size_must_be_positive() {
        let result = Client::builder(lazy_pool())
            .queue("default", QueueConfig::default())
            .dlq_cleanup_batch_size(0)
            .build();

        assert!(matches!(result, Err(BuildError::InvalidDlqBatchSize)));
    }

    #[tokio::test]
    async fn terminal_count_rollup_interval_must_be_positive() {
        let result = Client::builder(lazy_pool())
            .queue("default", QueueConfig::default())
            .terminal_count_rollup_interval(Duration::ZERO)
            .build();

        assert!(matches!(
            result,
            Err(BuildError::InvalidTerminalCountRollupInterval)
        ));
    }

    #[tokio::test]
    async fn health_check_reads_available_from_active_queue_storage() {
        let _guard = test_mutex().lock().await;
        let pool = setup_pool(4).await;
        reset_schema(&pool).await;
        migrations::run(&pool)
            .await
            .expect("migrations should succeed");
        // The explicit install remains an idempotence check now that
        // migrations materialize the default queue-storage substrate.

        let queue = "health_queue_storage";
        let client = Client::builder(pool.clone())
            .queue(queue, QueueConfig::default())
            .build()
            .expect("queue-storage health client should build");

        let store = client
            .storage
            .queue_storage_store()
            .expect("client should default to queue storage");
        store
            .install(&pool)
            .await
            .expect("queue storage install should succeed");

        insert_available_job(&pool, "cutover_short_job", queue).await;

        let health = client.health_check().await;
        let queue_health = health
            .queues
            .get(queue)
            .expect("queue should appear in health");
        assert_eq!(queue_health.available, 1);
    }

    #[derive(Clone, serde::Serialize, serde::Deserialize, awa_macros::JobArgs)]
    struct CutoverLongJob {}

    #[derive(Clone, serde::Serialize, serde::Deserialize, awa_macros::JobArgs)]
    struct CutoverShortJob {}

    #[tokio::test]
    async fn fresh_auto_finalize_uses_migrated_default_substrate_without_prepare_schema() {
        let _guard = test_mutex().lock().await;
        let pool = setup_pool(4).await;
        reset_schema(&pool).await;
        migrations::run(&pool)
            .await
            .expect("fresh 0.6 schema install should succeed");

        sqlx::raw_sql(
            r#"
            CREATE OR REPLACE FUNCTION awa.install_queue_storage_substrate(
                p_schema TEXT,
                p_queue_slot_count INT DEFAULT 16,
                p_lease_slot_count INT DEFAULT 8,
                p_claim_slot_count INT DEFAULT 8,
                p_lease_claim_receipts BOOLEAN DEFAULT TRUE
            )
            RETURNS VOID
            LANGUAGE plpgsql
            AS $$
            BEGIN
                RAISE EXCEPTION 'prepare_schema should not run when default queue-storage substrate is already ready'
                    USING ERRCODE = '55000';
            END
            $$;
            "#,
        )
        .execute(&pool)
        .await
        .expect("failed to poison queue-storage helper");

        let client = Client::builder(pool.clone())
            .queue(
                "fresh_auto_finalize",
                QueueConfig {
                    max_workers: 1,
                    poll_interval: Duration::from_millis(25),
                    deadline_duration: Duration::ZERO,
                    ..QueueConfig::default()
                },
            )
            .register::<CutoverShortJob, _, _>(move |_args, _ctx| async move {
                Ok(JobResult::Completed)
            })
            .promote_interval(Duration::from_millis(25))
            .leader_election_interval(Duration::from_millis(100))
            .leader_check_interval(Duration::from_millis(50))
            .runtime_snapshot_interval(Duration::from_millis(100))
            .build()
            .expect("Failed to build fresh auto-finalize client");

        client
            .start()
            .await
            .expect("fresh runtime should auto-finalize without prepare_schema");
        assert_eq!(
            active_queue_storage_schema(&pool).await.as_deref(),
            Some("awa")
        );
        client.shutdown(Duration::from_secs(5)).await;
    }

    #[tokio::test]
    async fn canonical_runtime_drains_in_flight_jobs_across_schema_upgrade_before_queue_storage_cutover(
    ) {
        let _guard = test_mutex().lock().await;
        let pool = setup_pool(8).await;
        let queue_storage_schema = "awa_cutover_runtime";
        reset_schema(&pool).await;
        drop_queue_storage_schema(&pool, queue_storage_schema).await;
        apply_migrations_through(&pool, 9).await;

        let long_started_flag = Arc::new(AtomicBool::new(false));
        let (long_started_tx_inner, long_started_rx) = oneshot::channel::<()>();
        let long_started_tx = Arc::new(Mutex::new(Some(long_started_tx_inner)));
        let long_release = Arc::new(Notify::new());
        let canonical_short_seen = Arc::new(AtomicUsize::new(0));
        let queue_storage_short_seen = Arc::new(AtomicUsize::new(0));

        let canonical_client = {
            let started = long_started_flag.clone();
            let started_tx = long_started_tx.clone();
            let release = long_release.clone();
            let canonical_short_seen = canonical_short_seen.clone();
            let builder = Client::builder(pool.clone())
                .queue(
                    "cutover",
                    QueueConfig {
                        max_workers: 2,
                        poll_interval: Duration::from_millis(25),
                        ..QueueConfig::default()
                    },
                )
                .register::<CutoverLongJob, _, _>(move |_args, _ctx| {
                    let started = started.clone();
                    let started_tx = started_tx.clone();
                    let release = release.clone();
                    async move {
                        started.store(true, Ordering::SeqCst);
                        if let Some(tx) =
                            started_tx.lock().expect("long-start mutex poisoned").take()
                        {
                            let _ = tx.send(());
                        }
                        release.notified().await;
                        Ok(JobResult::Completed)
                    }
                })
                .register::<CutoverShortJob, _, _>(move |_args, _ctx| {
                    let canonical_short_seen = canonical_short_seen.clone();
                    async move {
                        canonical_short_seen.fetch_add(1, Ordering::SeqCst);
                        Ok(JobResult::Completed)
                    }
                })
                .promote_interval(Duration::from_millis(25))
                .leader_election_interval(Duration::from_millis(100))
                .leader_check_interval(Duration::from_millis(50))
                .heartbeat_rescue_interval(Duration::from_millis(100))
                .deadline_rescue_interval(Duration::from_millis(100))
                .callback_rescue_interval(Duration::from_millis(100));
            force_canonical(builder)
                .build()
                .expect("Failed to build canonical client")
        };

        canonical_client
            .start()
            .await
            .expect("Failed to start canonical client");

        let long_id =
            insert_available_job(&pool, <CutoverLongJob as JobArgs>::kind(), "cutover").await;
        tokio::time::timeout(Duration::from_secs(5), long_started_rx)
            .await
            .expect("Timed out waiting for long canonical job to start")
            .expect("Long job start signal dropped");
        assert!(
            long_started_flag.load(Ordering::SeqCst),
            "long-running canonical job should be in flight before migration"
        );

        migrations::run(&pool)
            .await
            .expect("Schema upgrade from 0.5.x to 0.6 should succeed during canonical runtime");
        assert_eq!(
            active_queue_storage_schema(&pool).await,
            None,
            "schema upgrade alone must not activate queue storage"
        );

        let canonical_short_id =
            insert_available_job(&pool, <CutoverShortJob as JobArgs>::kind(), "cutover").await;
        let canonical_short_start = Instant::now();
        while canonical_short_seen.load(Ordering::SeqCst) == 0 {
            assert!(
                canonical_short_start.elapsed() <= Duration::from_secs(5),
                "canonical worker stopped processing new jobs after schema upgrade"
            );
            tokio::time::sleep(Duration::from_millis(25)).await;
        }
        wait_for_state(
            &pool,
            canonical_short_id,
            "completed",
            Duration::from_secs(5),
        )
        .await;

        long_release.notify_waiters();
        wait_for_state(&pool, long_id, "completed", Duration::from_secs(5)).await;
        canonical_client.shutdown(Duration::from_secs(5)).await;
        expire_runtime_instance(&pool, canonical_client.runtime_instance_id).await;

        // Pin the legacy lease-materialization path. The receipt-plane
        // fast path (now the default since ADR-023 Phase 6) requires
        // deadline_duration=0 on QueueConfig and would error every claim
        // here; this test exercises the canonical-drain → cutover flow
        // with the standard 60s deadline, so lease_claim_receipts stays
        // off. Tests that specifically cover the receipt path opt back
        // in and zero the deadline. The flag must be set when
        // `prepare_schema` runs because the receipts-vs-legacy claim CTE
        // is baked into the SQL function definition at that point.
        let store_config = QueueStorageConfig {
            schema: queue_storage_schema.to_string(),
            queue_slot_count: 4,
            lease_slot_count: 2,
            lease_claim_receipts: false,
            ..Default::default()
        };
        let _store =
            prepare_queue_storage_transition_with_config(&pool, store_config.clone()).await;
        assert_eq!(
            active_queue_storage_schema(&pool).await,
            None,
            "prepare alone must not activate queue storage routing"
        );
        let drain_only_client = {
            let queue_storage_short_seen = queue_storage_short_seen.clone();
            Client::builder(pool.clone())
                .queue(
                    "cutover",
                    QueueConfig {
                        max_workers: 2,
                        poll_interval: Duration::from_millis(25),
                        ..QueueConfig::default()
                    },
                )
                .queue_storage(
                    store_config.clone(),
                    Duration::from_millis(1_000),
                    Duration::from_millis(50),
                )
                .register::<CutoverShortJob, _, _>(move |_args, _ctx| {
                    let queue_storage_short_seen = queue_storage_short_seen.clone();
                    async move {
                        queue_storage_short_seen.fetch_add(1, Ordering::SeqCst);
                        Ok(JobResult::Completed)
                    }
                })
                .promote_interval(Duration::from_millis(25))
                .leader_election_interval(Duration::from_millis(100))
                .leader_check_interval(Duration::from_millis(50))
                .heartbeat_rescue_interval(Duration::from_millis(100))
                .deadline_rescue_interval(Duration::from_millis(100))
                .callback_rescue_interval(Duration::from_millis(100))
                .runtime_snapshot_interval(Duration::from_millis(100))
                .build()
                .expect("Failed to build queue storage client")
        };

        drain_only_client
            .start()
            .await
            .expect("Failed to start queue storage client");
        wait_for_runtime_capability(
            &pool,
            drain_only_client.runtime_instance_id,
            StorageCapability::QueueStorage,
            Duration::from_secs(5),
        )
        .await;
        assert_eq!(
            active_queue_storage_schema(&pool).await,
            None,
            "prepared queue storage runtime must stay canonical until mixed transition"
        );

        let prepared_short_id =
            insert_available_job(&pool, <CutoverShortJob as JobArgs>::kind(), "cutover").await;
        let queue_storage_start = Instant::now();
        while queue_storage_short_seen.load(Ordering::SeqCst) == 0 {
            assert!(
                queue_storage_start.elapsed() <= Duration::from_secs(5),
                "queue-storage-capable runtime failed to process canonical work before mixed transition"
            );
            tokio::time::sleep(Duration::from_millis(25)).await;
        }
        wait_for_state(
            &pool,
            prepared_short_id,
            "completed",
            Duration::from_secs(5),
        )
        .await;

        // The mixed-transition gate now requires at least one runtime
        // running with `transition_role=queue_storage_target` (auto-role
        // runtimes downgrade to drain-only after the routing flip and
        // would leave the cluster with no queue-storage executor). This
        // test focuses on the canonical-drain side; insert a fake target
        // row so the gate passes without standing up a second real client.
        insert_fake_queue_storage_target(&pool).await;

        storage::enter_mixed_transition(&pool)
            .await
            .expect("enter_mixed_transition should succeed once only 0.6 workers remain");
        assert_eq!(
            active_queue_storage_schema(&pool).await,
            Some(queue_storage_schema.to_string()),
            "mixed transition should activate queue storage routing"
        );
        wait_for_runtime_capability(
            &pool,
            drain_only_client.runtime_instance_id,
            StorageCapability::CanonicalDrainOnly,
            Duration::from_secs(5),
        )
        .await;

        let canonical_drain_id =
            insert_canonical_available_job(&pool, <CutoverShortJob as JobArgs>::kind(), "cutover")
                .await;
        wait_for_state(
            &pool,
            canonical_drain_id,
            "completed",
            Duration::from_secs(5),
        )
        .await;

        drain_only_client.shutdown(Duration::from_secs(5)).await;

        let queue_storage_client = {
            let queue_storage_short_seen = queue_storage_short_seen.clone();
            Client::builder(pool.clone())
                .queue(
                    "cutover",
                    QueueConfig {
                        max_workers: 2,
                        poll_interval: Duration::from_millis(25),
                        ..QueueConfig::default()
                    },
                )
                .queue_storage(
                    store_config.clone(),
                    Duration::from_millis(1_000),
                    Duration::from_millis(50),
                )
                .register::<CutoverShortJob, _, _>(move |_args, _ctx| {
                    let queue_storage_short_seen = queue_storage_short_seen.clone();
                    async move {
                        queue_storage_short_seen.fetch_add(1, Ordering::SeqCst);
                        Ok(JobResult::Completed)
                    }
                })
                .promote_interval(Duration::from_millis(25))
                .leader_election_interval(Duration::from_millis(100))
                .leader_check_interval(Duration::from_millis(50))
                .heartbeat_rescue_interval(Duration::from_millis(100))
                .deadline_rescue_interval(Duration::from_millis(100))
                .callback_rescue_interval(Duration::from_millis(100))
                .runtime_snapshot_interval(Duration::from_millis(100))
                .build()
                .expect("Failed to build post-transition queue storage client")
        };

        queue_storage_client
            .start()
            .await
            .expect("Failed to start post-transition queue storage client");
        wait_for_runtime_capability(
            &pool,
            queue_storage_client.runtime_instance_id,
            StorageCapability::QueueStorage,
            Duration::from_secs(5),
        )
        .await;

        let before_queue_storage = queue_storage_short_seen.load(Ordering::SeqCst);
        let queue_storage_job_id =
            insert_available_job(&pool, <CutoverShortJob as JobArgs>::kind(), "cutover").await;
        let queue_storage_start = Instant::now();
        while queue_storage_short_seen.load(Ordering::SeqCst) == before_queue_storage {
            assert!(
                queue_storage_start.elapsed() <= Duration::from_secs(5),
                "queue storage runtime failed to process new work after cutover"
            );
            tokio::time::sleep(Duration::from_millis(25)).await;
        }
        wait_for_queue_storage_done(
            &pool,
            queue_storage_schema,
            queue_storage_job_id,
            Duration::from_secs(5),
        )
        .await;

        queue_storage_client.shutdown(Duration::from_secs(5)).await;
    }

    #[tokio::test]
    async fn queue_storage_target_started_before_mixed_transition_processes_new_work_immediately() {
        let _guard = test_mutex().lock().await;
        let pool = setup_pool(8).await;
        let queue_storage_schema = "awa_cutover_target_runtime";
        reset_schema(&pool).await;
        migrations::run(&pool)
            .await
            .expect("fresh 0.6 schema install should succeed");
        drop_queue_storage_schema(&pool, queue_storage_schema).await;

        let canonical_seen = Arc::new(AtomicUsize::new(0));
        let queue_storage_seen = Arc::new(AtomicUsize::new(0));
        // See the matching note in
        // canonical_runtime_drains_in_flight_jobs_across_schema_upgrade_before_queue_storage_cutover:
        // pin the legacy materialization path so a 60s deadline_duration
        // on QueueConfig::default() doesn't collide with receipt-plane
        // mode. The flag must be set when `prepare_schema` runs because
        // the receipts-vs-legacy claim CTE is baked into the SQL
        // function definition at that point.
        let store_config = QueueStorageConfig {
            schema: queue_storage_schema.to_string(),
            queue_slot_count: 4,
            lease_slot_count: 2,
            lease_claim_receipts: false,
            ..Default::default()
        };

        prepare_queue_storage_transition_with_config(&pool, store_config.clone()).await;
        assert_eq!(
            active_queue_storage_schema(&pool).await,
            None,
            "prepare should not activate queue storage routing"
        );

        let auto_client = {
            let canonical_seen = canonical_seen.clone();
            Client::builder(pool.clone())
                .queue(
                    "cutover",
                    QueueConfig {
                        max_workers: 2,
                        poll_interval: Duration::from_millis(25),
                        ..QueueConfig::default()
                    },
                )
                .queue_storage(
                    store_config.clone(),
                    Duration::from_millis(1_000),
                    Duration::from_millis(50),
                )
                .register::<CutoverShortJob, _, _>(move |_args, _ctx| {
                    let canonical_seen = canonical_seen.clone();
                    async move {
                        canonical_seen.fetch_add(1, Ordering::SeqCst);
                        Ok(JobResult::Completed)
                    }
                })
                .promote_interval(Duration::from_millis(25))
                .leader_election_interval(Duration::from_millis(100))
                .leader_check_interval(Duration::from_millis(50))
                .heartbeat_rescue_interval(Duration::from_millis(100))
                .deadline_rescue_interval(Duration::from_millis(100))
                .callback_rescue_interval(Duration::from_millis(100))
                .runtime_snapshot_interval(Duration::from_millis(100))
                .build()
                .expect("Failed to build auto cutover client")
        };
        auto_client
            .start()
            .await
            .expect("Failed to start auto cutover client");

        let target_client = {
            let queue_storage_seen = queue_storage_seen.clone();
            Client::builder(pool.clone())
                .queue(
                    "cutover",
                    QueueConfig {
                        max_workers: 2,
                        poll_interval: Duration::from_millis(25),
                        ..QueueConfig::default()
                    },
                )
                .queue_storage(
                    store_config.clone(),
                    Duration::from_millis(1_000),
                    Duration::from_millis(50),
                )
                .transition_role(TransitionWorkerRole::QueueStorageTarget)
                .register::<CutoverShortJob, _, _>(move |_args, _ctx| {
                    let queue_storage_seen = queue_storage_seen.clone();
                    async move {
                        queue_storage_seen.fetch_add(1, Ordering::SeqCst);
                        Ok(JobResult::Completed)
                    }
                })
                .promote_interval(Duration::from_millis(25))
                .leader_election_interval(Duration::from_millis(100))
                .leader_check_interval(Duration::from_millis(50))
                .heartbeat_rescue_interval(Duration::from_millis(100))
                .deadline_rescue_interval(Duration::from_millis(100))
                .callback_rescue_interval(Duration::from_millis(100))
                .runtime_snapshot_interval(Duration::from_millis(100))
                .build()
                .expect("Failed to build queue-storage target client")
        };
        target_client
            .start()
            .await
            .expect("Failed to start queue-storage target client");

        wait_for_runtime_capability(
            &pool,
            auto_client.runtime_instance_id,
            StorageCapability::QueueStorage,
            Duration::from_secs(5),
        )
        .await;
        wait_for_runtime_capability(
            &pool,
            target_client.runtime_instance_id,
            StorageCapability::QueueStorage,
            Duration::from_secs(5),
        )
        .await;

        let canonical_job_id =
            insert_available_job(&pool, <CutoverShortJob as JobArgs>::kind(), "cutover").await;
        let canonical_start = Instant::now();
        while canonical_seen.load(Ordering::SeqCst) == 0 {
            assert!(
                canonical_start.elapsed() <= Duration::from_secs(5),
                "auto client failed to process canonical work before mixed transition"
            );
            tokio::time::sleep(Duration::from_millis(25)).await;
        }
        wait_for_state(&pool, canonical_job_id, "completed", Duration::from_secs(5)).await;
        assert_eq!(
            queue_storage_seen.load(Ordering::SeqCst),
            0,
            "queue-storage target should stay idle before routing flips"
        );

        storage::enter_mixed_transition(&pool)
            .await
            .expect("enter_mixed_transition should succeed with prepared 0.6 fleet");
        wait_for_runtime_capability(
            &pool,
            auto_client.runtime_instance_id,
            StorageCapability::CanonicalDrainOnly,
            Duration::from_secs(5),
        )
        .await;
        wait_for_runtime_capability(
            &pool,
            target_client.runtime_instance_id,
            StorageCapability::QueueStorage,
            Duration::from_secs(5),
        )
        .await;
        assert_eq!(
            active_queue_storage_schema(&pool).await,
            Some(queue_storage_schema.to_string()),
            "mixed transition should activate queue storage routing"
        );

        let before_queue_storage = queue_storage_seen.load(Ordering::SeqCst);
        let queue_storage_job_id =
            insert_available_job(&pool, <CutoverShortJob as JobArgs>::kind(), "cutover").await;
        let queue_storage_start = Instant::now();
        while queue_storage_seen.load(Ordering::SeqCst) == before_queue_storage {
            assert!(
                queue_storage_start.elapsed() <= Duration::from_secs(5),
                "queue-storage target failed to process new work after routing flip"
            );
            tokio::time::sleep(Duration::from_millis(25)).await;
        }
        wait_for_queue_storage_done(
            &pool,
            queue_storage_schema,
            queue_storage_job_id,
            Duration::from_secs(5),
        )
        .await;

        let canonical_drain_id =
            insert_canonical_available_job(&pool, <CutoverShortJob as JobArgs>::kind(), "cutover")
                .await;
        wait_for_state(
            &pool,
            canonical_drain_id,
            "completed",
            Duration::from_secs(5),
        )
        .await;

        target_client.shutdown(Duration::from_secs(5)).await;
        auto_client.shutdown(Duration::from_secs(5)).await;
    }
}