napi-async-runtime 0.2.2

//! napi adapter: registers the shared scheduler as the addon's
//! [`AsyncRuntime`] backend and exports the JavaScript-facing host protocol
//! (CurrentThread task host, JS timer host + relay, config/metrics).
//!
//! Hosts call [`install`] from their own `#[napi_derive::module_init]` hook.
//! This module deliberately ships no `module_init` of its own: the host must
//! resolve its runtime configuration (environment variables, defaults) first
//! and stay in charge of that resolution order.

use std::{future::Future, pin::Pin, ptr};

use napi::bindgen_prelude::{
  AsyncRuntime, AsyncRuntimeRejection, AsyncRuntimeTask, FnArgs, Promise, Unknown,
  register_async_runtime,
};
use napi::threadsafe_function::ThreadsafeFunctionCallMode;
use napi_derive::napi;

use crate::MAX_ASYNC_RUNTIME_WORKER_THREADS;
use crate::async_runtime::{
  CurrentThreadTaskDelivery, CurrentThreadTaskDriver, CurrentThreadTaskDriverId, RuntimeFlavor,
  RuntimeMetricsSnapshot, RuntimeOptions, RuntimeOptionsPatch, TimerDriver, TimerDriverId, TimerId,
  acknowledge_current_thread_task_delivery, configure, configure_partial, configured_options,
  drive_current_thread_tasks, fail_current_thread_task_delivery, metrics,
  register_current_thread_task_driver, register_timer_driver, request_current_thread_task_drain,
  reset_metrics, shutdown, start, try_block_on_dyn, try_spawn, try_spawn_blocking,
  try_spawn_detached, unregister_current_thread_task_driver, unregister_timer_driver,
};
use crate::js_callback::{InvalidReturnValue, JsCallback};

struct SharedAsyncRuntime;

// SAFETY: See the implementation notes in this crate's README. Shutdown closes
// admission, waits for the scheduler generation to quiesce, joins native
// workers, and releases active resources. Independently, napi-rs permanently
// retains the native image after a module that registered this backend exports
// successfully, so externally cloned wakers cannot call into unmapped code.
unsafe impl AsyncRuntime for SharedAsyncRuntime {
  fn spawn(
    &self,
    task: AsyncRuntimeTask,
  ) -> std::result::Result<(), AsyncRuntimeRejection<AsyncRuntimeTask>> {
    match try_spawn(task) {
      Ok(handle) => {
        handle.detach();
        Ok(())
      }
      Err((error, task)) => Err(AsyncRuntimeRejection::new(
        task,
        napi::Error::from_reason(error.to_string()),
      )),
    }
  }

  fn block_on(&self, future: Pin<&mut dyn Future<Output = ()>>) -> napi::Result<()> {
    try_block_on_dyn(future).map_err(|error| napi::Error::from_reason(error.to_string()))
  }

  fn spawn_blocking(
    &self,
    work: Box<dyn FnOnce() + Send + 'static>,
  ) -> std::result::Result<(), AsyncRuntimeRejection<Box<dyn FnOnce() + Send + 'static>>> {
    // Route blocking work submitted through this SPI to the same bounded lane
    // as the scheduler's spawn_blocking facade. See the implementation notes in this crate's README.
    match try_spawn_blocking(work) {
      Ok(handle) => {
        handle.detach();
        Ok(())
      }
      Err((error, work)) => Err(AsyncRuntimeRejection::new(
        work,
        napi::Error::from_reason(error.to_string()),
      )),
    }
  }

  fn start(&self) -> napi::Result<()> {
    start().map_err(|error| napi::Error::from_reason(error.to_string()))
  }

  fn shutdown(&self) -> napi::Result<()> {
    shutdown().map_err(|error| napi::Error::from_reason(error.to_string()))
  }
}

#[napi(string_enum)]
#[derive(Clone, Copy)]
pub enum BindingRuntimeFlavor {
  CurrentThread,
  MultiThread,
}

impl From<BindingRuntimeFlavor> for RuntimeFlavor {
  fn from(value: BindingRuntimeFlavor) -> Self {
    match value {
      BindingRuntimeFlavor::CurrentThread => Self::CurrentThread,
      BindingRuntimeFlavor::MultiThread => Self::MultiThread,
    }
  }
}

impl From<RuntimeFlavor> for BindingRuntimeFlavor {
  fn from(value: RuntimeFlavor) -> Self {
    match value {
      RuntimeFlavor::CurrentThread => Self::CurrentThread,
      RuntimeFlavor::MultiThread => Self::MultiThread,
    }
  }
}

#[napi(object)]
pub struct BindingRuntimeOptions {
  pub flavor: Option<BindingRuntimeFlavor>,
  /// Positive integer worker count. Values above 256 are rejected.
  pub worker_threads: Option<f64>,
  /// Positive integer blocking-task limit. Values above 256 are rejected.
  pub max_blocking_tasks: Option<f64>,
}

impl TryFrom<BindingRuntimeOptions> for RuntimeOptionsPatch {
  type Error = napi::Error;

  fn try_from(value: BindingRuntimeOptions) -> Result<Self, Self::Error> {
    Ok(Self {
      flavor: value.flavor.map(Into::into),
      worker_threads: validate_binding_thread_count(
        "workerThreads",
        value.worker_threads,
        MAX_ASYNC_RUNTIME_WORKER_THREADS,
      )?,
      max_blocking_tasks: validate_binding_thread_count(
        "maxBlockingTasks",
        value.max_blocking_tasks,
        MAX_ASYNC_RUNTIME_WORKER_THREADS,
      )?,
      // Not exposed through the JavaScript configure surface; the addon-load
      // resolution (`install` + `DRAIN_LINGER_ENV`) owns this knob.
      drain_linger: None,
    })
  }
}

fn validate_binding_thread_count(
  field: &str,
  value: Option<f64>,
  maximum: usize,
) -> napi::Result<Option<usize>> {
  value
    .map(|count| {
      #[expect(
        clippy::cast_precision_loss,
        reason = "the small production thread limit is exactly representable as f64"
      )]
      let maximum_as_f64 = maximum as f64;
      if !count.is_finite() || count < 1.0 || count.fract() != 0.0 || count > maximum_as_f64 {
        return Err(napi::Error::from_reason(format!(
          "`{field}` must be a positive integer no greater than {maximum}"
        )));
      }

      #[expect(
        clippy::cast_possible_truncation,
        clippy::cast_sign_loss,
        reason = "the finite positive integer was range-checked against the usize-backed limit"
      )]
      let count = count as usize;
      Ok(count)
    })
    .transpose()
}

#[napi(object)]
pub struct BindingRuntimeConfig {
  pub flavor: BindingRuntimeFlavor,
  pub worker_threads: u32,
  pub max_blocking_tasks: u32,
  /// Drainer idle-linger budget in microseconds; `0` means lingering is
  /// disabled. Reported from the validated (ceiling-clamped) options.
  pub drain_linger_us: u32,
}

impl From<RuntimeOptions> for BindingRuntimeConfig {
  fn from(value: RuntimeOptions) -> Self {
    Self {
      flavor: value.flavor.into(),
      worker_threads: saturating_u32(value.worker_threads as u64),
      max_blocking_tasks: saturating_u32(value.max_blocking_tasks as u64),
      drain_linger_us: saturating_u32(
        u64::try_from(value.drain_linger.as_micros()).unwrap_or(u64::MAX),
      ),
    }
  }
}

#[napi(object)]
pub struct BindingRuntimeMetrics {
  pub flavor: BindingRuntimeFlavor,
  pub worker_threads: u32,
  pub max_blocking_tasks: u32,
  pub tasks_spawned: f64,
  pub tasks_completed: f64,
  pub tasks_panicked: f64,
  pub runnable_schedules: f64,
  pub runnable_polls: f64,
  pub queued_runnables: f64,
  pub max_queued_runnables: f64,
  pub active_runnables: f64,
  pub max_active_runnables: f64,
  pub blocking_tasks_started: f64,
  pub blocking_tasks_completed: f64,
  pub active_blocking_tasks: f64,
  pub max_active_blocking_tasks: f64,
}

impl From<RuntimeMetricsSnapshot> for BindingRuntimeMetrics {
  fn from(value: RuntimeMetricsSnapshot) -> Self {
    Self {
      flavor: value.flavor.into(),
      worker_threads: saturating_u32(value.worker_threads as u64),
      max_blocking_tasks: saturating_u32(value.max_blocking_tasks as u64),
      tasks_spawned: safe_js_number(value.tasks_spawned),
      tasks_completed: safe_js_number(value.tasks_completed),
      tasks_panicked: safe_js_number(value.tasks_panicked),
      runnable_schedules: safe_js_number(value.runnable_schedules),
      runnable_polls: safe_js_number(value.runnable_polls),
      queued_runnables: safe_js_number(value.queued_runnables),
      max_queued_runnables: safe_js_number(value.max_queued_runnables),
      active_runnables: safe_js_number(value.active_runnables),
      max_active_runnables: safe_js_number(value.max_active_runnables),
      blocking_tasks_started: safe_js_number(value.blocking_tasks_started),
      blocking_tasks_completed: safe_js_number(value.blocking_tasks_completed),
      active_blocking_tasks: safe_js_number(value.active_blocking_tasks),
      max_active_blocking_tasks: safe_js_number(value.max_active_blocking_tasks),
    }
  }
}

fn saturating_u32(value: u64) -> u32 {
  u32::try_from(value).unwrap_or(u32::MAX)
}

const MAX_SAFE_JS_INTEGER: u64 = (1_u64 << 53) - 1;

#[expect(
  clippy::cast_precision_loss,
  reason = "the value is clamped to JavaScript's exactly representable integer range"
)]
fn safe_js_number(value: u64) -> f64 {
  value.min(MAX_SAFE_JS_INTEGER) as f64
}

#[napi]
/// Override the shared async runtime's flavor and thread counts.
///
/// Must be called before the first async binding call.
pub fn configure_async_runtime(options: BindingRuntimeOptions) -> napi::Result<()> {
  configure_partial(options.try_into()?)
    .map_err(|error| napi::Error::from_reason(error.to_string()))
}

#[napi]
/// Return the effective async runtime configuration.
///
/// Reports the runtime controller's validated options, including a
/// pre-first-use `configureAsyncRuntime` override. The environment is never
/// re-read.
pub fn get_async_runtime_config() -> BindingRuntimeConfig {
  configured_options().into()
}

#[napi]
/// Return a snapshot of the shared async runtime's task and scheduler counters.
pub fn get_async_runtime_metrics() -> BindingRuntimeMetrics {
  metrics().into()
}

#[napi]
/// Reset cumulative async runtime event counters to zero.
///
/// Live gauges and their lifetime high-water marks are preserved so active
/// task guards can complete without corrupting concurrent observations.
pub fn reset_async_runtime_metrics() {
  reset_metrics();
}

#[napi(object)]
pub struct BindingHostRegistration {
  pub high: u32,
  pub low: u32,
}

impl BindingHostRegistration {
  fn from_id(id: u64) -> Self {
    Self {
      high: (id >> 32) as u32,
      low: u32::try_from(id & u64::from(u32::MAX))
        .expect("the masked host registration low word must fit in u32"),
    }
  }

  fn id(high: u32, low: u32) -> u64 {
    (u64::from(high) << 32) | u64::from(low)
  }
}

static NEXT_HOST_REGISTRATION_ID: std::sync::atomic::AtomicU64 =
  std::sync::atomic::AtomicU64::new(1);

static RESERVED_HOST_REGISTRATIONS: std::sync::LazyLock<
  std::sync::Mutex<rustc_hash::FxHashSet<u64>>,
> = std::sync::LazyLock::new(|| std::sync::Mutex::new(rustc_hash::FxHashSet::default()));

fn reserve_host_registration_id() -> napi::Result<u64> {
  let id = NEXT_HOST_REGISTRATION_ID
    .fetch_update(
      std::sync::atomic::Ordering::SeqCst,
      std::sync::atomic::Ordering::SeqCst,
      |current| current.checked_add(1),
    )
    .map_err(|_| {
      napi::Error::new(
        napi::Status::GenericFailure,
        "JavaScript host registration id space exhausted",
      )
    })?;
  RESERVED_HOST_REGISTRATIONS
    .lock()
    .unwrap_or_else(std::sync::PoisonError::into_inner)
    .insert(id);
  Ok(id)
}

fn claim_host_registration_id(registration_high: u32, registration_low: u32) -> napi::Result<u64> {
  let id = BindingHostRegistration::id(registration_high, registration_low);
  if RESERVED_HOST_REGISTRATIONS
    .lock()
    .unwrap_or_else(std::sync::PoisonError::into_inner)
    .remove(&id)
  {
    Ok(id)
  } else {
    Err(napi::Error::new(
      napi::Status::InvalidArg,
      "CurrentThread host registration was not reserved or was already consumed",
    ))
  }
}

fn release_host_registration_id(id: u64) {
  RESERVED_HOST_REGISTRATIONS
    .lock()
    .unwrap_or_else(std::sync::PoisonError::into_inner)
    .remove(&id);
}

#[napi]
/// Reserve an exact CurrentThread host registration capability before either
/// task or timer installation performs side effects. The JavaScript package
/// validates the returned words and passes them back to one registration call.
pub fn reserve_current_thread_host_registration() -> napi::Result<BindingHostRegistration> {
  reserve_host_registration_id().map(BindingHostRegistration::from_id)
}

fn current_thread_task_host_napi_result(
  status: napi::sys::napi_status,
  context: &'static str,
) -> napi::Result<()> {
  if status == napi::sys::Status::napi_ok {
    Ok(())
  } else {
    Err(napi::Error::new(napi::Status::from(status), context))
  }
}

fn contain_current_thread_task_host_unwind<T>(operation: impl FnOnce() -> T) -> Option<T> {
  match std::panic::catch_unwind(std::panic::AssertUnwindSafe(operation)) {
    Ok(value) => Some(value),
    Err(payload) => {
      if let Err(nested) = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| drop(payload)))
      {
        std::mem::forget(nested);
      }
      None
    }
  }
}

#[cfg(test)]
thread_local! {
  static NATIVE_TASK_HOST_AFTER_DRIVE_TEST_HOOK:
    std::cell::RefCell<Option<Box<dyn FnOnce()>>> =
      const { std::cell::RefCell::new(None) };
  static NATIVE_TASK_HOST_AFTER_PAYLOAD_DROP_TEST_HOOK:
    std::cell::RefCell<Option<Box<dyn FnOnce()>>> =
      const { std::cell::RefCell::new(None) };
}

#[cfg(test)]
fn run_native_task_host_after_drive_test_hook() {
  if let Some(hook) = NATIVE_TASK_HOST_AFTER_DRIVE_TEST_HOOK.with(|slot| slot.borrow_mut().take()) {
    hook();
  }
}

#[cfg(test)]
fn run_native_task_host_after_payload_drop_test_hook() {
  if let Some(hook) =
    NATIVE_TASK_HOST_AFTER_PAYLOAD_DROP_TEST_HOOK.with(|slot| slot.borrow_mut().take())
  {
    hook();
  }
}

unsafe extern "C" fn finalize_native_current_thread_task_host(
  _env: napi::sys::napi_env,
  finalize_data: *mut std::ffi::c_void,
  _finalize_hint: *mut std::ffi::c_void,
) {
  if finalize_data.is_null() {
    return;
  }
  let _ = contain_current_thread_task_host_unwind(|| {
    let weak = unsafe {
      std::sync::Weak::<NativeCurrentThreadTaskHostInner>::from_raw(finalize_data.cast())
    };
    if let Some(inner) = weak.upgrade() {
      inner.finalized();
    }
  });
}

unsafe extern "C" fn call_native_current_thread_task_host(
  env: napi::sys::napi_env,
  _js_callback: napi::sys::napi_value,
  _context: *mut std::ffi::c_void,
  data: *mut std::ffi::c_void,
) {
  if data.is_null() {
    return;
  }

  let payload = unsafe { Box::<NativeCurrentThreadTaskHostPayload>::from_raw(data.cast()) };
  let delivery = payload.delivery;
  let callback_lease = if env.is_null() {
    None
  } else {
    contain_current_thread_task_host_unwind(|| {
      let lease = drive_current_thread_tasks(delivery.capability());
      #[cfg(test)]
      if lease.is_some() {
        run_native_task_host_after_drive_test_hook();
      }
      lease
    })
    .flatten()
  };
  let claimed = callback_lease.is_some();
  let completed = contain_current_thread_task_host_unwind(|| {
    if claimed {
      acknowledge_current_thread_task_delivery(delivery);
    } else {
      fail_current_thread_task_delivery(delivery);
    }
  });
  if completed.is_none() {
    let _ = contain_current_thread_task_host_unwind(|| {
      fail_current_thread_task_delivery(delivery);
    });
  }
  let _ = contain_current_thread_task_host_unwind(|| drop(payload));
  #[cfg(test)]
  run_native_task_host_after_payload_drop_test_hook();
  let _ = contain_current_thread_task_host_unwind(|| drop(callback_lease));
}

struct NativeCurrentThreadTaskHostPayload {
  delivery: CurrentThreadTaskDelivery,
  #[cfg(test)]
  drop_observer: Option<std::sync::Arc<std::sync::atomic::AtomicUsize>>,
}

impl NativeCurrentThreadTaskHostPayload {
  fn new(delivery: CurrentThreadTaskDelivery) -> Self {
    Self {
      delivery,
      #[cfg(test)]
      drop_observer: None,
    }
  }

  #[cfg(test)]
  fn with_drop_observer(
    delivery: CurrentThreadTaskDelivery,
    drop_observer: std::sync::Arc<std::sync::atomic::AtomicUsize>,
  ) -> Self {
    Self {
      delivery,
      drop_observer: Some(drop_observer),
    }
  }
}

#[cfg(test)]
impl Drop for NativeCurrentThreadTaskHostPayload {
  fn drop(&mut self) {
    if let Some(observer) = &self.drop_observer {
      observer.fetch_add(1, std::sync::atomic::Ordering::SeqCst);
    }
  }
}

// See the implementation notes in this crate's README.
struct NativeCurrentThreadTaskHost {
  inner: std::sync::Arc<NativeCurrentThreadTaskHostInner>,
}

struct NativeCurrentThreadTaskHostInner {
  threadsafe_function: std::sync::Mutex<Option<usize>>,
  dead: std::sync::atomic::AtomicBool,
  environment_closing: std::sync::atomic::AtomicBool,
  host_registration: u64,
  registration: std::sync::Mutex<Option<CurrentThreadTaskDriverId>>,
}

static NATIVE_CURRENT_THREAD_TASK_HOSTS: std::sync::LazyLock<
  std::sync::Mutex<rustc_hash::FxHashMap<u64, std::sync::Weak<NativeCurrentThreadTaskHostInner>>>,
> = std::sync::LazyLock::new(|| std::sync::Mutex::new(rustc_hash::FxHashMap::default()));

fn registered_current_thread_task_host(
  id: u64,
) -> Option<std::sync::Arc<NativeCurrentThreadTaskHostInner>> {
  let mut registrations = NATIVE_CURRENT_THREAD_TASK_HOSTS
    .lock()
    .unwrap_or_else(std::sync::PoisonError::into_inner);
  let inner = registrations.get(&id).and_then(std::sync::Weak::upgrade);
  if inner.is_none() {
    registrations.remove(&id);
  }
  inner
}

impl NativeCurrentThreadTaskHostInner {
  fn new(env: &napi::Env, host_registration: u64) -> napi::Result<std::sync::Arc<Self>> {
    const ASYNC_RESOURCE_NAME: &[u8] = b"napi_async_runtime_current_thread_task_host";
    let name_len = isize::try_from(ASYNC_RESOURCE_NAME.len())
      .expect("the CurrentThread task-host resource name length must fit");
    let mut async_resource_name = ptr::null_mut();
    current_thread_task_host_napi_result(
      unsafe {
        napi::sys::napi_create_string_utf8(
          env.raw(),
          ASYNC_RESOURCE_NAME.as_ptr().cast(),
          name_len,
          &raw mut async_resource_name,
        )
      },
      "Failed to create the CurrentThread task-host async resource name",
    )?;

    let inner = std::sync::Arc::new(Self {
      threadsafe_function: std::sync::Mutex::default(),
      dead: std::sync::atomic::AtomicBool::new(false),
      environment_closing: std::sync::atomic::AtomicBool::new(false),
      host_registration,
      registration: std::sync::Mutex::default(),
    });
    let finalize_data = std::sync::Weak::into_raw(std::sync::Arc::downgrade(&inner))
      .cast_mut()
      .cast();
    let mut threadsafe_function = ptr::null_mut();
    let create_status = unsafe {
      napi::sys::napi_create_threadsafe_function(
        env.raw(),
        ptr::null_mut(),
        ptr::null_mut(),
        async_resource_name,
        1,
        1,
        finalize_data,
        Some(finalize_native_current_thread_task_host),
        ptr::null_mut(),
        Some(call_native_current_thread_task_host),
        &raw mut threadsafe_function,
      )
    };
    if create_status != napi::sys::Status::napi_ok {
      unsafe {
        drop(std::sync::Weak::<Self>::from_raw(finalize_data.cast()));
      }
      return Err(napi::Error::new(
        napi::Status::from(create_status),
        "Failed to create the native CurrentThread task host",
      ));
    }
    *inner
      .threadsafe_function
      .lock()
      .unwrap_or_else(std::sync::PoisonError::into_inner) = Some(threadsafe_function as usize);

    if let Err(error) = current_thread_task_host_napi_result(
      unsafe { napi::sys::napi_unref_threadsafe_function(env.raw(), threadsafe_function) },
      "Failed to unref the native CurrentThread task host",
    ) {
      inner.release_threadsafe_function(napi::sys::ThreadsafeFunctionReleaseMode::abort);
      return Err(error);
    }
    Ok(inner)
  }

  fn is_live(&self) -> bool {
    !self.dead.load(std::sync::atomic::Ordering::SeqCst)
      && !self
        .environment_closing
        .load(std::sync::atomic::Ordering::SeqCst)
      && self
        .threadsafe_function
        .lock()
        .unwrap_or_else(std::sync::PoisonError::into_inner)
        .is_some()
  }

  fn call_threadsafe_function_with(
    &self,
    data: *mut std::ffi::c_void,
    call: impl FnOnce(
      napi::sys::napi_threadsafe_function,
      *mut std::ffi::c_void,
    ) -> napi::sys::napi_status,
  ) -> Option<napi::sys::napi_status> {
    let mut slot = self
      .threadsafe_function
      .lock()
      .unwrap_or_else(std::sync::PoisonError::into_inner);
    if self.dead.load(std::sync::atomic::Ordering::SeqCst)
      || self
        .environment_closing
        .load(std::sync::atomic::Ordering::SeqCst)
    {
      return None;
    }
    let threadsafe_function = (*slot)? as napi::sys::napi_threadsafe_function;
    let status = call(threadsafe_function, data);
    if status == napi::sys::Status::napi_closing {
      // Node-API decrements this caller's acquisition before returning
      // `napi_closing`; the pointer is no longer safe for any later API call.
      self
        .environment_closing
        .store(true, std::sync::atomic::Ordering::SeqCst);
      slot.take();
    }
    Some(status)
  }

  fn dispatch(&self, delivery: CurrentThreadTaskDelivery) -> bool {
    let data: *mut std::ffi::c_void =
      Box::into_raw(Box::new(NativeCurrentThreadTaskHostPayload::new(delivery))).cast();
    let Some(status) =
      self.call_threadsafe_function_with(data, |threadsafe_function, data| unsafe {
        napi::sys::napi_call_threadsafe_function(
          threadsafe_function,
          data,
          napi::sys::ThreadsafeFunctionCallMode::nonblocking,
        )
      })
    else {
      unsafe {
        drop(Box::<NativeCurrentThreadTaskHostPayload>::from_raw(
          data.cast(),
        ));
      }
      return false;
    };
    if status != napi::sys::Status::napi_ok {
      unsafe {
        drop(Box::<NativeCurrentThreadTaskHostPayload>::from_raw(
          data.cast(),
        ));
      }
      return false;
    }
    true
  }

  fn finalized(&self) {
    self
      .environment_closing
      .store(true, std::sync::atomic::Ordering::SeqCst);
    // Finalization means Node is already destroying the TSFN. Invalidate the
    // pointer without calling back into Node from its own finalizer.
    self
      .threadsafe_function
      .lock()
      .unwrap_or_else(std::sync::PoisonError::into_inner)
      .take();
    self.evict_inner(true, false);
  }

  fn environment_cleanup(&self) {
    self
      .environment_closing
      .store(true, std::sync::atomic::Ordering::SeqCst);
    // Cleanup is registered after the TSFN's own cleanup hook, so it normally
    // runs first. Keep owner retirement independent from registry eviction:
    // even a contained eviction panic must not retain the initial acquisition.
    let _ = contain_current_thread_task_host_unwind(|| self.evict_inner(true, false));
    self.release_threadsafe_function(napi::sys::ThreadsafeFunctionReleaseMode::release);
  }

  fn evict_after_sweep(&self) {
    let abort = !self
      .environment_closing
      .load(std::sync::atomic::Ordering::SeqCst);
    self.evict_inner(false, abort);
  }

  fn rollback(&self) {
    self.evict_inner(true, true);
  }

  fn evict_inner(&self, request_redispatch: bool, abort: bool) {
    self.dead.store(true, std::sync::atomic::Ordering::SeqCst);
    NATIVE_CURRENT_THREAD_TASK_HOSTS
      .lock()
      .unwrap_or_else(std::sync::PoisonError::into_inner)
      .remove(&self.host_registration);
    let registration = self
      .registration
      .lock()
      .unwrap_or_else(std::sync::PoisonError::into_inner)
      .take();
    if let Some(id) = registration {
      unregister_current_thread_task_driver(id);
      if request_redispatch {
        // The dead host may have accepted a weak-TSFN callback that its env
        // discarded before delivery. Republish the same internal capability to
        // remaining live hosts through their exact delivery identities.
        request_current_thread_task_drain();
      }
    }
    if abort {
      self.release_threadsafe_function(napi::sys::ThreadsafeFunctionReleaseMode::abort);
    }
  }

  fn release_threadsafe_function(&self, mode: napi::sys::napi_threadsafe_function_release_mode) {
    self.release_threadsafe_function_with(mode, |threadsafe_function, mode| unsafe {
      napi::sys::napi_release_threadsafe_function(threadsafe_function, mode)
    });
  }

  fn release_threadsafe_function_with(
    &self,
    mode: napi::sys::napi_threadsafe_function_release_mode,
    release: impl FnOnce(
      napi::sys::napi_threadsafe_function,
      napi::sys::napi_threadsafe_function_release_mode,
    ) -> napi::sys::napi_status,
  ) {
    let threadsafe_function = self
      .threadsafe_function
      .lock()
      .unwrap_or_else(std::sync::PoisonError::into_inner)
      .take();
    let Some(threadsafe_function) = threadsafe_function else {
      return;
    };
    let status = release(
      threadsafe_function as napi::sys::napi_threadsafe_function,
      mode,
    );
    if status == napi::sys::Status::napi_closing {
      self
        .environment_closing
        .store(true, std::sync::atomic::Ordering::SeqCst);
    }
  }
}

impl CurrentThreadTaskDriver for NativeCurrentThreadTaskHost {
  fn dispatch(&self, delivery: CurrentThreadTaskDelivery) -> bool {
    self.inner.dispatch(delivery)
  }

  fn is_live(&self) -> bool {
    self.inner.is_live()
  }

  fn on_swept(&self) {
    // The registry is already selecting/dispatching a fallback. Avoid
    // recursively starting another selection pass from its sweep callback.
    self.inner.evict_after_sweep();
  }
}

fn reject_current_thread_task_host_callback(dispatch: Option<Unknown<'_>>) -> napi::Result<()> {
  if dispatch.is_none() {
    Ok(())
  } else {
    Err(napi::Error::new(
      napi::Status::InvalidArg,
      "registerCurrentThreadTaskHost does not accept a JavaScript callback",
    ))
  }
}

const CURRENT_THREAD_TASK_HOST_CONTRACT_VERSION: u32 = 4;

#[napi]
/// Return the native CurrentThread task-host ABI expected by the JavaScript
/// package before it invokes either async-runtime host registration. Version 4
/// reserves and validates an exact registration capability before host
/// installation performs side effects.
pub fn get_current_thread_task_host_contract_version() -> u32 {
  CURRENT_THREAD_TASK_HOST_CONTRACT_VERSION
}

#[napi]
/// Return whether one exact CurrentThread task- or timer-host registration is
/// still live. The JavaScript package revalidates its process-global marker on
/// every module evaluation so native eviction cannot leave a stale installed
/// bit that permanently suppresses replacement registration.
pub fn is_current_thread_host_registration_active(
  registration_high: u32,
  registration_low: u32,
) -> bool {
  let id = BindingHostRegistration::id(registration_high, registration_low);
  registered_current_thread_task_host(id).is_some_and(|inner| inner.is_live())
    || registered_timer_host(id).is_some_and(|inner| inner.is_live())
}

fn install_cleanup_hook_or_rollback<T>(
  install: impl FnOnce() -> napi::Result<T>,
  rollback: impl FnOnce(),
) -> napi::Result<()> {
  match install() {
    Ok(_) => Ok(()),
    Err(error) => {
      rollback();
      Err(error)
    }
  }
}

fn install_host_driver_registration<T>(
  dead: &std::sync::atomic::AtomicBool,
  registration: &std::sync::Mutex<Option<T>>,
  install: impl FnOnce() -> T,
  publish: impl FnOnce(),
  rollback: impl FnOnce(T),
) -> bool {
  // `install` may synchronously wake arbitrary task wakers. Keep it outside
  // the exact-registration mutex so a reentrant liveness sweep can evict this
  // host instead of deadlocking on the half-published registration.
  let mut installed = Some(install());
  {
    let mut slot = registration
      .lock()
      .unwrap_or_else(std::sync::PoisonError::into_inner);
    if !dead.load(std::sync::atomic::Ordering::SeqCst) {
      *slot = installed.take();
      // Publication is serialized with eviction by `registration`. This
      // callback must only update binding-owned indexes.
      publish();
    }
  }
  if let Some(installed) = installed {
    rollback(installed);
    false
  } else {
    true
  }
}

#[napi(ts_args_type = "registrationHigh: number, registrationLow: number, dispatch?: never")]
/// Install a native-owned host turn used to poll CurrentThread runnables
/// without re-entering arbitrary future waker locks. Called once per importing
/// environment. JavaScript callbacks are rejected synchronously.
pub fn register_current_thread_task_host(
  env: &napi::Env,
  registration_high: u32,
  registration_low: u32,
  dispatch: Option<Unknown<'_>>,
) -> napi::Result<()> {
  reject_current_thread_task_host_callback(dispatch)?;
  let host_registration = claim_host_registration_id(registration_high, registration_low)?;
  let inner = NativeCurrentThreadTaskHostInner::new(env, host_registration)?;
  {
    let mut slot = inner
      .registration
      .lock()
      .unwrap_or_else(std::sync::PoisonError::into_inner);
    *slot = Some(register_current_thread_task_driver(std::sync::Arc::new(
      NativeCurrentThreadTaskHost {
        inner: std::sync::Arc::clone(&inner),
      },
    )));
  }
  NATIVE_CURRENT_THREAD_TASK_HOSTS
    .lock()
    .unwrap_or_else(std::sync::PoisonError::into_inner)
    .insert(host_registration, std::sync::Arc::downgrade(&inner));
  request_current_thread_task_drain();
  let hook_inner = std::sync::Arc::clone(&inner);
  install_cleanup_hook_or_rollback(
    || {
      env.add_env_cleanup_hook(hook_inner, |inner| {
        let _ = contain_current_thread_task_host_unwind(|| inner.environment_cleanup());
      })
    },
    || inner.rollback(),
  )?;
  Ok(())
}

#[napi]
/// Evict exactly one native host installed by `registerCurrentThreadTaskHost`.
/// Managed workerd disposal uses this before environment cleanup so a later
/// throwing cleanup hook cannot leave the process-global driver selected.
pub fn unregister_current_thread_task_host(registration_high: u32, registration_low: u32) {
  let id = BindingHostRegistration::id(registration_high, registration_low);
  release_host_registration_id(id);
  if let Some(inner) = registered_current_thread_task_host(id) {
    inner.rollback();
  }
}

/// Host timer driver for the shared runtime's CurrentThread flavor:
/// `sleep_until` on the single-thread executor cannot park a
/// helper thread (none exists on threadless wasm), so it delegates each timer
/// to the host event loop through the JS callback registered at import --
/// `(id, ms) => new Promise((resolve) => setTimeout(resolve, ms))`, paired
/// with a cancellation callback that clears the timeout and resolves the
/// relay promise.
///
/// Per timer id: the FIRST `register` arms one host timeout via a detached
/// relay task; re-registers (`Sleep` re-polls) only refresh the stored waker.
/// `cancel` removes the pending waker and either invokes the host cancellation
/// callback or leaves a pre-arm tombstone for the accepted TSFN delivery. The
/// JS side clears the timeout and resolves its promise, so a dropped sleep
/// leaves neither a live timeout nor a detached relay task.
///
/// LIFETIME: each importing napi env registers its own
/// host, and a host dies WITH its env -- the weak threadsafe function does
/// not keep a worker's event loop alive, so a worker that imported the
/// binding can exit at any time and orphan its host. A dead host must never
/// keep timer duty (the registry would busy-fail every debounce against it),
/// so it is EVICTED -- proactively by the env-cleanup hook installed at
/// registration, and reactively by the `is_live` probe (the threadsafe
/// function's `aborted` flag) and by relay-call failure. Eviction wakes every
/// sleep armed here so each re-polls onto the registry's next live registrant
/// (see `TimerDriverRegistry`).
struct JsTimerHost {
  inner: std::sync::Arc<JsTimerHostInner>,
}

#[derive(Default)]
struct RelayIdAllocator {
  next: std::sync::atomic::AtomicU64,
}

impl RelayIdAllocator {
  fn reserve(&self) -> Result<u32, RelayIdExhausted> {
    self
      .next
      .fetch_update(
        std::sync::atomic::Ordering::Relaxed,
        std::sync::atomic::Ordering::Relaxed,
        |next| u32::try_from(next).is_ok().then_some(next + 1),
      )
      .map(|id| u32::try_from(id).expect("the checked relay id must fit in u32"))
      .map_err(|_| RelayIdExhausted)
  }
}

#[derive(Clone, Copy, Debug, PartialEq, Eq)]
struct RelayIdExhausted;

#[derive(Clone, Copy, Debug, PartialEq, Eq)]
#[repr(u8)]
enum RelayCancellationAccounting {
  HostHealth = 0,
  CleanupOnly = 1,
}

struct HostTimerRelayHealth {
  cancellation_accounting: std::sync::atomic::AtomicU8,
  failure_recorded: std::sync::atomic::AtomicBool,
}

impl Default for HostTimerRelayHealth {
  fn default() -> Self {
    Self {
      cancellation_accounting: std::sync::atomic::AtomicU8::new(
        RelayCancellationAccounting::HostHealth as u8,
      ),
      failure_recorded: std::sync::atomic::AtomicBool::new(false),
    }
  }
}

impl HostTimerRelayHealth {
  fn set_cancellation_accounting(&self, accounting: RelayCancellationAccounting) {
    self
      .cancellation_accounting
      .store(accounting as u8, std::sync::atomic::Ordering::Release);
  }

  fn cancellation_accounting(&self) -> RelayCancellationAccounting {
    match self
      .cancellation_accounting
      .load(std::sync::atomic::Ordering::Acquire)
    {
      value if value == RelayCancellationAccounting::CleanupOnly as u8 => {
        RelayCancellationAccounting::CleanupOnly
      }
      _ => RelayCancellationAccounting::HostHealth,
    }
  }

  fn claim_failure(&self) -> bool {
    self
      .failure_recorded
      .compare_exchange(
        false,
        true,
        std::sync::atomic::Ordering::AcqRel,
        std::sync::atomic::Ordering::Acquire,
      )
      .is_ok()
  }

  fn failure_recorded(&self) -> bool {
    self
      .failure_recorded
      .load(std::sync::atomic::Ordering::Acquire)
  }
}

/// Consecutive NON-LIFETIME relay failures tolerated on one live host before
/// eviction. A transient failure (a one-off JS rejection, a queueing hiccup)
/// must not poison a live driver -- on a
/// main-only process that would leave NO driver and every later CT sleep
/// would hit the loud no-driver panic. But a PERSISTENTLY failing live
/// callback can never fire a timer either, so after this many consecutive
/// failures the host is evicted anyway (announced in the log). Reset on any
/// successful relay. Small on purpose: each strike costs one wasted arm/wake
/// round-trip for the affected sleep.
const HOST_TIMER_MAX_TRANSIENT_FAILURES: u32 = 3;

/// Does this relay error mean the HOST IS GONE (evict immediately), as
/// opposed to a callback failure on a live host (strike-counted)?
///
/// ERROR CLASSIFICATION: no message strings. A rejected JS promise is
/// coerced into `GenericFailure` CARRYING THE JS REJECTION STRING (pinned
/// napi 3.10, error.rs `From<Unknown> for Error`: native coerces the value
/// to a string, wasm reads `.message` -- both always `GenericFailure`), so
/// any message match is forgeable by a live callback rejecting with a
/// colliding string (e.g. `Error('oneshot canceled')`) and would evict a
/// live host, bypassing the strike budget. Two string-free authorities
/// instead:
/// - `Status::Closing` is LIFETIME: it originates only from the TSFN layer
///   (aborted pre-check, raw `napi_closing`), and the JS coercion above can
///   never produce it -- unforgeable.
/// - Everything else defers to the LIVENESS PROBE (`is_live` = the dead
///   latch + the threadsafe function's own `aborted` flag): the genuine
///   teardown shapes (queue drained at env teardown, env died before the JS
///   promise settled) all coincide with the env being torn down, which the
///   probe observes directly.
///
/// Race walk: env dies between the error and the probe read -> the probe
/// reads dead -> evict: correct. Env alive at the probe but dying a
/// microsecond later -> strike now (the affected sleep is re-woken); the
/// death is then caught by the env-cleanup hook, by the aborted-probe sweep
/// at the next selection, or by the next relay failure (which will probe
/// dead) -> bounded, correct. A LIVE host's failure -- whatever its message
/// says -- takes the strike path.
fn should_evict_for_relay_error(status: napi::Status, host_is_live: bool) -> bool {
  status == napi::Status::Closing || !host_is_live
}

#[derive(Clone, Copy, Debug, PartialEq, Eq)]
enum HostTimerFailureAction {
  Duplicate,
  EvictHost,
  EvictHostAfterStrikes(u32),
  Retry(u32),
}

fn record_host_timer_failure(
  transient_failures: &std::sync::atomic::AtomicU32,
  relay_health: &HostTimerRelayHealth,
  status: napi::Status,
  host_is_live: bool,
) -> HostTimerFailureAction {
  if !relay_health.claim_failure() {
    return HostTimerFailureAction::Duplicate;
  }
  if should_evict_for_relay_error(status, host_is_live) {
    return HostTimerFailureAction::EvictHost;
  }
  let strikes = transient_failures.fetch_add(1, std::sync::atomic::Ordering::SeqCst) + 1;
  if strikes >= HOST_TIMER_MAX_TRANSIENT_FAILURES {
    HostTimerFailureAction::EvictHostAfterStrikes(strikes)
  } else {
    HostTimerFailureAction::Retry(strikes)
  }
}

fn reset_host_timer_failures_after_success(
  transient_failures: &std::sync::atomic::AtomicU32,
  relay_health: &HostTimerRelayHealth,
) {
  if !relay_health.failure_recorded() {
    transient_failures.store(0, std::sync::atomic::Ordering::SeqCst);
  }
}

struct JsTimerHostInner {
  callback: JsCallback<FnArgs<(u32, f64)>, Promise<()>>,
  cancel_callback: JsCallback<FnArgs<(u32,)>, ()>,
  pending: std::sync::Mutex<rustc_hash::FxHashMap<TimerId, PendingHostTimer>>,
  relay_ids: RelayIdAllocator,
  /// Latched by [`JsTimerHostInner::evict`]: this host's env is gone (or its
  /// callback failed) and it must never serve another timer.
  dead: std::sync::atomic::AtomicBool,
  /// Exact JavaScript-facing capability used to unregister this host before
  /// emnapi starts running fallible environment cleanup hooks.
  host_registration: u64,
  /// This host's registration in the global driver registry. Installation and
  /// eviction use this mutex as their publication boundary, but the core
  /// registration call itself runs outside it because that call may wake
  /// arbitrary task wakers.
  registration: std::sync::Mutex<Option<TimerDriverId>>,
  /// Consecutive non-lifetime relay failures (see
  /// [`HOST_TIMER_MAX_TRANSIENT_FAILURES`]); reset on success.
  transient_failures: std::sync::atomic::AtomicU32,
}

static JS_TIMER_HOSTS: std::sync::LazyLock<
  std::sync::Mutex<rustc_hash::FxHashMap<u64, std::sync::Weak<JsTimerHostInner>>>,
> = std::sync::LazyLock::new(|| std::sync::Mutex::new(rustc_hash::FxHashMap::default()));

fn registered_timer_host(id: u64) -> Option<std::sync::Arc<JsTimerHostInner>> {
  let mut registrations = JS_TIMER_HOSTS
    .lock()
    .unwrap_or_else(std::sync::PoisonError::into_inner);
  let inner = registrations.get(&id).and_then(std::sync::Weak::upgrade);
  if inner.is_none() {
    registrations.remove(&id);
  }
  inner
}

struct PendingHostTimer {
  cancellation: Option<futures::channel::oneshot::Sender<()>>,
  relay_health: std::sync::Arc<HostTimerRelayHealth>,
  relay_id: u32,
  waker: std::task::Waker,
  schedule_state: RelayScheduleState,
}

impl PendingHostTimer {
  fn signal_native_cancellation(&mut self) {
    if let Some(sender) = self.cancellation.take() {
      let _ = sender.send(());
    }
  }
}

#[derive(Clone, Copy, Debug, PartialEq, Eq)]
enum RelayScheduleState {
  AwaitingCallback,
  CallbackComplete,
}

trait PendingRelayState: Send + Sync + 'static {
  fn take_pending_relay(
    &self,
    id: TimerId,
    relay_id: u32,
    accounting: RelayCancellationAccounting,
  ) -> Option<PendingHostTimer>;
  /// Record that the JavaScript schedule callback has returned. `true` means
  /// Rust cleanup already removed this exact relay, so cancellation must be
  /// submitted now that it cannot overtake timer creation.
  fn mark_relay_callback_complete(&self, id: TimerId, relay_id: u32) -> bool;
  fn cancel_relay(
    self: &std::sync::Arc<Self>,
    relay_id: u32,
    relay_health: std::sync::Arc<HostTimerRelayHealth>,
  );
}

struct PendingRelayDropGuard<T: PendingRelayState> {
  state: std::sync::Arc<T>,
  id: TimerId,
  relay_id: u32,
  cleanup_on_drop: bool,
}

impl<T: PendingRelayState> PendingRelayDropGuard<T> {
  fn new(state: std::sync::Arc<T>, id: TimerId, relay_id: u32) -> Self {
    Self {
      state,
      id,
      relay_id,
      cleanup_on_drop: true,
    }
  }

  fn disarm(&mut self) {
    self.cleanup_on_drop = false;
  }
}

impl<T: PendingRelayState> Drop for PendingRelayDropGuard<T> {
  fn drop(&mut self) {
    if !self.cleanup_on_drop {
      return;
    }
    let Some(pending) = self.state.take_pending_relay(
      self.id,
      self.relay_id,
      RelayCancellationAccounting::HostHealth,
    ) else {
      return;
    };
    // An awaiting callback owns its exact, never-reused relay id and will
    // submit cancellation after JavaScript returns. Waking may re-enter timer
    // registration, but a later timer can never receive this id.
    retire_pending_relay(&self.state, pending);
  }
}

enum PendingHostTimerRegistration {
  Refreshed(std::task::Waker),
  Armed {
    relay_id: u32,
    cancellation: futures::channel::oneshot::Receiver<()>,
    relay_health: std::sync::Arc<HostTimerRelayHealth>,
  },
  Exhausted(std::task::Waker),
}

fn register_pending_host_timer_locked(
  pending: &mut rustc_hash::FxHashMap<TimerId, PendingHostTimer>,
  relay_ids: &RelayIdAllocator,
  id: TimerId,
  waker: std::task::Waker,
) -> PendingHostTimerRegistration {
  if let Some(slot) = pending.get_mut(&id) {
    return PendingHostTimerRegistration::Refreshed(std::mem::replace(&mut slot.waker, waker));
  }
  let Ok(relay_id) = relay_ids.reserve() else {
    return PendingHostTimerRegistration::Exhausted(waker);
  };
  let (cancellation_sender, cancellation) = futures::channel::oneshot::channel();
  let relay_health = std::sync::Arc::new(HostTimerRelayHealth::default());
  pending.insert(
    id,
    PendingHostTimer {
      cancellation: Some(cancellation_sender),
      relay_health: std::sync::Arc::clone(&relay_health),
      relay_id,
      waker,
      schedule_state: RelayScheduleState::AwaitingCallback,
    },
  );
  PendingHostTimerRegistration::Armed {
    relay_id,
    cancellation,
    relay_health,
  }
}

#[cfg(test)]
fn register_pending_host_timer(
  pending: &std::sync::Mutex<rustc_hash::FxHashMap<TimerId, PendingHostTimer>>,
  relay_ids: &RelayIdAllocator,
  id: TimerId,
  waker: std::task::Waker,
) -> PendingHostTimerRegistration {
  let mut pending = pending
    .lock()
    .unwrap_or_else(std::sync::PoisonError::into_inner);
  register_pending_host_timer_locked(&mut pending, relay_ids, id, waker)
}

fn register_pending_host_timer_if_live(
  pending: &std::sync::Mutex<rustc_hash::FxHashMap<TimerId, PendingHostTimer>>,
  relay_ids: &RelayIdAllocator,
  id: TimerId,
  waker: std::task::Waker,
  is_live: impl FnOnce() -> bool,
) -> Result<PendingHostTimerRegistration, std::task::Waker> {
  let mut pending = pending
    .lock()
    .unwrap_or_else(std::sync::PoisonError::into_inner);
  if !is_live() {
    return Err(waker);
  }
  Ok(register_pending_host_timer_locked(
    &mut pending,
    relay_ids,
    id,
    waker,
  ))
}

fn take_pending_host_timers(
  pending: &std::sync::Mutex<rustc_hash::FxHashMap<TimerId, PendingHostTimer>>,
  accounting: RelayCancellationAccounting,
) -> rustc_hash::FxHashMap<TimerId, PendingHostTimer> {
  let mut pending = pending
    .lock()
    .unwrap_or_else(std::sync::PoisonError::into_inner);
  for timer in pending.values() {
    timer.relay_health.set_cancellation_accounting(accounting);
  }
  std::mem::take(&mut *pending)
}

fn take_pending_host_timer(
  pending: &std::sync::Mutex<rustc_hash::FxHashMap<TimerId, PendingHostTimer>>,
  id: TimerId,
  relay_id: u32,
  accounting: RelayCancellationAccounting,
) -> Option<PendingHostTimer> {
  let mut pending = pending
    .lock()
    .unwrap_or_else(std::sync::PoisonError::into_inner);
  if pending
    .get(&id)
    .is_some_and(|slot| slot.relay_id == relay_id)
  {
    pending
      .get(&id)
      .expect("the matching pending relay must remain present")
      .relay_health
      .set_cancellation_accounting(accounting);
    pending.remove(&id)
  } else {
    None
  }
}

fn run_host_timer_cleanup_safely(cleanup: impl FnOnce()) {
  if let Err(payload) = std::panic::catch_unwind(std::panic::AssertUnwindSafe(cleanup))
    && let Err(nested_payload) =
      std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| drop(payload)))
  {
    // The original unwind is contained. A hostile payload destructor may
    // panic again; quarantine only that nested payload so its destructor
    // cannot unwind through the napi env cleanup hook either.
    std::mem::forget(nested_payload);
  }
}

fn retire_pending_relay<T: PendingRelayState>(
  state: &std::sync::Arc<T>,
  mut pending: PendingHostTimer,
) {
  // The detached relay must stop depending on JavaScript before cancellation
  // crosses the fallible TSFN boundary. A queue rejection or callback throw can
  // then affect diagnostics only, never relay retirement or runtime shutdown.
  pending.signal_native_cancellation();
  if pending.schedule_state == RelayScheduleState::CallbackComplete {
    let relay_health = std::sync::Arc::clone(&pending.relay_health);
    run_host_timer_cleanup_safely(|| state.cancel_relay(pending.relay_id, relay_health));
  }
  wake_host_timer_safely(pending.waker);
}

fn retire_pending_relays<T: PendingRelayState>(
  state: &std::sync::Arc<T>,
  pending: rustc_hash::FxHashMap<TimerId, PendingHostTimer>,
) {
  for (_, pending) in pending {
    retire_pending_relay(state, pending);
  }
}

fn complete_relay_schedule_callback<T: PendingRelayState, R>(
  state: &std::sync::Arc<T>,
  id: TimerId,
  relay_id: u32,
  relay_health: &std::sync::Arc<HostTimerRelayHealth>,
  result: R,
  deliver: impl FnOnce(R),
) {
  // This runs from napi-rs's return-value callback on the JavaScript call
  // stack, after the schedule function has returned but before delivery wakes
  // the Rust relay future. A terminal runnable drop therefore sees
  // CallbackComplete, while cleanup that won earlier is completed here.
  run_host_timer_cleanup_safely(|| {
    if state.mark_relay_callback_complete(id, relay_id) {
      state.cancel_relay(relay_id, std::sync::Arc::clone(relay_health));
    }
  });
  run_host_timer_cleanup_safely(|| deliver(result));
}

fn normalize_timer_schedule_result(
  result: napi::Result<napi::Either<Promise<()>, InvalidReturnValue>>,
) -> napi::Result<Promise<()>> {
  match result {
    Ok(napi::Either::A(promise)) => Ok(promise),
    Ok(napi::Either::B(invalid)) => Err(napi::Error::new(
      napi::Status::InvalidArg,
      format!(
        "The function returned `{}`, but expected `object`.",
        invalid.value_type.to_string().to_ascii_lowercase()
      ),
    )),
    Err(error) => Err(error),
  }
}

fn wake_host_timer_safely(waker: std::task::Waker) {
  // Host eviction runs from a napi env cleanup hook. A custom RawWaker must
  // not unwind through that FFI boundary or abort the remaining timer drain.
  // Borrowing for the wake keeps a panicking wake and a panicking destructor
  // in separate containment boundaries.
  run_host_timer_cleanup_safely(|| waker.wake_by_ref());
  run_host_timer_cleanup_safely(|| drop(waker));
}

fn drop_host_timer_waker_safely(waker: std::task::Waker) {
  run_host_timer_cleanup_safely(|| drop(waker));
}

fn recover_host_timer_failure(recover: impl FnOnce(), diagnostic: std::fmt::Arguments<'_>) {
  recover();
  run_host_timer_cleanup_safely(|| {
    use std::io::Write as _;
    let _ = writeln!(std::io::stderr().lock(), "{diagnostic}");
  });
}

impl JsTimerHostInner {
  fn lock_pending(
    &self,
  ) -> std::sync::MutexGuard<'_, rustc_hash::FxHashMap<TimerId, PendingHostTimer>> {
    self
      .pending
      .lock()
      .unwrap_or_else(std::sync::PoisonError::into_inner)
  }

  fn register_pending(
    &self,
    id: TimerId,
    waker: std::task::Waker,
  ) -> Result<PendingHostTimerRegistration, std::task::Waker> {
    // Eviction publishes `dead` before taking this same pending-map mutex to
    // drain. Registration therefore either enters the map before that drain,
    // or observes the dead/aborted host afterward and returns its waker intact.
    register_pending_host_timer_if_live(&self.pending, &self.relay_ids, id, waker, || {
      self.is_live()
    })
  }

  /// Can this host still deliver wakes? The `aborted` probe reads the
  /// threadsafe function's own closing flag, so a dying env is detected even
  /// before any eviction path ran.
  fn is_live(&self) -> bool {
    !self.dead.load(std::sync::atomic::Ordering::SeqCst)
      && !self.callback.aborted()
      && !self.cancel_callback.aborted()
  }

  fn handle_cancellation_failure(
    self: &std::sync::Arc<Self>,
    relay_id: u32,
    relay_health: &HostTimerRelayHealth,
    error: napi::Error,
    queue_failure: bool,
  ) {
    let operation = if queue_failure {
      "could not be queued"
    } else {
      "failed"
    };
    if relay_health.cancellation_accounting() == RelayCancellationAccounting::CleanupOnly {
      recover_host_timer_failure(
        || {},
        format_args!(
          "napi-async-runtime: host timer cancellation callback {operation} for relay {relay_id} during \
           cleanup: {error}"
        ),
      );
      return;
    }

    match record_host_timer_failure(
      &self.transient_failures,
      relay_health,
      error.status,
      self.is_live(),
    ) {
      HostTimerFailureAction::Duplicate => {
        recover_host_timer_failure(
          || {},
          format_args!(
            "napi-async-runtime: host timer cancellation callback {operation} for relay {relay_id} after \
             this relay failure was already accounted: {error}"
          ),
        );
      }
      HostTimerFailureAction::EvictHost => {
        recover_host_timer_failure(
          || self.evict(),
          format_args!(
            "napi-async-runtime: host timer cancellation callback {operation} for relay {relay_id} (host \
             gone, evicting): {error}"
          ),
        );
      }
      HostTimerFailureAction::EvictHostAfterStrikes(strikes) => {
        recover_host_timer_failure(
          || self.evict(),
          format_args!(
            "napi-async-runtime: host timer cancellation callback {operation} {strikes} times in a row, \
             evicting this timer host (relay {relay_id}): {error}"
          ),
        );
      }
      HostTimerFailureAction::Retry(strikes) => {
        recover_host_timer_failure(
          || {},
          format_args!(
            "napi-async-runtime: host timer cancellation callback {operation} for relay {relay_id} \
             ({strikes}/{HOST_TIMER_MAX_TRANSIENT_FAILURES} before eviction): {error}"
          ),
        );
      }
    }
  }

  fn cancel_relay(
    self: &std::sync::Arc<Self>,
    relay_id: u32,
    relay_health: std::sync::Arc<HostTimerRelayHealth>,
  ) {
    let callback_inner = std::sync::Arc::clone(self);
    let callback_health = std::sync::Arc::clone(&relay_health);
    let status = self.cancel_callback.call_with_return_value(
      FnArgs { data: (relay_id,) },
      ThreadsafeFunctionCallMode::NonBlocking,
      move |result, _| {
        let error = match result {
          Ok(napi::Either::A(())) => {
            if callback_health.cancellation_accounting() == RelayCancellationAccounting::HostHealth
            {
              reset_host_timer_failures_after_success(
                &callback_inner.transient_failures,
                &callback_health,
              );
            }
            return Ok(());
          }
          Ok(napi::Either::B(invalid)) => napi::Error::new(
            napi::Status::InvalidArg,
            format!(
              "The timer cancellation callback returned `{}`, but expected `undefined`.",
              invalid.value_type.to_string().to_ascii_lowercase()
            ),
          ),
          Err(error) => error,
        };
        callback_inner.handle_cancellation_failure(relay_id, &callback_health, error, false);
        Ok(())
      },
    );
    if status != napi::Status::Ok {
      let error = napi::Error::new(status, "Threadsafe timer cancellation callback call failed");
      self.handle_cancellation_failure(relay_id, &relay_health, error, true);
    }
  }

  fn mark_relay_callback_complete(&self, id: TimerId, relay_id: u32) -> bool {
    let mut pending = self.lock_pending();
    match pending.get_mut(&id) {
      Some(slot) if slot.relay_id == relay_id => {
        slot.schedule_state = RelayScheduleState::CallbackComplete;
        false
      }
      _ => true,
    }
  }

  fn take_pending_relay(
    &self,
    id: TimerId,
    relay_id: u32,
    accounting: RelayCancellationAccounting,
  ) -> Option<PendingHostTimer> {
    take_pending_host_timer(&self.pending, id, relay_id, accounting)
  }

  async fn invoke_schedule_callback(
    self: &std::sync::Arc<Self>,
    id: TimerId,
    relay_id: u32,
    ms: f64,
    relay_health: std::sync::Arc<HostTimerRelayHealth>,
  ) -> napi::Result<Promise<()>> {
    let (sender, receiver) = futures::channel::oneshot::channel();
    let callback_inner = std::sync::Arc::clone(self);
    let status = self.callback.call_with_return_value(
      FnArgs {
        data: (relay_id, ms),
      },
      ThreadsafeFunctionCallMode::NonBlocking,
      move |result, _| {
        complete_relay_schedule_callback(
          &callback_inner,
          id,
          relay_id,
          &relay_health,
          result,
          move |result| {
            let _ = sender.send(normalize_timer_schedule_result(result));
          },
        );
        Ok(())
      },
    );
    if status != napi::Status::Ok {
      return Err(napi::Error::new(
        status,
        "Threadsafe function call_async_catch failed",
      ));
    }
    receiver.await.map_err(|_| {
      napi::Error::new(
        napi::Status::GenericFailure,
        "Receive value from threadsafe function sender failed",
      )
    })?
  }

  /// Remove this host from timer duty: latch `dead`, drop the registry
  /// entry, and wake every sleep armed here so each re-polls onto the next
  /// live registrant (absolute deadlines preserve the remaining time; with no
  /// live registrant left the re-poll fails LOUD in the scheduler).
  /// Idempotent -- the cleanup hook, the `is_live` race path, and the
  /// relay-failure backstop may all reach it.
  fn evict(self: &std::sync::Arc<Self>) {
    self.dead.store(true, std::sync::atomic::Ordering::SeqCst);
    let registration = self
      .registration
      .lock()
      .unwrap_or_else(std::sync::PoisonError::into_inner)
      .take();
    JS_TIMER_HOSTS
      .lock()
      .unwrap_or_else(std::sync::PoisonError::into_inner)
      .remove(&self.host_registration);
    if let Some(id) = registration {
      unregister_timer_driver(id);
    }
    let pending = take_pending_host_timers(&self.pending, RelayCancellationAccounting::CleanupOnly);
    retire_pending_relays(self, pending);
  }
}

impl PendingRelayState for JsTimerHostInner {
  fn take_pending_relay(
    &self,
    id: TimerId,
    relay_id: u32,
    accounting: RelayCancellationAccounting,
  ) -> Option<PendingHostTimer> {
    JsTimerHostInner::take_pending_relay(self, id, relay_id, accounting)
  }

  fn mark_relay_callback_complete(&self, id: TimerId, relay_id: u32) -> bool {
    JsTimerHostInner::mark_relay_callback_complete(self, id, relay_id)
  }

  fn cancel_relay(
    self: &std::sync::Arc<Self>,
    relay_id: u32,
    relay_health: std::sync::Arc<HostTimerRelayHealth>,
  ) {
    JsTimerHostInner::cancel_relay(self, relay_id, relay_health);
  }
}

impl TimerDriver for JsTimerHost {
  fn register(&self, id: TimerId, deadline: std::time::Instant, waker: std::task::Waker) {
    if !self.inner.is_live() {
      // The selecting poll raced this host's death: make sure the eviction
      // bookkeeping ran, then wake immediately so the sleep re-selects from
      // the registry (which no longer offers this host).
      self.inner.evict();
      wake_host_timer_safely(waker);
      return;
    }
    let (relay_id, cancellation, relay_health) = match self.inner.register_pending(id, waker) {
      Err(waker) => {
        self.inner.evict();
        wake_host_timer_safely(waker);
        return;
      }
      Ok(PendingHostTimerRegistration::Refreshed(replaced_waker)) => {
        // A custom RawWaker destructor may re-enter timer code. `register_pending`
        // returns the old waker so its destructor runs after `pending` unlocks.
        drop_host_timer_waker_safely(replaced_waker);
        return;
      }
      Ok(PendingHostTimerRegistration::Armed {
        relay_id,
        cancellation,
        relay_health,
      }) => (relay_id, cancellation, relay_health),
      Ok(PendingHostTimerRegistration::Exhausted(waker)) => {
        recover_host_timer_failure(
          || {
            self.inner.evict();
            wake_host_timer_safely(waker);
          },
          format_args!(
            "napi-async-runtime: host timer relay id space exhausted, evicting this timer host"
          ),
        );
        return;
      }
    };
    let ms = deadline
      .saturating_duration_since(std::time::Instant::now())
      .as_secs_f64()
      * 1000.0;
    let inner = std::sync::Arc::clone(&self.inner);
    // Build the guard outside the async body so destroying a never-polled
    // relay runnable still retires its pending timer.
    let relay_drop_guard = PendingRelayDropGuard::new(std::sync::Arc::clone(&inner), id, relay_id);
    let submission = try_spawn_detached(async move {
      let mut relay_drop_guard = relay_drop_guard;
      let schedule_started = std::sync::atomic::AtomicBool::new(false);
      let schedule = async {
        schedule_started.store(true, std::sync::atomic::Ordering::Release);
        match inner
          .invoke_schedule_callback(id, relay_id, ms, std::sync::Arc::clone(&relay_health))
          .await
        {
          Ok(promise) => promise.await,
          Err(error) => Err(error),
        }
      };
      futures::pin_mut!(cancellation);
      futures::pin_mut!(schedule);
      let result = match futures::future::select(cancellation, schedule).await {
        futures::future::Either::Left((_cancelled, schedule)) => {
          // Native cancellation won the race. The pending entry is already
          // gone (`cancel` removes it before signalling this oneshot), so the
          // drop guard has nothing left to clean up either way.
          relay_drop_guard.disarm();
          if !schedule_started.load(std::sync::atomic::Ordering::Acquire) {
            // The schedule callback was never invoked (the cancellation was
            // already pending at this task's first poll), so there is no host
            // result to observe -- and invoking the callback now would arm a
            // JS timer that nothing will ever cancel.
            return;
          }
          // The schedule callback is in flight. Conforming hosts settle the
          // schedule promise on cancellation (see `cancelTimer` in
          // workerd-timer-host.ts and the node host in timer-host.ts), and
          // that settlement shares this relay's health record with the
          // cancellation callback: whichever failure reaches Rust first
          // consumes the relay's one strike and the other is reported as
          // diagnostic-only (`Duplicate`, "already accounted"). Dropping the
          // schedule future here instead would leave the schedule-side
          // result unobserved and the shared accounting blind.
          schedule.await
        }
        futures::future::Either::Right((result, _cancellation)) => result,
      };
      match result {
        Ok(()) => {
          reset_host_timer_failures_after_success(&inner.transient_failures, &relay_health);
          if let Some(pending) =
            inner.take_pending_relay(id, relay_id, RelayCancellationAccounting::CleanupOnly)
          {
            wake_host_timer_safely(pending.waker);
          }
          relay_drop_guard.disarm();
        }
        Err(error) => {
          let action = record_host_timer_failure(
            &inner.transient_failures,
            &relay_health,
            error.status,
            inner.is_live(),
          );
          match action {
            HostTimerFailureAction::EvictHost => {
              recover_host_timer_failure(
                || inner.evict(),
                format_args!(
                  "napi-async-runtime: host timer callback failed (host gone, evicting): {error}"
                ),
              );
            }
            HostTimerFailureAction::EvictHostAfterStrikes(strikes) => {
              recover_host_timer_failure(
                || inner.evict(),
                format_args!(
                  "napi-async-runtime: host timer callback failed {strikes} times in a row, evicting this \
                   timer host: {error}"
                ),
              );
            }
            HostTimerFailureAction::Retry(strikes) => {
              recover_host_timer_failure(
                || {
                  if let Some(pending) =
                    inner.take_pending_relay(id, relay_id, RelayCancellationAccounting::CleanupOnly)
                  {
                    // The callback may have synchronously armed a timeout before
                    // throwing, returning the wrong type, or producing a
                    // rejected Promise. Cleanup cancellation must not add a
                    // second strike for this same relay failure.
                    retire_pending_relay(&inner, pending);
                  }
                },
                format_args!(
                  "napi-async-runtime: host timer callback failed \
                   ({strikes}/{HOST_TIMER_MAX_TRANSIENT_FAILURES} before eviction): {error}"
                ),
              );
            }
            HostTimerFailureAction::Duplicate => {
              recover_host_timer_failure(
                || {
                  if let Some(pending) =
                    inner.take_pending_relay(id, relay_id, RelayCancellationAccounting::CleanupOnly)
                  {
                    retire_pending_relay(&inner, pending);
                  }
                },
                format_args!(
                  "napi-async-runtime: host timer callback failed after this relay failure was already \
                   accounted: {error}"
                ),
              );
            }
          }
          relay_drop_guard.disarm();
        }
      }
    });
    if let Err(rejected_relay) = submission {
      // The rejected future owns the exact-match guard. Dropping it removes
      // and wakes the pending timer before returning to the caller.
      drop(rejected_relay);
    }
  }

  fn cancel(&self, id: TimerId) {
    let pending = {
      let mut pending = self.inner.lock_pending();
      if let Some(timer) = pending.get(&id) {
        timer
          .relay_health
          .set_cancellation_accounting(RelayCancellationAccounting::HostHealth);
      }
      pending.remove(&id)
    };
    if let Some(mut pending) = pending {
      pending.signal_native_cancellation();
      if pending.schedule_state == RelayScheduleState::CallbackComplete {
        let relay_health = std::sync::Arc::clone(&pending.relay_health);
        self.inner.cancel_relay(pending.relay_id, relay_health);
      }
      drop_host_timer_waker_safely(pending.waker);
    }
  }

  fn is_live(&self) -> bool {
    self.inner.is_live()
  }

  fn on_swept(&self) {
    // The registry's selection sweep noticed this host's death (the
    // `aborted` probe can fire before the env-cleanup hook runs): run the
    // full eviction so every sleep pending here is woken into re-selection
    // instead of stranded. Idempotent with the hook and the relay backstop.
    self.inner.evict();
  }
}

#[napi(
  ts_args_type = "registrationHigh: number, registrationLow: number, schedule: (id: number, ms: number) => Promise<void>, cancel: (id: number) => void"
)]
/// Install the host timer callback backing the shared async runtime's
/// CurrentThread timers. Called at import by every
/// binding-loading JS entry with paired setTimeout/clearTimeout callbacks; each
/// importing env (main thread and workers alike) registers its own host, and
/// every live host receives each timer.
pub fn register_timer_host(
  env: &napi::Env,
  registration_high: u32,
  registration_low: u32,
  schedule: JsCallback<FnArgs<(u32, f64)>, Promise<()>>,
  cancel: JsCallback<FnArgs<(u32,)>, ()>,
) -> napi::Result<()> {
  let host_registration = claim_host_registration_id(registration_high, registration_low)?;
  let inner = std::sync::Arc::new(JsTimerHostInner {
    callback: schedule,
    cancel_callback: cancel,
    pending: std::sync::Mutex::default(),
    relay_ids: RelayIdAllocator::default(),
    dead: std::sync::atomic::AtomicBool::new(false),
    host_registration,
    registration: std::sync::Mutex::default(),
    transient_failures: std::sync::atomic::AtomicU32::new(0),
  });
  if !install_host_driver_registration(
    &inner.dead,
    &inner.registration,
    || {
      register_timer_driver(std::sync::Arc::new(JsTimerHost {
        inner: std::sync::Arc::clone(&inner),
      }))
    },
    || {
      JS_TIMER_HOSTS
        .lock()
        .unwrap_or_else(std::sync::PoisonError::into_inner)
        .insert(host_registration, std::sync::Arc::downgrade(&inner));
    },
    unregister_timer_driver,
  ) {
    return Err(napi::Error::new(
      napi::Status::GenericFailure,
      "The CurrentThread timer host was evicted during registration",
    ));
  }
  // Proactive eviction at env teardown (worker exit): the primary lifetime
  // mechanism; the `aborted` probe and the relay-failure path in the driver
  // are the backstops for anything the hook cannot reach in time.
  let hook_inner = std::sync::Arc::clone(&inner);
  install_cleanup_hook_or_rollback(
    || {
      env.add_env_cleanup_hook(hook_inner, |inner| {
        run_host_timer_cleanup_safely(|| inner.evict());
      })
    },
    || run_host_timer_cleanup_safely(|| inner.evict()),
  )?;
  Ok(())
}

#[napi]
/// Evict exactly one callback installed by `registerTimerHost`.
/// Pending sleeps are woken so they can reselect another live environment.
pub fn unregister_timer_host(registration_high: u32, registration_low: u32) {
  let id = BindingHostRegistration::id(registration_high, registration_low);
  release_host_registration_id(id);
  if let Some(inner) = registered_timer_host(id) {
    inner.evict();
  }
}

/// Configure the shared scheduler and register it as this addon's napi
/// [`AsyncRuntime`] backend.
///
/// Call this from the host's own `#[napi_derive::module_init]` hook, after
/// resolving the process configuration (environment variables, defaults) into
/// [`RuntimeOptions`]. The scheduler never reads the environment itself; the
/// host owns that resolution -- with ONE adapter-owned convenience: `install`
/// resolves the conventional [`DRAIN_LINGER_ENV`] override into
/// [`RuntimeOptions::drain_linger`] here, at the same addon-load boundary
/// where the host resolves its variables (a present numeric value wins over
/// the passed field; absent or non-numeric keeps it -- see
/// [`resolve_drain_linger`]). Hosts that call [`configure`] directly instead
/// of `install` own that resolution too. `configure` validates and normalizes
/// the options (clamping `drain_linger` to its ceiling), and the runtime
/// controller's options remain the reporting authority for
/// `getAsyncRuntimeConfig`.
pub fn install(mut options: RuntimeOptions) -> napi::Result<()> {
  if let Ok(raw) = std::env::var(crate::async_runtime::DRAIN_LINGER_ENV) {
    options.drain_linger = resolve_drain_linger(Some(raw), options.drain_linger);
  }
  configure(options).map_err(|error| napi::Error::from_reason(error.to_string()))?;
  register_async_runtime(SharedAsyncRuntime);
  Ok(())
}

/// Resolve an env-derived thread count: parse the raw value, treat a missing,
/// non-numeric, OR zero value as "unset" and fall back to `default`, then
/// clamp the result to `maximum`. A `0` must never survive here because the
/// host's module-init path feeds this count into a constructor that rejects
/// `0`: [`RuntimeOptions`] validation rejects a `0` thread count, which would
/// panic an `expect()` during addon load over what may be a simple typo in an
/// environment variable.
pub fn resolve_thread_count(raw: Option<String>, default: usize, maximum: usize) -> usize {
  assert!(default > 0, "the default thread count must be positive");
  assert!(maximum > 0, "the maximum thread count must be positive");
  raw
    .and_then(|value| value.parse::<usize>().ok())
    .filter(|&value| value != 0)
    .unwrap_or(default)
    .min(maximum)
}

/// Resolve an env-derived drainer idle-linger budget (a raw
/// [`DRAIN_LINGER_ENV`]-style value, MICROSECONDS) into
/// [`RuntimeOptions::drain_linger`]: a missing or non-numeric value keeps
/// `fallback` (the host-provided field, [`DEFAULT_DRAIN_LINGER_MICROS`] by
/// default); `0` disables lingering; anything else is clamped to
/// [`MAX_DRAIN_LINGER_MICROS`] so an extreme value (e.g. `u64::MAX`) tunes
/// the wave optimization instead of pinning pool workers on never-expiring
/// lingers. [`RuntimeOptions`] validation re-applies the same clamp, so a
/// host bypassing this resolver cannot smuggle an unbounded budget either.
pub fn resolve_drain_linger(
  raw: Option<String>,
  fallback: std::time::Duration,
) -> std::time::Duration {
  raw
    .and_then(|value| value.trim().parse::<u64>().ok())
    .map_or(fallback, std::time::Duration::from_micros)
    .min(std::time::Duration::from_micros(
      crate::async_runtime::MAX_DRAIN_LINGER_MICROS,
    ))
}

#[cfg(test)]
mod tests {
  use super::{
    BindingHostRegistration, BindingRuntimeFlavor, BindingRuntimeOptions,
    HOST_TIMER_MAX_TRANSIENT_FAILURES, HostTimerFailureAction, HostTimerRelayHealth,
    MAX_SAFE_JS_INTEGER, NATIVE_TASK_HOST_AFTER_DRIVE_TEST_HOOK,
    NATIVE_TASK_HOST_AFTER_PAYLOAD_DROP_TEST_HOOK, NativeCurrentThreadTaskHostInner,
    NativeCurrentThreadTaskHostPayload, PendingHostTimer, PendingHostTimerRegistration,
    PendingRelayDropGuard, PendingRelayState, RelayCancellationAccounting, RelayIdAllocator,
    RelayScheduleState, SharedAsyncRuntime, call_native_current_thread_task_host,
    claim_host_registration_id, complete_relay_schedule_callback,
    get_current_thread_task_host_contract_version, install_cleanup_hook_or_rollback,
    install_host_driver_registration, record_host_timer_failure, recover_host_timer_failure,
    register_pending_host_timer, register_pending_host_timer_if_live,
    reserve_current_thread_host_registration, reset_host_timer_failures_after_success,
    resolve_thread_count, retire_pending_relay, retire_pending_relays, safe_js_number,
    take_pending_host_timers, unregister_current_thread_task_host, wake_host_timer_safely,
  };

  #[derive(Default)]
  struct TestPendingRelayState {
    pending:
      std::sync::Mutex<rustc_hash::FxHashMap<crate::async_runtime::TimerId, PendingHostTimer>>,
    relay_ids: RelayIdAllocator,
    cancelled_relays: std::sync::Mutex<Vec<u32>>,
    panic_on_cancel: std::sync::atomic::AtomicBool,
  }

  impl PendingRelayState for TestPendingRelayState {
    fn take_pending_relay(
      &self,
      id: crate::async_runtime::TimerId,
      relay_id: u32,
      accounting: RelayCancellationAccounting,
    ) -> Option<PendingHostTimer> {
      let mut pending = self
        .pending
        .lock()
        .unwrap_or_else(std::sync::PoisonError::into_inner);
      if pending
        .get(&id)
        .is_some_and(|slot| slot.relay_id == relay_id)
      {
        pending
          .get(&id)
          .expect("the matching test relay must remain present")
          .relay_health
          .set_cancellation_accounting(accounting);
        pending.remove(&id)
      } else {
        None
      }
    }

    fn mark_relay_callback_complete(
      &self,
      id: crate::async_runtime::TimerId,
      relay_id: u32,
    ) -> bool {
      let mut pending = self
        .pending
        .lock()
        .unwrap_or_else(std::sync::PoisonError::into_inner);
      match pending.get_mut(&id) {
        Some(slot) if slot.relay_id == relay_id => {
          slot.schedule_state = RelayScheduleState::CallbackComplete;
          false
        }
        _ => true,
      }
    }

    fn cancel_relay(
      self: &std::sync::Arc<Self>,
      relay_id: u32,
      _relay_health: std::sync::Arc<HostTimerRelayHealth>,
    ) {
      self
        .cancelled_relays
        .lock()
        .unwrap_or_else(std::sync::PoisonError::into_inner)
        .push(relay_id);
      assert!(
        !self
          .panic_on_cancel
          .load(std::sync::atomic::Ordering::SeqCst),
        "intentional relay cancellation panic"
      );
    }
  }

  fn test_pending_relay_guard(
    state: &std::sync::Arc<TestPendingRelayState>,
    id: crate::async_runtime::TimerId,
    waker: std::task::Waker,
    schedule_state: RelayScheduleState,
  ) -> (
    u32,
    std::sync::Arc<HostTimerRelayHealth>,
    PendingRelayDropGuard<TestPendingRelayState>,
  ) {
    let relay_id = state
      .relay_ids
      .reserve()
      .expect("the test relay id must be available");
    let relay_health = std::sync::Arc::new(HostTimerRelayHealth::default());
    let replaced = state
      .pending
      .lock()
      .unwrap_or_else(std::sync::PoisonError::into_inner)
      .insert(
        id,
        PendingHostTimer {
          cancellation: None,
          relay_health: std::sync::Arc::clone(&relay_health),
          relay_id,
          waker,
          schedule_state,
        },
      );
    assert!(replaced.is_none(), "the test timer id must be vacant");
    (
      relay_id,
      relay_health,
      PendingRelayDropGuard::new(std::sync::Arc::clone(state), id, relay_id),
    )
  }

  struct BudgetBoundaryTimerWait {
    state: std::sync::Arc<TestPendingRelayState>,
    timer_id: crate::async_runtime::TimerId,
    schedule_state: RelayScheduleState,
    registered: bool,
  }

  impl std::future::Future for BudgetBoundaryTimerWait {
    type Output = ();

    fn poll(
      mut self: std::pin::Pin<&mut Self>,
      cx: &mut std::task::Context<'_>,
    ) -> std::task::Poll<Self::Output> {
      assert!(
        !self.registered,
        "the timer wait must be cancelled before a second poll"
      );
      self.registered = true;
      let state = std::sync::Arc::clone(&self.state);
      let timer_id = self.timer_id;
      let (_, _, relay_drop_guard) =
        test_pending_relay_guard(&state, timer_id, cx.waker().clone(), self.schedule_state);
      crate::async_runtime::spawn_detached(async move {
        let mut relay_drop_guard = relay_drop_guard;
        std::future::pending::<()>().await;
        relay_drop_guard.disarm();
      });
      std::task::Poll::Pending
    }
  }

  #[test]
  fn current_thread_task_host_contract_version_is_stable() {
    assert_eq!(get_current_thread_task_host_contract_version(), 4);
  }

  #[test]
  fn host_registration_reservations_are_exact_and_single_use() {
    let claimed = reserve_current_thread_host_registration().unwrap();
    let claimed_id = BindingHostRegistration::id(claimed.high, claimed.low);
    assert_eq!(
      claim_host_registration_id(claimed.high, claimed.low).unwrap(),
      claimed_id,
      "the exact reserved capability must be claimable once"
    );
    assert!(
      claim_host_registration_id(claimed.high, claimed.low).is_err(),
      "a consumed registration capability must not be reusable"
    );

    let released = reserve_current_thread_host_registration().unwrap();
    unregister_current_thread_task_host(released.high, released.low);
    assert!(
      claim_host_registration_id(released.high, released.low).is_err(),
      "unregister must release a reservation before installation"
    );
  }

  fn native_task_host_with_raw_owner(raw: usize) -> NativeCurrentThreadTaskHostInner {
    NativeCurrentThreadTaskHostInner {
      threadsafe_function: std::sync::Mutex::new(Some(raw)),
      dead: std::sync::atomic::AtomicBool::new(false),
      environment_closing: std::sync::atomic::AtomicBool::new(false),
      host_registration: u64::MAX,
      registration: std::sync::Mutex::default(),
    }
  }

  #[test]
  fn native_task_host_competing_release_paths_retire_the_raw_owner_once() {
    let inner = std::sync::Arc::new(native_task_host_with_raw_owner(1));
    let releases = std::sync::Arc::new(std::sync::atomic::AtomicUsize::new(0));
    let barrier = std::sync::Arc::new(std::sync::Barrier::new(3));
    let mut threads = Vec::new();

    for mode in [
      napi::sys::ThreadsafeFunctionReleaseMode::release,
      napi::sys::ThreadsafeFunctionReleaseMode::abort,
    ] {
      let inner = std::sync::Arc::clone(&inner);
      let releases = std::sync::Arc::clone(&releases);
      let barrier = std::sync::Arc::clone(&barrier);
      threads.push(std::thread::spawn(move || {
        barrier.wait();
        inner.release_threadsafe_function_with(mode, |threadsafe_function, _| {
          assert_eq!(threadsafe_function as usize, 1);
          releases.fetch_add(1, std::sync::atomic::Ordering::SeqCst);
          napi::sys::Status::napi_ok
        });
      }));
    }

    barrier.wait();
    for thread in threads {
      thread.join().unwrap();
    }
    assert_eq!(releases.load(std::sync::atomic::Ordering::SeqCst), 1);
    assert!(
      inner
        .threadsafe_function
        .lock()
        .unwrap_or_else(std::sync::PoisonError::into_inner)
        .is_none()
    );
  }

  #[test]
  fn native_task_host_closing_call_retires_owner_without_a_second_release() {
    let inner = native_task_host_with_raw_owner(1);
    let status = inner
      .call_threadsafe_function_with(std::ptr::null_mut(), |threadsafe_function, _| {
        assert_eq!(threadsafe_function as usize, 1);
        napi::sys::Status::napi_closing
      })
      .expect("the fake TSFN owner must be callable");

    assert_eq!(status, napi::sys::Status::napi_closing);
    assert!(
      inner
        .environment_closing
        .load(std::sync::atomic::Ordering::SeqCst)
    );
    let releases = std::sync::atomic::AtomicUsize::new(0);
    inner.release_threadsafe_function_with(
      napi::sys::ThreadsafeFunctionReleaseMode::release,
      |_, _| {
        releases.fetch_add(1, std::sync::atomic::Ordering::SeqCst);
        napi::sys::Status::napi_ok
      },
    );
    assert_eq!(
      releases.load(std::sync::atomic::Ordering::SeqCst),
      0,
      "napi_closing already retired the initial acquisition and invalidated the pointer"
    );
  }

  #[test]
  fn host_registration_capability_round_trips_without_precision_loss() {
    for id in [1, u64::from(u32::MAX), u64::from(u32::MAX) + 1, u64::MAX] {
      let registration = BindingHostRegistration::from_id(id);
      assert_eq!(
        BindingHostRegistration::id(registration.high, registration.low),
        id
      );
    }
  }

  #[test]
  fn binding_runtime_options_convert_to_a_partial_core_patch() {
    use crate::async_runtime::{RuntimeFlavor, RuntimeOptionsPatch};

    let patch = RuntimeOptionsPatch::try_from(BindingRuntimeOptions {
      flavor: Some(BindingRuntimeFlavor::CurrentThread),
      worker_threads: None,
      max_blocking_tasks: Some(7.0),
    })
    .expect("valid binding options must convert");

    assert!(matches!(patch.flavor, Some(RuntimeFlavor::CurrentThread)));
    assert_eq!(patch.worker_threads, None);
    assert_eq!(patch.max_blocking_tasks, Some(7));
  }

  #[test]
  fn binding_runtime_options_reject_unsafe_thread_counts() {
    use crate::async_runtime::RuntimeOptionsPatch;

    for value in [
      -1.0,
      0.0,
      1.5,
      f64::NAN,
      f64::INFINITY,
      257.0,
      f64::from(u32::MAX) + 1.0,
    ] {
      let error = RuntimeOptionsPatch::try_from(BindingRuntimeOptions {
        flavor: None,
        worker_threads: Some(value),
        max_blocking_tasks: None,
      })
      .expect_err("invalid worker thread counts must be rejected");
      assert!(
        error
          .reason
          .contains("`workerThreads` must be a positive integer"),
        "unexpected validation error: {}",
        error.reason
      );
    }

    for value in [-1.0, 257.0] {
      let error = RuntimeOptionsPatch::try_from(BindingRuntimeOptions {
        flavor: None,
        worker_threads: None,
        max_blocking_tasks: Some(value),
      })
      .expect_err("invalid blocking task counts must be rejected");
      assert!(
        error
          .reason
          .contains("`maxBlockingTasks` must be a positive integer"),
        "unexpected validation error: {}",
        error.reason
      );
    }
  }

  #[test]
  fn runtime_metrics_preserve_safe_javascript_integer_range() {
    const BEYOND_U32_JS_NUMBER: f64 = 4_294_967_296.0;
    const MAX_SAFE_JS_NUMBER: f64 = 9_007_199_254_740_991.0;

    let beyond_u32 = u64::from(u32::MAX) + 1;
    assert_eq!(
      safe_js_number(beyond_u32).to_bits(),
      BEYOND_U32_JS_NUMBER.to_bits()
    );
    assert_eq!(
      safe_js_number(MAX_SAFE_JS_INTEGER).to_bits(),
      MAX_SAFE_JS_NUMBER.to_bits()
    );
    assert_eq!(
      safe_js_number(MAX_SAFE_JS_INTEGER + 1).to_bits(),
      MAX_SAFE_JS_NUMBER.to_bits(),
      "metrics must remain exact JavaScript numbers instead of silently losing integer precision"
    );
  }

  #[test]
  fn relay_ids_fail_closed_without_reuse_at_u32_exhaustion() {
    let allocator = RelayIdAllocator::default();
    allocator
      .next
      .store(u64::from(u32::MAX), std::sync::atomic::Ordering::Relaxed);

    assert_eq!(allocator.reserve(), Ok(u32::MAX));
    assert!(
      allocator.reserve().is_err(),
      "the allocator must fail instead of wrapping to a reusable id"
    );
    assert!(
      allocator.reserve().is_err(),
      "exhaustion must remain stable instead of spinning or advancing toward wraparound"
    );

    let pending = std::sync::Mutex::new(rustc_hash::FxHashMap::default());
    let waker = futures::task::noop_waker();
    let exhausted_waker = match register_pending_host_timer(&pending, &allocator, 1, waker.clone())
    {
      PendingHostTimerRegistration::Exhausted(waker) => waker,
      PendingHostTimerRegistration::Refreshed(_) | PendingHostTimerRegistration::Armed { .. } => {
        panic!("an exhausted allocator must reject a new pending timer")
      }
    };
    assert!(
      exhausted_waker.will_wake(&waker),
      "allocator failure must return the attempted timer's waker intact"
    );
    assert!(
      pending
        .lock()
        .unwrap_or_else(std::sync::PoisonError::into_inner)
        .is_empty(),
      "an exhausted allocator must not publish a pending timer without a relay capability"
    );
  }

  #[test]
  fn host_timer_wake_and_drop_panics_are_contained_separately() {
    struct PanickingWakeAndDrop;

    impl std::task::Wake for PanickingWakeAndDrop {
      fn wake(self: std::sync::Arc<Self>) {
        panic!("intentional host timer waker panic");
      }

      fn wake_by_ref(self: &std::sync::Arc<Self>) {
        panic!("intentional host timer borrowed waker panic");
      }
    }

    impl Drop for PanickingWakeAndDrop {
      fn drop(&mut self) {
        panic!("intentional host timer waker destructor panic");
      }
    }

    let waker = std::task::Waker::from(std::sync::Arc::new(PanickingWakeAndDrop));
    wake_host_timer_safely(waker);
  }

  #[test]
  fn relay_drop_guard_cleans_only_its_matching_pending_relay() {
    struct CountingWake(std::sync::Arc<std::sync::atomic::AtomicUsize>);

    impl std::task::Wake for CountingWake {
      fn wake(self: std::sync::Arc<Self>) {
        self.0.fetch_add(1, std::sync::atomic::Ordering::SeqCst);
      }
    }

    let state = std::sync::Arc::new(TestPendingRelayState::default());
    let wakes = std::sync::Arc::new(std::sync::atomic::AtomicUsize::new(0));
    let (relay_id, _, guard) = test_pending_relay_guard(
      &state,
      1,
      std::task::Waker::from(std::sync::Arc::new(CountingWake(std::sync::Arc::clone(
        &wakes,
      )))),
      RelayScheduleState::CallbackComplete,
    );
    drop(guard);

    assert!(
      state
        .pending
        .lock()
        .unwrap_or_else(std::sync::PoisonError::into_inner)
        .is_empty()
    );
    assert_eq!(
      *state
        .cancelled_relays
        .lock()
        .unwrap_or_else(std::sync::PoisonError::into_inner),
      [relay_id],
      "an armed relay must be cancelled before its waiter is woken"
    );
    assert_eq!(
      wakes.load(std::sync::atomic::Ordering::SeqCst),
      1,
      "terminal relay cleanup must wake the pending timer"
    );

    let stale_wakes = std::sync::Arc::new(std::sync::atomic::AtomicUsize::new(0));
    let (stale_relay_id, _, stale_guard) = test_pending_relay_guard(
      &state,
      2,
      std::task::Waker::from(std::sync::Arc::new(CountingWake(std::sync::Arc::clone(
        &stale_wakes,
      )))),
      RelayScheduleState::AwaitingCallback,
    );
    let replacement_relay_id = state
      .relay_ids
      .reserve()
      .expect("the replacement relay id must be available");
    assert_ne!(
      replacement_relay_id, stale_relay_id,
      "relay ids must never be reused within one host"
    );
    let replaced = state
      .pending
      .lock()
      .unwrap_or_else(std::sync::PoisonError::into_inner)
      .insert(
        2,
        PendingHostTimer {
          cancellation: None,
          relay_health: std::sync::Arc::new(HostTimerRelayHealth::default()),
          relay_id: replacement_relay_id,
          waker: futures::task::noop_waker(),
          schedule_state: RelayScheduleState::CallbackComplete,
        },
      )
      .expect("the original relay must still be pending");
    drop(replaced);
    drop(stale_guard);

    let replacement = state
      .pending
      .lock()
      .unwrap_or_else(std::sync::PoisonError::into_inner)
      .remove(&2)
      .expect("a stale guard must preserve the replacement relay");
    assert_eq!(replacement.relay_id, replacement_relay_id);
    assert_eq!(stale_wakes.load(std::sync::atomic::Ordering::SeqCst), 0);
    assert_eq!(
      *state
        .cancelled_relays
        .lock()
        .unwrap_or_else(std::sync::PoisonError::into_inner),
      [relay_id],
      "a stale guard must not cancel a replacement relay"
    );
    drop(replacement);

    let disarmed_wakes = std::sync::Arc::new(std::sync::atomic::AtomicUsize::new(0));
    let (disarmed_relay_id, _, mut disarmed_guard) = test_pending_relay_guard(
      &state,
      3,
      std::task::Waker::from(std::sync::Arc::new(CountingWake(std::sync::Arc::clone(
        &disarmed_wakes,
      )))),
      RelayScheduleState::CallbackComplete,
    );
    disarmed_guard.disarm();
    drop(disarmed_guard);
    let disarmed_pending = state
      .pending
      .lock()
      .unwrap_or_else(std::sync::PoisonError::into_inner)
      .remove(&3)
      .expect("normal relay completion must retain ownership of pending cleanup");
    assert_eq!(disarmed_pending.relay_id, disarmed_relay_id);
    assert_eq!(disarmed_wakes.load(std::sync::atomic::Ordering::SeqCst), 0);
    assert_eq!(
      *state
        .cancelled_relays
        .lock()
        .unwrap_or_else(std::sync::PoisonError::into_inner),
      [relay_id],
      "a disarmed guard must not change normal relay cleanup semantics"
    );
    drop(disarmed_pending);
  }

  #[test]
  fn relay_callback_completion_precedes_delivery_and_closes_terminal_drop_window() {
    struct CountingWake(std::sync::Arc<std::sync::atomic::AtomicUsize>);

    impl std::task::Wake for CountingWake {
      fn wake(self: std::sync::Arc<Self>) {
        self.0.fetch_add(1, std::sync::atomic::Ordering::SeqCst);
      }
    }

    let state = std::sync::Arc::new(TestPendingRelayState::default());
    let wakes = std::sync::Arc::new(std::sync::atomic::AtomicUsize::new(0));
    let (relay_id, relay_health, guard) = test_pending_relay_guard(
      &state,
      1,
      std::task::Waker::from(std::sync::Arc::new(CountingWake(std::sync::Arc::clone(
        &wakes,
      )))),
      RelayScheduleState::AwaitingCallback,
    );

    // Model napi-rs after the JavaScript schedule function returned and armed
    // its timeout, but before result delivery can re-poll the Rust relay.
    complete_relay_schedule_callback(&state, 1, relay_id, &relay_health, guard, drop);

    assert!(
      state
        .pending
        .lock()
        .unwrap_or_else(std::sync::PoisonError::into_inner)
        .is_empty()
    );
    assert_eq!(
      *state
        .cancelled_relays
        .lock()
        .unwrap_or_else(std::sync::PoisonError::into_inner),
      [relay_id],
      "terminal drop during return-value delivery must cancel the synchronously armed timeout"
    );
    assert_eq!(wakes.load(std::sync::atomic::Ordering::SeqCst), 1);
  }

  #[test]
  fn relay_callback_completion_finishes_cleanup_that_won_before_js_returned() {
    struct CountingWake(std::sync::Arc<std::sync::atomic::AtomicUsize>);

    impl std::task::Wake for CountingWake {
      fn wake(self: std::sync::Arc<Self>) {
        self.0.fetch_add(1, std::sync::atomic::Ordering::SeqCst);
      }
    }

    let state = std::sync::Arc::new(TestPendingRelayState::default());
    let wakes = std::sync::Arc::new(std::sync::atomic::AtomicUsize::new(0));
    let (relay_id, relay_health, guard) = test_pending_relay_guard(
      &state,
      1,
      std::task::Waker::from(std::sync::Arc::new(CountingWake(std::sync::Arc::clone(
        &wakes,
      )))),
      RelayScheduleState::AwaitingCallback,
    );

    drop(guard);
    assert!(
      state
        .cancelled_relays
        .lock()
        .unwrap_or_else(std::sync::PoisonError::into_inner)
        .is_empty(),
      "cancellation must not overtake a schedule callback that has not returned"
    );

    complete_relay_schedule_callback(&state, 1, relay_id, &relay_health, (), drop);
    assert_eq!(
      *state
        .cancelled_relays
        .lock()
        .unwrap_or_else(std::sync::PoisonError::into_inner),
      [relay_id],
      "the callback return handshake must finish exact cancellation after timer creation"
    );
    assert_eq!(wakes.load(std::sync::atomic::Ordering::SeqCst), 1);
  }

  #[test]
  fn relay_terminal_cleanup_contains_cancel_panics_and_still_wakes() {
    struct CountingWake(std::sync::Arc<std::sync::atomic::AtomicUsize>);

    impl std::task::Wake for CountingWake {
      fn wake(self: std::sync::Arc<Self>) {
        self.0.fetch_add(1, std::sync::atomic::Ordering::SeqCst);
      }
    }

    let state = std::sync::Arc::new(TestPendingRelayState::default());
    state
      .panic_on_cancel
      .store(true, std::sync::atomic::Ordering::SeqCst);
    let wakes = std::sync::Arc::new(std::sync::atomic::AtomicUsize::new(0));
    let (relay_id, _, guard) = test_pending_relay_guard(
      &state,
      1,
      std::task::Waker::from(std::sync::Arc::new(CountingWake(std::sync::Arc::clone(
        &wakes,
      )))),
      RelayScheduleState::CallbackComplete,
    );

    drop(guard);

    assert_eq!(
      *state
        .cancelled_relays
        .lock()
        .unwrap_or_else(std::sync::PoisonError::into_inner),
      [relay_id]
    );
    assert_eq!(
      wakes.load(std::sync::atomic::Ordering::SeqCst),
      1,
      "a panicking cancellation path must not suppress the pending timer wake"
    );
  }

  #[test]
  fn relay_native_cancellation_settles_before_fallible_host_cancellation() {
    use futures::FutureExt as _;

    let state = std::sync::Arc::new(TestPendingRelayState::default());
    state
      .panic_on_cancel
      .store(true, std::sync::atomic::Ordering::SeqCst);
    let registration = register_pending_host_timer(
      &state.pending,
      &state.relay_ids,
      1,
      futures::task::noop_waker(),
    );
    let cancellation = match registration {
      PendingHostTimerRegistration::Armed { cancellation, .. } => cancellation,
      PendingHostTimerRegistration::Refreshed(_) | PendingHostTimerRegistration::Exhausted(_) => {
        panic!("the first timer registration must reserve a native cancellation relay")
      }
    };
    let mut pending = state
      .pending
      .lock()
      .unwrap_or_else(std::sync::PoisonError::into_inner)
      .remove(&1)
      .expect("the registered timer must remain pending");
    pending.schedule_state = RelayScheduleState::CallbackComplete;

    retire_pending_relay(&state, pending);

    assert!(
      matches!(cancellation.now_or_never(), Some(Ok(()))),
      "native relay retirement must publish before the contained host cancellation panic"
    );
  }

  #[test]
  fn relay_bulk_eviction_contains_each_cancel_panic_and_wakes_every_timer() {
    struct CountingWake(std::sync::Arc<std::sync::atomic::AtomicUsize>);

    impl std::task::Wake for CountingWake {
      fn wake(self: std::sync::Arc<Self>) {
        self.0.fetch_add(1, std::sync::atomic::Ordering::SeqCst);
      }
    }

    let state = std::sync::Arc::new(TestPendingRelayState::default());
    state
      .panic_on_cancel
      .store(true, std::sync::atomic::Ordering::SeqCst);
    let wakes = std::sync::Arc::new(std::sync::atomic::AtomicUsize::new(0));
    let pending = rustc_hash::FxHashMap::from_iter([7, 8].map(|relay_id| {
      (
        u64::from(relay_id),
        PendingHostTimer {
          cancellation: None,
          relay_health: std::sync::Arc::new(HostTimerRelayHealth::default()),
          relay_id,
          waker: std::task::Waker::from(std::sync::Arc::new(CountingWake(std::sync::Arc::clone(
            &wakes,
          )))),
          schedule_state: RelayScheduleState::CallbackComplete,
        },
      )
    }));

    retire_pending_relays(&state, pending);

    let mut cancelled = state
      .cancelled_relays
      .lock()
      .unwrap_or_else(std::sync::PoisonError::into_inner)
      .clone();
    cancelled.sort_unstable();
    assert_eq!(cancelled, [7, 8]);
    assert_eq!(
      wakes.load(std::sync::atomic::Ordering::SeqCst),
      2,
      "one panicking cancellation must not suppress any later pending timer wake"
    );
  }

  #[test]
  fn failed_relay_retirement_cancels_only_after_the_schedule_callback_returns() {
    struct CountingWake(std::sync::Arc<std::sync::atomic::AtomicUsize>);

    impl std::task::Wake for CountingWake {
      fn wake(self: std::sync::Arc<Self>) {
        self.0.fetch_add(1, std::sync::atomic::Ordering::SeqCst);
      }
    }

    let state = std::sync::Arc::new(TestPendingRelayState::default());
    let wakes = std::sync::Arc::new(std::sync::atomic::AtomicUsize::new(0));
    retire_pending_relay(
      &state,
      PendingHostTimer {
        cancellation: None,
        relay_health: std::sync::Arc::new(HostTimerRelayHealth::default()),
        relay_id: 7,
        waker: std::task::Waker::from(std::sync::Arc::new(CountingWake(std::sync::Arc::clone(
          &wakes,
        )))),
        schedule_state: RelayScheduleState::CallbackComplete,
      },
    );
    retire_pending_relay(
      &state,
      PendingHostTimer {
        cancellation: None,
        relay_health: std::sync::Arc::new(HostTimerRelayHealth::default()),
        relay_id: 8,
        waker: std::task::Waker::from(std::sync::Arc::new(CountingWake(std::sync::Arc::clone(
          &wakes,
        )))),
        schedule_state: RelayScheduleState::AwaitingCallback,
      },
    );

    assert_eq!(
      *state
        .cancelled_relays
        .lock()
        .unwrap_or_else(std::sync::PoisonError::into_inner),
      [7],
      "a callback-complete failure may have armed a timeout, while a failed TSFN submission did not"
    );
    assert_eq!(
      wakes.load(std::sync::atomic::Ordering::SeqCst),
      2,
      "both failed relays must wake their timer operations"
    );
  }

  #[test]
  fn host_timer_cleanup_contains_panicking_payload_drop_across_ffi() {
    const CHILD_ENV: &str = "NAPI_RUNTIME_TEST_HOST_TIMER_PANIC_PAYLOAD_CHILD";

    if std::env::var_os(CHILD_ENV).is_some() {
      struct PanicOnDrop;

      impl Drop for PanicOnDrop {
        fn drop(&mut self) {
          panic!("intentional host timer panic payload destructor panic");
        }
      }

      struct PanickingWake;

      impl std::task::Wake for PanickingWake {
        fn wake(self: std::sync::Arc<Self>) {
          std::panic::panic_any(PanicOnDrop);
        }
      }

      extern "C" fn cleanup_shim() {
        let waker = std::task::Waker::from(std::sync::Arc::new(PanickingWake));
        wake_host_timer_safely(waker);
      }

      cleanup_shim();
      return;
    }

    let output = std::process::Command::new(std::env::current_exe().unwrap())
      .arg("--exact")
      .arg("adapter::tests::host_timer_cleanup_contains_panicking_payload_drop_across_ffi")
      .arg("--nocapture")
      .env(CHILD_ENV, "1")
      .output()
      .expect("the host timer cleanup subprocess must start");
    assert!(
      output.status.success(),
      "host timer cleanup must contain panic-payload destruction across the C ABI; status={:?}\nstdout={}\nstderr={}",
      output.status.code(),
      String::from_utf8_lossy(&output.stdout),
      String::from_utf8_lossy(&output.stderr)
    );
  }

  #[test]
  fn host_timer_failure_recovery_precedes_and_survives_diagnostic_panics() {
    struct PanickingDiagnostic;

    impl std::fmt::Display for PanickingDiagnostic {
      fn fmt(&self, _formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        panic!("intentional host timer diagnostic panic");
      }
    }

    let recovered = std::sync::atomic::AtomicBool::new(false);
    let diagnostic = PanickingDiagnostic;
    recover_host_timer_failure(
      || recovered.store(true, std::sync::atomic::Ordering::SeqCst),
      format_args!("{diagnostic}"),
    );
    assert!(
      recovered.load(std::sync::atomic::Ordering::SeqCst),
      "timer recovery must complete before fallible diagnostics run"
    );
  }

  #[test]
  fn host_timer_registry_drops_wakers_after_unlock() {
    struct LockCheckingDrop {
      pending: std::sync::Weak<
        std::sync::Mutex<rustc_hash::FxHashMap<crate::async_runtime::TimerId, PendingHostTimer>>,
      >,
      dropped: std::sync::Arc<std::sync::atomic::AtomicBool>,
    }

    #[expect(
      clippy::manual_noop_waker,
      reason = "the Arc-backed no-op waker exists to exercise its re-entrant destructor"
    )]
    impl std::task::Wake for LockCheckingDrop {
      fn wake(self: std::sync::Arc<Self>) {}
    }

    impl Drop for LockCheckingDrop {
      fn drop(&mut self) {
        let pending = self
          .pending
          .upgrade()
          .expect("the pending registry must outlive its waker");
        let _guard = pending.try_lock().unwrap_or_else(|_| {
          panic!("a host-timer waker was dropped while the pending registry was locked")
        });
        self
          .dropped
          .store(true, std::sync::atomic::Ordering::SeqCst);
      }
    }

    let pending = std::sync::Arc::new(std::sync::Mutex::new(rustc_hash::FxHashMap::default()));
    let relay_ids = RelayIdAllocator::default();
    let dropped = std::sync::Arc::new(std::sync::atomic::AtomicBool::new(false));
    let old_waker = std::task::Waker::from(std::sync::Arc::new(LockCheckingDrop {
      pending: std::sync::Arc::downgrade(&pending),
      dropped: std::sync::Arc::clone(&dropped),
    }));
    assert!(matches!(
      register_pending_host_timer(&pending, &relay_ids, 1, old_waker),
      PendingHostTimerRegistration::Armed { .. }
    ));

    let replaced_waker =
      match register_pending_host_timer(&pending, &relay_ids, 1, futures::task::noop_waker()) {
        PendingHostTimerRegistration::Refreshed(waker) => waker,
        PendingHostTimerRegistration::Armed { .. } => {
          panic!("a re-poll must refresh the existing timer")
        }
        PendingHostTimerRegistration::Exhausted(_) => {
          panic!("the test allocator must not be exhausted")
        }
      };
    assert!(
      !dropped.load(std::sync::atomic::Ordering::SeqCst),
      "the helper must return ownership instead of dropping under its lock"
    );
    drop(replaced_waker);
    assert!(dropped.load(std::sync::atomic::Ordering::SeqCst));

    let evicted_waker_dropped = std::sync::Arc::new(std::sync::atomic::AtomicBool::new(false));
    let evicted_waker = std::task::Waker::from(std::sync::Arc::new(LockCheckingDrop {
      pending: std::sync::Arc::downgrade(&pending),
      dropped: std::sync::Arc::clone(&evicted_waker_dropped),
    }));
    assert!(matches!(
      register_pending_host_timer(&pending, &relay_ids, 2, evicted_waker),
      PendingHostTimerRegistration::Armed { .. }
    ));
    let evicted = take_pending_host_timers(&pending, RelayCancellationAccounting::CleanupOnly);
    assert!(
      !evicted_waker_dropped.load(std::sync::atomic::Ordering::SeqCst),
      "bulk eviction must move wakers out instead of dropping under its lock"
    );
    drop(evicted);
    assert!(evicted_waker_dropped.load(std::sync::atomic::Ordering::SeqCst));
  }

  #[test]
  fn host_timer_eviction_prevents_pending_registration_after_its_drain() {
    let pending = std::sync::Arc::new(std::sync::Mutex::new(rustc_hash::FxHashMap::default()));
    let relay_ids = std::sync::Arc::new(RelayIdAllocator::default());
    let dead = std::sync::Arc::new(std::sync::atomic::AtomicBool::new(false));
    let mut pending_guard = pending
      .lock()
      .unwrap_or_else(std::sync::PoisonError::into_inner);
    let (started_tx, started_rx) = std::sync::mpsc::sync_channel(1);
    let registration_pending = std::sync::Arc::clone(&pending);
    let registration_relay_ids = std::sync::Arc::clone(&relay_ids);
    let registration_dead = std::sync::Arc::clone(&dead);
    let registration = std::thread::spawn(move || {
      started_tx.send(()).unwrap();
      register_pending_host_timer_if_live(
        &registration_pending,
        &registration_relay_ids,
        1,
        futures::task::noop_waker(),
        || !registration_dead.load(std::sync::atomic::Ordering::SeqCst),
      )
    });

    started_rx.recv().unwrap();
    dead.store(true, std::sync::atomic::Ordering::SeqCst);
    let drained = std::mem::take(&mut *pending_guard);
    drop(pending_guard);
    drop(drained);

    assert!(
      registration.join().unwrap().is_err(),
      "registration arriving after death publication and the pending drain must be rejected"
    );
    assert!(
      pending
        .lock()
        .unwrap_or_else(std::sync::PoisonError::into_inner)
        .is_empty(),
      "a stale selected driver must not publish a new pending arm after eviction drained the map"
    );
  }

  #[test]
  fn cleanup_hook_registration_failure_rolls_back_host_registration() {
    let rollback_calls = std::sync::atomic::AtomicUsize::new(0);
    let result = install_cleanup_hook_or_rollback(
      || -> napi::Result<()> {
        Err(napi::Error::new(
          napi::Status::GenericFailure,
          "intentional cleanup-hook registration failure",
        ))
      },
      || {
        rollback_calls.fetch_add(1, std::sync::atomic::Ordering::SeqCst);
      },
    );

    assert!(result.is_err());
    assert_eq!(rollback_calls.load(std::sync::atomic::Ordering::SeqCst), 1);
  }

  #[test]
  fn host_driver_installation_allows_reentrant_eviction_before_publication() {
    let dead = std::sync::atomic::AtomicBool::new(false);
    let registration = std::sync::Mutex::new(None);
    let published = std::sync::atomic::AtomicBool::new(false);
    let rolled_back = std::sync::atomic::AtomicUsize::new(0);

    let installed = install_host_driver_registration(
      &dead,
      &registration,
      || {
        dead.store(true, std::sync::atomic::Ordering::SeqCst);
        let removed = registration
          .try_lock()
          .expect("the arbitrary installation callback must run outside the registration mutex")
          .take();
        assert!(
          removed.is_none(),
          "reentrant eviction must see no half-published registration"
        );
        17_u64
      },
      || {
        published.store(true, std::sync::atomic::Ordering::SeqCst);
      },
      |id| {
        assert_eq!(id, 17);
        let guard = registration
          .try_lock()
          .expect("rollback must run outside the registration mutex");
        drop(guard);
        rolled_back.fetch_add(1, std::sync::atomic::Ordering::SeqCst);
      },
    );

    assert!(
      !installed,
      "reentrant eviction must reject the exported registration"
    );
    assert!(!published.load(std::sync::atomic::Ordering::SeqCst));
    assert_eq!(rolled_back.load(std::sync::atomic::Ordering::SeqCst), 1);
    assert!(registration.lock().unwrap().is_none());
  }

  #[test]
  fn shared_runtime_rejects_after_shutdown_and_accepts_after_restart() {
    use std::{
      sync::{
        Arc,
        atomic::{AtomicBool, Ordering},
        mpsc,
      },
      time::Duration,
    };

    napi::bindgen_prelude::AsyncRuntime::start(&SharedAsyncRuntime)
      .expect("the adapter runtime must start");
    crate::async_runtime::reset_metrics();
    let (first_tx, first_rx) = mpsc::channel();
    napi::bindgen_prelude::AsyncRuntime::spawn_blocking(
      &SharedAsyncRuntime,
      Box::new(move || {
        first_tx
          .send(())
          .expect("test receiver must still be listening");
      }),
    )
    .unwrap_or_else(|_| panic!("the shared blocking lane must accept napi work"));
    first_rx
      .recv_timeout(Duration::from_secs(5))
      .expect("the shared blocking lane must execute accepted napi work");

    napi::bindgen_prelude::AsyncRuntime::shutdown(&SharedAsyncRuntime)
      .expect("the adapter runtime must shut down");
    let block_on_ran = Arc::new(AtomicBool::new(false));
    let block_on_ran_future = Arc::clone(&block_on_ran);
    let mut block_on_future = std::pin::pin!(async move {
      block_on_ran_future.store(true, Ordering::SeqCst);
    });
    let block_on_error =
      napi::bindgen_prelude::AsyncRuntime::block_on(&SharedAsyncRuntime, block_on_future.as_mut())
        .expect_err("the adapter must reject block_on while stopped");
    assert_eq!(
      block_on_error.reason,
      "the async runtime is stopped; call start before submitting work"
    );
    assert!(
      !block_on_ran.load(Ordering::SeqCst),
      "rejected block_on must not poll its future"
    );

    let rejected_ran = Arc::new(AtomicBool::new(false));
    let rejected_ran_work = Arc::clone(&rejected_ran);
    let (rejected_retry_tx, rejected_retry_rx) = mpsc::channel();
    let rejected = napi::bindgen_prelude::AsyncRuntime::spawn_blocking(
      &SharedAsyncRuntime,
      Box::new(move || {
        rejected_ran_work.store(true, Ordering::SeqCst);
        rejected_retry_tx
          .send(())
          .expect("test receiver must still be listening");
      }),
    )
    .expect_err("the adapter must reject work while stopped");
    assert!(!rejected_ran.load(Ordering::SeqCst));
    let (rejected, error) = rejected.into_parts();
    assert_eq!(
      error.reason,
      "the async runtime is stopped; call start before submitting work"
    );

    napi::bindgen_prelude::AsyncRuntime::start(&SharedAsyncRuntime)
      .expect("the adapter runtime must restart");
    napi::bindgen_prelude::AsyncRuntime::block_on(&SharedAsyncRuntime, block_on_future.as_mut())
      .expect("the restarted adapter runtime must accept the retained future");
    assert!(block_on_ran.load(Ordering::SeqCst));

    napi::bindgen_prelude::AsyncRuntime::spawn_blocking(&SharedAsyncRuntime, rejected)
      .unwrap_or_else(|_| panic!("the restarted runtime must accept the retained napi work"));
    rejected_retry_rx
      .recv_timeout(Duration::from_secs(5))
      .expect("the restarted runtime must execute the retained napi work");
    assert!(
      rejected_ran.load(Ordering::SeqCst),
      "rejected work must be returned intact"
    );
    assert!(
      crate::async_runtime::metrics().blocking_tasks_started >= 2,
      "both accepted adapter submissions should reach the shared blocking scheduler"
    );
  }

  #[cfg(not(target_family = "wasm"))]
  #[test]
  fn native_task_host_null_env_retires_payload_and_recovers_exact_delivery() {
    const CHILD_ENV: &str = "NAPI_RUNTIME_TEST_NATIVE_TASK_HOST_NULL_ENV_CHILD";

    if std::env::var_os(CHILD_ENV).is_some() {
      use std::sync::{Arc, mpsc};
      use std::time::Duration;

      use crate::async_runtime::{
        CurrentThreadTaskDelivery, CurrentThreadTaskDriver, RuntimeFlavor, RuntimeOptions,
        configure, register_current_thread_task_driver, shutdown, spawn_detached, start,
        unregister_current_thread_task_driver,
      };

      struct RecordingTaskDriver {
        dispatches: mpsc::Sender<CurrentThreadTaskDelivery>,
      }

      impl CurrentThreadTaskDriver for RecordingTaskDriver {
        fn dispatch(&self, delivery: CurrentThreadTaskDelivery) -> bool {
          self.dispatches.send(delivery).is_ok()
        }
      }

      configure(RuntimeOptions {
        flavor: RuntimeFlavor::CurrentThread,
        worker_threads: 1,
        max_blocking_tasks: 1,
        ..RuntimeOptions::default()
      })
      .expect("the isolated runtime must accept CurrentThread configuration");
      start().expect("the isolated CurrentThread runtime must start");

      let (dispatch_tx, dispatch_rx) = mpsc::channel();
      let driver_id = register_current_thread_task_driver(Arc::new(RecordingTaskDriver {
        dispatches: dispatch_tx,
      }));
      let (completed_tx, completed_rx) = mpsc::channel();
      spawn_detached(async move {
        completed_tx
          .send(())
          .expect("the completion observer must remain live");
      });

      let payload_drops = Arc::new(std::sync::atomic::AtomicUsize::new(0));
      let first_delivery = dispatch_rx
        .recv_timeout(Duration::from_secs(2))
        .expect("the queued task must publish its first host delivery");
      let first_payload = Box::into_raw(Box::new(
        NativeCurrentThreadTaskHostPayload::with_drop_observer(
          first_delivery,
          Arc::clone(&payload_drops),
        ),
      ))
      .cast();
      unsafe {
        call_native_current_thread_task_host(
          std::ptr::null_mut(),
          std::ptr::null_mut(),
          std::ptr::null_mut(),
          first_payload,
        );
      }
      assert_eq!(
        payload_drops.load(std::sync::atomic::Ordering::SeqCst),
        1,
        "the null-env callback must destroy its queued payload exactly once"
      );

      let replacement_delivery = dispatch_rx
        .recv_timeout(Duration::from_secs(2))
        .expect("the exact failed delivery must publish one replacement");
      assert_ne!(
        replacement_delivery, first_delivery,
        "recovery must use a fresh registration-scoped delivery capability"
      );
      let replacement_payload = Box::into_raw(Box::new(
        NativeCurrentThreadTaskHostPayload::with_drop_observer(
          replacement_delivery,
          Arc::clone(&payload_drops),
        ),
      ))
      .cast();
      let fake_env = std::ptr::NonNull::<std::ffi::c_void>::dangling()
        .as_ptr()
        .cast();
      unsafe {
        call_native_current_thread_task_host(
          fake_env,
          std::ptr::null_mut(),
          std::ptr::null_mut(),
          replacement_payload,
        );
      }

      completed_rx
        .recv_timeout(Duration::from_secs(2))
        .expect("the replacement host delivery must run the queued task");
      assert_eq!(
        payload_drops.load(std::sync::atomic::Ordering::SeqCst),
        2,
        "successful recovery must also destroy its queued payload exactly once"
      );
      assert!(
        matches!(dispatch_rx.try_recv(), Err(mpsc::TryRecvError::Empty)),
        "successful replacement acknowledgement must not publish another delivery"
      );

      unregister_current_thread_task_driver(driver_id);
      shutdown().expect("the isolated CurrentThread runtime must shut down cleanly");
      return;
    }

    let output = std::process::Command::new(std::env::current_exe().unwrap())
      .arg("--exact")
      .arg("adapter::tests::native_task_host_null_env_retires_payload_and_recovers_exact_delivery")
      .arg("--nocapture")
      .env(CHILD_ENV, "1")
      .output()
      .expect("the null-env task-host subprocess must start");
    assert!(
      output.status.success(),
      "the null-env task-host regression failed; status={:?}\nstdout={}\nstderr={}",
      output.status.code(),
      String::from_utf8_lossy(&output.stdout),
      String::from_utf8_lossy(&output.stderr)
    );
  }

  #[cfg(not(target_family = "wasm"))]
  #[test]
  fn native_task_host_callback_lease_covers_ack_payload_and_restart() {
    const CHILD_ENV: &str = "NAPI_RUNTIME_TEST_NATIVE_TASK_HOST_CALLBACK_LEASE_CHILD";

    if std::env::var_os(CHILD_ENV).is_some() {
      use std::sync::{Arc, mpsc};
      use std::time::{Duration, Instant};

      use crate::async_runtime::{
        CurrentThreadTaskDelivery, CurrentThreadTaskDriver, RuntimeFlavor, RuntimeOptions,
        configure, register_current_thread_task_driver, shutdown, spawn_detached, start,
        try_spawn_detached, unregister_current_thread_task_driver,
      };

      struct RecordingTaskDriver {
        dispatches: mpsc::Sender<CurrentThreadTaskDelivery>,
      }

      impl CurrentThreadTaskDriver for RecordingTaskDriver {
        fn dispatch(&self, delivery: CurrentThreadTaskDelivery) -> bool {
          self.dispatches.send(delivery).is_ok()
        }
      }

      configure(RuntimeOptions {
        flavor: RuntimeFlavor::CurrentThread,
        worker_threads: 1,
        max_blocking_tasks: 1,
        ..RuntimeOptions::default()
      })
      .expect("the isolated runtime must accept CurrentThread configuration");
      start().expect("the isolated CurrentThread runtime must start");

      let (dispatch_tx, dispatch_rx) = mpsc::channel();
      let driver_id = register_current_thread_task_driver(Arc::new(RecordingTaskDriver {
        dispatches: dispatch_tx,
      }));
      let (task_completed_tx, task_completed_rx) = mpsc::channel();
      spawn_detached(async move {
        task_completed_tx
          .send(())
          .expect("the task completion observer must remain live");
      });
      let delivery = dispatch_rx
        .recv_timeout(Duration::from_secs(2))
        .expect("the queued task must publish one native host delivery");

      let payload_drops = Arc::new(std::sync::atomic::AtomicUsize::new(0));
      let payload = Box::into_raw(Box::new(
        NativeCurrentThreadTaskHostPayload::with_drop_observer(
          delivery,
          Arc::clone(&payload_drops),
        ),
      ))
      .cast::<std::ffi::c_void>() as usize;
      let (after_drive_tx, after_drive_rx) = mpsc::channel();
      let (release_after_drive_tx, release_after_drive_rx) = mpsc::channel();
      let (after_payload_tx, after_payload_rx) = mpsc::channel();
      let (release_after_payload_tx, release_after_payload_rx) = mpsc::channel();
      let callback = std::thread::spawn(move || {
        NATIVE_TASK_HOST_AFTER_DRIVE_TEST_HOOK.with(|slot| {
          *slot.borrow_mut() = Some(Box::new(move || {
            after_drive_tx.send(()).unwrap();
            release_after_drive_rx.recv().unwrap();
          }));
        });
        NATIVE_TASK_HOST_AFTER_PAYLOAD_DROP_TEST_HOOK.with(|slot| {
          *slot.borrow_mut() = Some(Box::new(move || {
            after_payload_tx.send(()).unwrap();
            release_after_payload_rx.recv().unwrap();
          }));
        });
        let fake_env = std::ptr::NonNull::<std::ffi::c_void>::dangling()
          .as_ptr()
          .cast();
        unsafe {
          call_native_current_thread_task_host(
            fake_env,
            std::ptr::null_mut(),
            std::ptr::null_mut(),
            payload as *mut std::ffi::c_void,
          );
        }
      });

      after_drive_rx
        .recv_timeout(Duration::from_secs(2))
        .expect("the native callback must pause after driving the host turn");
      task_completed_rx
        .recv_timeout(Duration::from_secs(2))
        .expect("the host turn must run the queued task before acknowledgement");

      let (shutdown_tx, shutdown_rx) = mpsc::channel();
      let shutdown_thread = std::thread::spawn(move || {
        shutdown_tx.send(shutdown()).unwrap();
      });
      // Hang detector: the freshly spawned shutdown thread must get CPU and
      // publish the stopping lifecycle -- generous for starved CI runners.
      let stopping_deadline = Instant::now() + Duration::from_secs(10);
      loop {
        match try_spawn_detached(async {}) {
          Ok(()) => {
            assert!(
              Instant::now() < stopping_deadline,
              "shutdown did not publish the stopping lifecycle"
            );
            std::thread::yield_now();
          }
          Err(future) => {
            drop(future);
            break;
          }
        }
      }
      assert!(
        shutdown_rx
          .recv_timeout(Duration::from_millis(200))
          .is_err(),
        "shutdown must wait after drive until delivery acknowledgement and payload destruction"
      );

      let (restart_tx, restart_rx) = mpsc::channel();
      let restart_thread = std::thread::spawn(move || {
        start().expect("the runtime must restart after the old callback retires");
        restart_tx.send(()).unwrap();
      });
      assert!(
        restart_rx.recv_timeout(Duration::from_millis(200)).is_err(),
        "restart must not overlap the old callback before acknowledgement"
      );

      release_after_drive_tx.send(()).unwrap();
      after_payload_rx
        .recv_timeout(Duration::from_secs(2))
        .expect("the callback must pause after acknowledging and destroying its payload");
      assert_eq!(
        payload_drops.load(std::sync::atomic::Ordering::SeqCst),
        1,
        "the acknowledged callback must destroy its payload before releasing the lease"
      );
      assert!(
        shutdown_rx
          .recv_timeout(Duration::from_millis(200))
          .is_err(),
        "shutdown must remain blocked after acknowledgement and payload destruction"
      );
      assert!(
        restart_rx.recv_timeout(Duration::from_millis(200)).is_err(),
        "restart must remain blocked until the callback lease retires"
      );

      release_after_payload_tx.send(()).unwrap();
      callback.join().unwrap();
      shutdown_rx
        .recv_timeout(Duration::from_secs(2))
        .expect("shutdown must finish after the callback lease retires")
        .unwrap();
      restart_rx
        .recv_timeout(Duration::from_secs(2))
        .expect("restart must finish after old-generation shutdown");
      shutdown_thread.join().unwrap();
      restart_thread.join().unwrap();

      unregister_current_thread_task_driver(driver_id);
      shutdown().expect("the replacement CurrentThread runtime must shut down cleanly");
      return;
    }

    let output = std::process::Command::new(std::env::current_exe().unwrap())
      .arg("--exact")
      .arg("adapter::tests::native_task_host_callback_lease_covers_ack_payload_and_restart")
      .arg("--nocapture")
      .env(CHILD_ENV, "1")
      .output()
      .expect("the native task-host callback lease subprocess must start");
    assert!(
      output.status.success(),
      "the native task-host callback lease regression failed; status={:?}\nstdout={}\nstderr={}",
      output.status.code(),
      String::from_utf8_lossy(&output.stdout),
      String::from_utf8_lossy(&output.stderr)
    );
  }

  #[cfg(not(target_family = "wasm"))]
  #[test]
  fn timer_relay_terminal_drop_at_host_turn_budget_settles_without_later_work() {
    const CHILD_ENV: &str = "NAPI_RUNTIME_TEST_TIMER_RELAY_TERMINAL_DROP_CHILD";
    const HOST_TURN_RUNNABLE_BUDGET: usize = 64;

    if let Some(relay_state) = std::env::var_os(CHILD_ENV) {
      use std::sync::{Arc, mpsc};
      use std::time::Duration;

      use crate::async_runtime::{
        CurrentThreadTaskDelivery, CurrentThreadTaskDriver, RuntimeFlavor, RuntimeOptions,
        acknowledge_current_thread_task_delivery, cancel_current_thread_task_dispatch, configure,
        drive_current_thread_tasks, metrics, register_current_thread_task_driver, reset_metrics,
        shutdown, spawn, spawn_detached, start, unregister_current_thread_task_driver,
      };
      use futures::FutureExt as _;

      struct RecordingTaskDriver {
        dispatches: mpsc::Sender<CurrentThreadTaskDelivery>,
      }

      impl CurrentThreadTaskDriver for RecordingTaskDriver {
        fn dispatch(&self, delivery: CurrentThreadTaskDelivery) -> bool {
          self.dispatches.send(delivery).is_ok()
        }
      }

      let schedule_state = if relay_state == "complete" {
        RelayScheduleState::CallbackComplete
      } else {
        RelayScheduleState::AwaitingCallback
      };

      let options = RuntimeOptions {
        flavor: RuntimeFlavor::CurrentThread,
        worker_threads: 1,
        max_blocking_tasks: 1,
        ..RuntimeOptions::default()
      };
      configure(options).expect("the isolated runtime must accept CurrentThread configuration");
      start().expect("the isolated CurrentThread runtime must start");
      reset_metrics();

      let (dispatch_tx, dispatch_rx) = mpsc::channel();
      let driver_id = register_current_thread_task_driver(Arc::new(RecordingTaskDriver {
        dispatches: dispatch_tx,
      }));

      for _ in 0..HOST_TURN_RUNNABLE_BUDGET - 1 {
        spawn_detached(async {});
      }
      let state = Arc::new(TestPendingRelayState::default());
      let timer_task = spawn(BudgetBoundaryTimerWait {
        state: Arc::clone(&state),
        timer_id: 1,
        schedule_state,
        registered: false,
      });
      let first_dispatch = dispatch_rx
        .recv_timeout(Duration::from_secs(2))
        .expect("the initial host turn must be published");
      assert!(
        matches!(dispatch_rx.try_recv(), Err(mpsc::TryRecvError::Empty)),
        "all 64 initial runnables must coalesce behind one host turn"
      );

      let callback_lease = drive_current_thread_tasks(first_dispatch.capability())
        .expect("the exact host delivery must be admitted");
      acknowledge_current_thread_task_delivery(first_dispatch);
      drop(callback_lease);
      assert_eq!(
        metrics().runnable_polls,
        HOST_TURN_RUNNABLE_BUDGET as u64,
        "the timer operation must occupy the last poll in the host-turn budget"
      );
      assert_eq!(
        state
          .pending
          .lock()
          .unwrap_or_else(std::sync::PoisonError::into_inner)
          .len(),
        1,
        "the timer poll must register its pending relay"
      );
      assert_eq!(
        metrics().queued_runnables,
        1,
        "the unpolled relay must be the sole runnable beyond the budget boundary"
      );

      let failed_dispatch = dispatch_rx
        .recv_timeout(Duration::from_secs(2))
        .expect("budget exhaustion must publish a fresh relay turn");
      cancel_current_thread_task_dispatch(failed_dispatch.capability());
      let replacement_dispatch = dispatch_rx
        .recv_timeout(Duration::from_secs(2))
        .expect("the first scheduler failure must publish one replacement");
      assert_ne!(replacement_dispatch, failed_dispatch);
      cancel_current_thread_task_dispatch(replacement_dispatch.capability());

      assert!(
        state
          .pending
          .lock()
          .unwrap_or_else(std::sync::PoisonError::into_inner)
          .is_empty(),
        "terminal queue destruction must retire the matching pending relay"
      );
      let cancelled_relays = state
        .cancelled_relays
        .lock()
        .unwrap_or_else(std::sync::PoisonError::into_inner);
      if schedule_state == RelayScheduleState::CallbackComplete {
        assert_eq!(
          cancelled_relays.len(),
          1,
          "a callback-complete relay must queue exactly one host cancellation"
        );
      } else {
        assert!(
          cancelled_relays.is_empty(),
          "cleanup must not overtake a schedule callback that has not returned"
        );
      }
      drop(cancelled_relays);
      assert_eq!(metrics().queued_runnables, 0);
      assert!(
        matches!(dispatch_rx.try_recv(), Err(mpsc::TryRecvError::Empty)),
        "terminal recovery must stay bounded after the replacement failure"
      );
      let task_result = timer_task
        .now_or_never()
        .expect("the original timer operation must settle immediately");
      assert_eq!(
        task_result
          .expect_err("terminal recovery must cancel the pending timer operation")
          .to_string(),
        "async runtime stopped before the task completed"
      );

      unregister_current_thread_task_driver(driver_id);
      shutdown().expect("the isolated CurrentThread runtime must shut down cleanly");
      return;
    }

    for relay_state in ["awaiting", "complete"] {
      let output = std::process::Command::new(std::env::current_exe().unwrap())
        .arg("--exact")
        .arg(
          "adapter::tests::timer_relay_terminal_drop_at_host_turn_budget_settles_without_later_work",
        )
        .arg("--nocapture")
        .env(CHILD_ENV, relay_state)
        .output()
        .expect("the timer relay regression subprocess must start");
      assert!(
        output.status.success(),
        "{relay_state} timer relay terminal-drop regression failed; status={:?}\nstdout={}\nstderr={}",
        output.status.code(),
        String::from_utf8_lossy(&output.stdout),
        String::from_utf8_lossy(&output.stderr)
      );
    }
  }

  /// Relay-eviction invariant: the relay must evict ONLY on host death, and
  /// the decision must be STRING-FREE. A rejected JS promise coerces to
  /// `GenericFailure` carrying the JS-controlled rejection string (pinned
  /// napi 3.10 error.rs), so message matching is forgeable by a live
  /// callback. The two authorities: the unforgeable `Closing` status, and
  /// the liveness probe.
  #[test]
  fn relay_eviction_is_decided_by_status_and_probe_never_by_message() {
    use napi::Status;

    use super::should_evict_for_relay_error;

    // `Closing` is authoritative host death: it originates only from the
    // TSFN layer (aborted pre-check, raw napi_closing) and JS coercion can
    // never produce it -- evict even if the probe still reads live (the
    // finalize flag can lag the raw status).
    assert!(
      should_evict_for_relay_error(Status::Closing, true),
      "Closing overrides a live probe"
    );
    assert!(should_evict_for_relay_error(Status::Closing, false));

    // A DEAD probe evicts regardless of status. Queue drain during env
    // teardown and env death before a JS promise settles both coincide with
    // the env being torn down, which the probe observes directly.
    assert!(should_evict_for_relay_error(Status::GenericFailure, false));
    assert!(should_evict_for_relay_error(
      Status::PendingException,
      false
    ));

    // A LIVE host's failure takes the strike path no matter what the error
    // says. A live callback can reject with
    // `Promise.reject(new Error('oneshot canceled'))`, which coerces to
    // GenericFailure + "Error: oneshot canceled" and must not be mistaken for
    // environment teardown.
    assert!(
      !should_evict_for_relay_error(Status::GenericFailure, true),
      "a live host's GenericFailure -- e.g. a forged 'oneshot canceled' rejection -- must strike"
    );
    assert!(
      !should_evict_for_relay_error(Status::PendingException, true),
      "a JS throw on a live host must strike"
    );
    assert!(
      !should_evict_for_relay_error(Status::InvalidArg, true),
      "a wrong return type on a live host must strike"
    );
    assert!(
      !should_evict_for_relay_error(Status::QueueFull, true),
      "queue pressure on a live host must strike"
    );
  }

  #[test]
  fn cancellation_failures_share_the_three_strike_budget_without_cleanup_double_counting() {
    use std::sync::atomic::{AtomicU32, Ordering};

    use napi::Status;

    let failures = AtomicU32::new(0);
    for expected_strikes in 1..HOST_TIMER_MAX_TRANSIENT_FAILURES {
      let relay_health = HostTimerRelayHealth::default();
      assert_eq!(
        record_host_timer_failure(&failures, &relay_health, Status::GenericFailure, true),
        HostTimerFailureAction::Retry(expected_strikes),
      );
      assert_eq!(
        record_host_timer_failure(&failures, &relay_health, Status::GenericFailure, true),
        HostTimerFailureAction::Duplicate,
        "one relay failure must consume at most one strike"
      );
    }

    let cleanup_health = HostTimerRelayHealth::default();
    cleanup_health.set_cancellation_accounting(RelayCancellationAccounting::CleanupOnly);
    assert_eq!(
      cleanup_health.cancellation_accounting(),
      RelayCancellationAccounting::CleanupOnly
    );
    assert_eq!(
      failures.load(Ordering::SeqCst),
      HOST_TIMER_MAX_TRANSIENT_FAILURES - 1,
      "schedule-failure and eviction cleanup must not mutate host health"
    );

    let final_health = HostTimerRelayHealth::default();
    assert_eq!(
      record_host_timer_failure(&failures, &final_health, Status::GenericFailure, true),
      HostTimerFailureAction::EvictHostAfterStrikes(HOST_TIMER_MAX_TRANSIENT_FAILURES),
    );
  }

  #[test]
  fn a_same_relay_success_cannot_erase_its_recorded_failure() {
    use std::sync::atomic::{AtomicU32, Ordering};

    use napi::Status;

    let failures = AtomicU32::new(1);
    let healthy_relay = HostTimerRelayHealth::default();
    reset_host_timer_failures_after_success(&failures, &healthy_relay);
    assert_eq!(failures.load(Ordering::SeqCst), 0);

    failures.store(1, Ordering::SeqCst);
    let failed_relay = HostTimerRelayHealth::default();
    assert_eq!(
      record_host_timer_failure(&failures, &failed_relay, Status::GenericFailure, true),
      HostTimerFailureAction::Retry(2),
    );
    reset_host_timer_failures_after_success(&failures, &failed_relay);
    assert_eq!(
      failures.load(Ordering::SeqCst),
      2,
      "a later success from the same relay must not hide its cancellation failure"
    );
  }

  #[test]
  fn resolve_thread_count_handles_env_inputs() {
    const DEFAULT: usize = 8;
    const MAXIMUM: usize = 16;
    // A `0` typo must be treated as unset, never fed to a constructor that
    // rejects 0 (`RuntimeOptions` validation), which would panic the host's
    // module_init `expect`.
    assert_eq!(
      resolve_thread_count(Some("0".to_string()), DEFAULT, MAXIMUM),
      DEFAULT
    );
    // Valid positive numbers pass through unchanged.
    assert_eq!(
      resolve_thread_count(Some("4".to_string()), DEFAULT, MAXIMUM),
      4
    );
    // Valid but unsafe values clamp to the construction limit.
    assert_eq!(
      resolve_thread_count(Some("1000".to_string()), DEFAULT, MAXIMUM),
      MAXIMUM
    );
    assert_eq!(
      resolve_thread_count(Some(usize::MAX.to_string()), DEFAULT, MAXIMUM),
      MAXIMUM
    );
    // Parse overflow is invalid input and falls back instead of wrapping.
    assert_eq!(
      resolve_thread_count(Some("18446744073709551616".to_string()), DEFAULT, MAXIMUM,),
      DEFAULT
    );
    // Non-numeric values fall back gracefully (unchanged behaviour).
    assert_eq!(
      resolve_thread_count(Some("abc".to_string()), DEFAULT, MAXIMUM),
      DEFAULT
    );
    // Unset falls back to the default.
    assert_eq!(resolve_thread_count(None, DEFAULT, MAXIMUM), DEFAULT);
    // Host-derived defaults are subject to the same production ceiling.
    assert_eq!(resolve_thread_count(None, 32, MAXIMUM), MAXIMUM);
  }

  #[test]
  #[should_panic(expected = "the default thread count must be positive")]
  fn resolve_thread_count_rejects_an_invalid_default() {
    let _ = resolve_thread_count(None, 0, 1);
  }

  #[test]
  fn resolve_drain_linger_handles_env_inputs() {
    use std::time::Duration;

    use crate::async_runtime::{DEFAULT_DRAIN_LINGER_MICROS, MAX_DRAIN_LINGER_MICROS};

    use super::resolve_drain_linger;

    let fallback = Duration::from_micros(DEFAULT_DRAIN_LINGER_MICROS);
    let ceiling = Duration::from_micros(MAX_DRAIN_LINGER_MICROS);
    // Valid values pass through unchanged.
    assert_eq!(
      resolve_drain_linger(Some("250".to_string()), fallback),
      Duration::from_micros(250)
    );
    // `0` disables lingering (it is a VALUE here, unlike thread counts).
    assert_eq!(
      resolve_drain_linger(Some("0".to_string()), fallback),
      Duration::ZERO
    );
    // u64::MAX -- the exact value that used to wedge the suite through the
    // removed first-drainer OnceLock -- clamps to the ceiling.
    assert_eq!(
      resolve_drain_linger(Some(u64::MAX.to_string()), fallback),
      ceiling
    );
    // Parse overflow is invalid input and keeps the host-provided field.
    assert_eq!(
      resolve_drain_linger(Some("18446744073709551616".to_string()), fallback),
      fallback
    );
    // Non-numeric and unset keep the host-provided field.
    assert_eq!(
      resolve_drain_linger(Some("abc".to_string()), fallback),
      fallback
    );
    assert_eq!(resolve_drain_linger(None, fallback), fallback);
    // Whitespace is tolerated, matching the removed in-scheduler parse.
    assert_eq!(
      resolve_drain_linger(Some(" 750 ".to_string()), fallback),
      Duration::from_micros(750)
    );
    // An out-of-range host fallback is bounded too.
    assert_eq!(
      resolve_drain_linger(None, Duration::from_micros(u64::MAX)),
      ceiling
    );
  }
}