feat(runtime,io): driver-enqueues + park/wake idle path — retire the wake pipe

The swap (RFC 018). Schedulers no longer sleep on a shared level-triggered
wake pipe — the herd source that made the default 8-thread config 7x
slower than 2 threads (E1). They park on per-thread futex parkers via the
coordination layer; IO backends become producers behind a two-call
contract (make runnable, then the enqueue tail wakes exactly one parked
scheduler).

Deleted: the drain lock and the one-winner phase-1 drain; the shared
completions VecDeque; the wake pipe fds, poll_wake, drain_wake_pipe,
wake_scheduler, the FdReady/Blocking Completion enum; the 100us idle nap;
the per-pop io.lock liveness read; io.rs's as_millis timeout truncation.

Added:
- enqueue wake tail (fixes the silent enqueue): wake_one_if_idle, a fence
  + one Relaxed mask load when everyone is busy — the pure-compute hot
  path pays almost nothing.
- driver-enqueues: the pool thread stashes its result in the slot,
  decrements io_outstanding, unparks; the epoll thread removes+DELs the
  waiter under the waiters lock and unparks. Both reach the runtime via a
  Weak (no Arc cycle). The waiters map moves behind its own Arc<Mutex> so
  the epoll thread never takes the runtime io lock (teardown holds it
  while joining that thread).
- io_outstanding / io_fd_waiters atomics: the termination verdict reads
  two atomics instead of taking io.lock on every pop.
- timekeeper idle path: at most one parked scheduler holds the timer
  deadline (an expiry wakes one, not a herd); everyone else parks
  indefinitely and is woken by the enqueue tail.
- busy-path timer due-check (ratified design point (a)): under saturation
  nobody parks and no timekeeper exists, yet due timers must still fire —
  one Relaxed load of the earliest-deadline snapshot per loop, clock read
  only when a timer is armed. Maintained under the timers mutex.
- chain rule: a scheduler that pops with more work queued and a sibling
  parked wakes one, so surplus runs in parallel rather than behind it.

tests/park_wake.rs pins the two new observable properties: timers fire
under full scheduler saturation, and sub-ms sleeps are prompt (the
as_millis truncation regression). Full suite + all loom models green;
clippy --lib clean.

.githooks/pre-commit

@@ -2,7 +2,23 @@
 # smarm pre-commit gate: clippy the library (src/) with warnings as errors.
 # unwrap_used / expect_used are denied (Cargo.toml [lints.clippy]): library
 # code must not hide a panic behind unwrap/expect. Tests/examples are not gated.
+#
+# Toolchain resolution: prefer an installed cargo-clippy; on machines whose
+# rust comes without the clippy component (e.g. NixOS home-manager), fall
+# back to an ephemeral nix-shell toolchain. The fallback uses its own target
+# dir (target/clippy) because the shell's rustc version may differ from the
+# default toolchain's — mixed-compiler artifacts in one target dir are an
+# E0514 hard error. MSRV (Cargo.toml rust-version) keeps the older shell
+# toolchain a legitimate gate.
 set -eu
 [ -f "$HOME/.cargo/env" ] && . "$HOME/.cargo/env"
 cd "$(git rev-parse --show-toplevel)"
+if cargo clippy --version >/dev/null 2>&1; then
     cargo clippy --lib -- -D warnings
+elif command -v nix-shell >/dev/null 2>&1; then
+    nix-shell -p clippy -p cargo -p rustc \
+        --run 'CARGO_TARGET_DIR=target/clippy cargo clippy --lib -- -D warnings'
+else
+    echo "pre-commit: cargo clippy unavailable and no nix-shell fallback" >&2
+    exit 1
+fi

src/io.rs

@@ -13,44 +13,68 @@
 //! leaves the actor, no copying through an intermediary thread. Built on
 //! these are the conveniences `read(fd, &mut buf)` and `write(fd, &buf)`.
 //!
-//! Architecture
-//! ============
-//! Per `run()`, two OS threads:
-//!   - **epoll thread**: owns the epollfd. Loops in `epoll_wait`. On a
-//!     ready fd, pushes `Completion::FdReady { pid, fd, events }` to the
-//!     shared completion queue and writes the scheduler-wake pipe. On the
-//!     shutdown pipe (also registered in epollfd), exits.
-//!   - **pool thread**: blocks on the request mpsc. Runs the closure
-//!     inside `catch_unwind`, pushes `Completion::Blocking { pid, result }`,
-//!     writes the scheduler-wake pipe.
+//! Architecture (RFC 018: driver-enqueues)
+//! =======================================
+//! Per `run()`, two OS threads, each a *producer* behind the runtime's
+//! two-call contract — make the actor runnable (`unpark_at`, whose enqueue
+//! tail wakes a parked scheduler), nothing else:
 //!
-//! Both threads share a single `completions: Arc<Mutex<VecDeque<Completion>>>`
-//! and the same scheduler-wake pipe.
+//!   - **epoll thread**: owns `epoll_wait` on the epollfd. On a ready fd it
+//!     removes the parked waiter from the shared `waiters` map and DELs the
+//!     fd (both under the waiters lock — see below), then unparks the
+//!     actor directly. On the shutdown pipe (also registered in the
+//!     epollfd), exits.
+//!   - **pool thread**: blocks on the request mpsc. Runs the closure inside
+//!     `catch_unwind`, stashes the result in the actor's slot
+//!     (`pending_io_result`, under the cold lock, generation-checked),
+//!     decrements the runtime's `io_outstanding`, and unparks the actor.
 //!
-//! `epoll_ctl` (register/unregister fd interest) is called by the
-//! scheduler thread *directly* on the epollfd. That's well-defined per
-//! `epoll_ctl(2)`: a thread may be calling `epoll_wait` on the epollfd
-//! while another thread calls `epoll_ctl`. Avoids needing a second mpsc
-//! and a second wake mechanism.
+//! There is no shared completion queue and no wake pipe: each producer
+//! routes its own completion, so the whole byte-vs-completion visibility
+//! discipline of the drain era — and the stranded-completion hazards it
+//! defended against — is unrepresentable. Producers reach the runtime
+//! through a `Weak<RuntimeInner>`: upgraded per completion (the path is
+//! syscall-bound; the refcount op is noise) and avoiding an Arc cycle
+//! through `RuntimeInner::io`.
+//!
+//! `epoll_ctl` (register fd interest) is called by the scheduler thread
+//! directly on the epollfd. That's well-defined per `epoll_ctl(2)`: a
+//! thread may be calling `epoll_wait` on the epollfd while another thread
+//! calls `epoll_ctl`.
 //!
 //! Epoll mode
 //! ==========
 //! Level-triggered with EPOLLONESHOT. After a wakeup the kernel
 //! auto-disarms the fd, so we never get two wakeups for one
-//! `wait_readable` call. The scheduler explicitly `EPOLL_CTL_DEL`s the fd
-//! on completion to free the slot for re-registration. Net effect: each
+//! `wait_readable` call. The epoll thread explicitly `EPOLL_CTL_DEL`s the
+//! fd on readiness to free the slot for re-registration. Net effect: each
 //! `wait_readable(fd)` is one ADD, one wakeup, one DEL — symmetric and
 //! stateless between calls.
 //!
+//! ## The waiters lock is the ADD/DEL serialization
+//!
+//! Registration (scheduler thread: check-vacant, defensive DEL, ADD,
+//! insert) and readiness consumption (epoll thread: remove, DEL) each run
+//! entirely under the `waiters` mutex. This is what makes the
+//! oneshot-rearm race unrepresentable: a woken actor re-registering the
+//! same fd cannot interleave with the epoll thread's DEL for the *previous*
+//! registration — whichever takes the lock second sees a consistent
+//! kernel-side state. Lock order: `io` (the runtime's outer mutex, held by
+//! scheduler-side callers) → `waiters` → slot/queue leaves via `unpark_at`.
+//! The epoll thread takes `waiters` without `io` — it must never take
+//! `io`, both for lock-order hygiene and because teardown holds `io` while
+//! joining it.
+//!
 //! Fd hygiene
 //! ==========
 //! An actor stopped while waiting on an fd unwinds out of `wait_fd`'s park;
 //! a drop guard there (armed after a successful register, forgotten on a
-//! normal wake) removes the `waiters` entry iff it is still that wait's
-//! `(pid, epoch)` and only then `EPOLL_CTL_DEL`s the fd — an entry already
-//! consumed by a racing `FdReady` means the fd may carry someone else's
-//! fresh registration, which must be left alone. `epoll_register` keeps a
-//! defensive bare DEL before ADD as belt-and-braces.
+//! normal wake) calls [`IoThread::cancel_waiter`], which removes the
+//! `waiters` entry iff it is still that wait's `(pid, epoch)` and only then
+//! `EPOLL_CTL_DEL`s the fd — an entry already consumed by the epoll thread
+//! means the fd may carry someone else's fresh registration, which must be
+//! left alone. `epoll_register` keeps a defensive bare DEL before ADD as
+//! belt-and-braces.
 //!
 //! Buffers used with `read`/`write` should be on fds opened with
 //! `O_NONBLOCK`. If they aren't, the syscall may block the scheduler
@@ -68,13 +92,14 @@
 //! they have no equivalent panic-propagation path.
 
 use crate::pid::Pid;
+use crate::runtime::RuntimeInner;
 use std::any::Any;
-use std::collections::{HashMap, VecDeque};
+use std::collections::HashMap;
 use std::io;
 use std::os::fd::RawFd;
 use std::panic;
-use std::sync::mpsc;
-use std::sync::{Arc, Mutex};
+use std::sync::atomic::Ordering;
+use std::sync::{mpsc, Arc, Mutex, Weak};
 use std::thread::JoinHandle as OsJoinHandle;
 
 // ---------------------------------------------------------------------------
@@ -86,42 +111,29 @@ use std::thread::JoinHandle as OsJoinHandle;
 pub type IoResult = Result<Box<dyn Any + Send>, Box<dyn Any + Send>>;
 
 struct Request {
-    /// The submitter's park-epoch — carried through to the `Blocking`
-    /// completion so the wake is epoch-matched.
+    /// The submitter's park-epoch — the eventual wake is epoch-matched.
     epoch: u32,
     pid: Pid,
     /// The work to perform. Returns the wire-form result directly.
     work: Box<dyn FnOnce() -> IoResult + Send>,
 }
 
-/// Completion message from either IO thread back to the scheduler.
-pub enum Completion {
-    /// A `block_on_io` closure has finished (Ok = return value, Err = panic
-    /// payload).
-    Blocking { pid: Pid, epoch: u32, result: IoResult },
-    /// An fd registered via `wait_readable`/`wait_writable` is ready. The
-    /// scheduler looks up the parked pid in `waiters`, unparks it, and
-    /// removes the entry. `pid` isn't in this variant because the epoll
-    /// thread doesn't have access to the `waiters` map; the scheduler
-    /// thread owns that.
-    FdReady { fd: RawFd, events: u32 },
-}
+/// The parked-waiter map, shared between scheduler-side registration and
+/// the epoll thread's readiness consumption. See the module docs on why
+/// this single lock is the ADD/DEL serialization.
+type Waiters = Arc<Mutex<HashMap<RawFd, (Pid, u32)>>>;
 
 // ---------------------------------------------------------------------------
-// IoThread — created per `run()`, owned by `SchedulerState`.
+// IoThread — created per `run()`, owned by `RuntimeInner::io`.
 // ---------------------------------------------------------------------------
 
 pub struct IoThread {
-    // ----- Channels & queues -----
-
     /// Submission queue into the blocking-work pool.
     tx: mpsc::Sender<Request>,
-    /// Shared completion queue, fed by both the pool and the epoll thread.
-    completions: Arc<Mutex<VecDeque<Completion>>>,
-    /// Pipe the scheduler polls in its idle path. Both IO threads write to
-    /// `wake_write` after pushing a completion.
-    wake_read: RawFd,
-    wake_write: RawFd,
+    /// One parked actor per registered fd. Populated by `epoll_register`,
+    /// consumed by the epoll thread on readiness or `cancel_waiter` on an
+    /// unwound wait.
+    waiters: Waiters,
 
     // ----- Epoll machinery -----
 
@@ -133,39 +145,25 @@ pub struct IoThread {
     /// shutdown.
     shutdown_read: RawFd,
     shutdown_write: RawFd,
-    /// One parked actor per registered fd. Populated by `wait_readable` /
-    /// `wait_writable` and drained by the scheduler when a `FdReady`
-    /// completion is processed.
-    pub waiters: HashMap<RawFd, (Pid, u32)>,
 
     // ----- Threads -----
 
     pool_thread: Option<OsJoinHandle<()>>,
     epoll_thread: Option<OsJoinHandle<()>>,
-
-    /// Number of `block_on_io` requests in-flight. Used by the scheduler's
-    /// idle path to decide whether to wait on the pipe or exit. Fd waits
-    /// are not counted here; they're counted by `waiters.len()`.
-    pub outstanding: u32,
 }
 
 impl IoThread {
-    pub fn start() -> io::Result<Self> {
-        // Scheduler-facing wake pipe.
-        let (wake_read, wake_write) = make_pipe()?;
-        // Pool submission channel + shared completion queue.
+    /// Start the pool and epoll threads. `rt` is the producers' route back
+    /// into the runtime (slot table + unpark protocol); a `Weak` so the
+    /// `RuntimeInner → IoThread → RuntimeInner` cycle never forms.
+    pub(crate) fn start(rt: Weak<RuntimeInner>) -> io::Result<Self> {
+        // Pool submission channel.
         let (tx, rx) = mpsc::channel::<Request>();
-        let completions: Arc<Mutex<VecDeque<Completion>>> =
-            Arc::new(Mutex::new(VecDeque::new()));
+        let waiters: Waiters = Arc::new(Mutex::new(HashMap::new()));
 
         // Epoll machinery.
         let epollfd = unsafe { libc::epoll_create1(libc::EPOLL_CLOEXEC) };
         if epollfd < 0 {
-            // Best-effort fd cleanup before bailing.
-            unsafe {
-                libc::close(wake_read);
-                libc::close(wake_write);
-            }
             return Err(io::Error::last_os_error());
         }
 
@@ -174,8 +172,6 @@ impl IoThread {
             Err(e) => {
                 unsafe {
                     libc::close(epollfd);
-                    libc::close(wake_read);
-                    libc::close(wake_write);
                 }
                 return Err(e);
             }
@@ -202,42 +198,37 @@ impl IoThread {
                 libc::close(epollfd);
                 libc::close(shutdown_read);
                 libc::close(shutdown_write);
-                libc::close(wake_read);
-                libc::close(wake_write);
             }
             return Err(e);
         }
 
         // Spawn pool thread.
-        let pool_comps = completions.clone();
+        let pool_rt = rt.clone();
         let pool_thread = std::thread::Builder::new()
             .name("smarm-io-pool".into())
-            .spawn(move || pool_loop(rx, pool_comps, wake_write))?;
+            .spawn(move || pool_loop(rx, pool_rt))?;
 
         // Spawn epoll thread.
-        let epoll_comps = completions.clone();
+        let epoll_waiters = waiters.clone();
         let epoll_thread = std::thread::Builder::new()
             .name("smarm-io-epoll".into())
-            .spawn(move || epoll_loop(epollfd, epoll_comps, wake_write))?;
+            .spawn(move || epoll_loop(epollfd, epoll_waiters, rt))?;
 
         Ok(Self {
             tx,
-            completions,
-            wake_read,
-            wake_write,
+            waiters,
             epollfd,
             shutdown_read,
             shutdown_write,
-            waiters: HashMap::new(),
             pool_thread: Some(pool_thread),
             epoll_thread: Some(epoll_thread),
-            outstanding: 0,
         })
     }
 
-    /// Hand a request to the pool. Increments `outstanding`.
+    /// Hand a request to the pool. The caller (scheduler.rs) increments
+    /// `io_outstanding` BEFORE calling — the pool decrements on completion,
+    /// and an increment that trailed the completion would underflow.
     pub fn submit(&mut self, pid: Pid, epoch: u32, work: Box<dyn FnOnce() -> IoResult + Send>) {
-        self.outstanding += 1;
         // Send can only fail if the pool has hung up, which only happens
         // on shutdown. submit during shutdown is a bug.
         if self.tx.send(Request { pid, epoch, work }).is_err() {
@@ -245,39 +236,13 @@ impl IoThread {
         }
     }
 
-    /// Drain every available completion. Caller (the scheduler) routes the
-    /// results and updates `outstanding` / `waiters` accordingly.
-    pub fn drain_completions(&mut self) -> Vec<Completion> {
-        let mut q = match self.completions.lock() {
-            Ok(g) => g,
-            Err(e) => panic!("smarm: io completions lock poisoned (core corrupt): {e}"),
-        };
-        let mut out = Vec::with_capacity(q.len());
-        while let Some(c) = q.pop_front() {
-            out.push(c);
-        }
-        out
-    }
-
-    pub fn wake_fd(&self) -> RawFd {
-        self.wake_read
-    }
-
-    /// Write the wake pipe directly: rouse every scheduler thread blocked in
-    /// its idle `poll_wake`. Used by the terminal (AllDone) path — an idle
-    /// sibling may be blocked on a snapshot that nothing will ever refresh
-    /// (an orphaned timer deadline, or `io_outstanding` from a waiter that
-    /// was stop-cancelled and so never produces a completion).
-    pub fn wake(&self) {
-        wake_scheduler(self.wake_write);
-    }
-
     /// Register interest in `fd` becoming readable/writable; record `pid`
-    /// as the parked waiter. The epoll thread will push a `FdReady`
-    /// completion when the kernel signals.
+    /// as the parked waiter. The epoll thread unparks it on readiness.
+    /// The caller increments `io_fd_waiters` BEFORE calling (mirror of
+    /// `submit`'s contract) and decrements it again if this errors.
     ///
-    /// EPOLLONESHOT: one wakeup per registration. The scheduler must
-    /// `epoll_del` on completion to free the slot for re-registration.
+    /// EPOLLONESHOT: one wakeup per registration; the epoll thread DELs on
+    /// readiness, `cancel_waiter` DELs on an unwound wait.
     pub fn epoll_register(
         &mut self,
         fd: RawFd,
@@ -286,20 +251,24 @@ impl IoThread {
         readable: bool,
         writable: bool,
     ) -> io::Result<()> {
+        let mut waiters = match self.waiters.lock() {
+            Ok(g) => g,
+            Err(e) => panic!("smarm: io waiters lock poisoned (core corrupt): {e}"),
+        };
         // Two actors waiting on the same fd would be a misuse: the kernel
         // delivers exactly one EPOLLONESHOT wakeup, so the second waiter
         // would hang. Reject up front.
-        if self.waiters.contains_key(&fd) {
+        if waiters.contains_key(&fd) {
             return Err(io::Error::new(
                 io::ErrorKind::AlreadyExists,
                 "fd already has a parked waiter",
             ));
         }
 
-        // Belt-and-braces: the unwind guard in `wait_fd` is responsible for
-        // cleaning up a stopped waiter's registration, but a bare DEL is
-        // harmless if the fd isn't registered (ENOENT) and removes any leak
-        // a path we haven't thought of might leave behind.
+        // Belt-and-braces: `cancel_waiter` is responsible for cleaning up a
+        // stopped waiter's registration, but a bare DEL is harmless if the
+        // fd isn't registered (ENOENT) and removes any leak a path we
+        // haven't thought of might leave behind.
         unsafe {
             libc::epoll_ctl(self.epollfd, libc::EPOLL_CTL_DEL, fd, std::ptr::null_mut());
         }
@@ -321,20 +290,30 @@ impl IoThread {
         if r < 0 {
             return Err(io::Error::last_os_error());
         }
-        self.waiters.insert(fd, (pid, epoch));
+        waiters.insert(fd, (pid, epoch));
         Ok(())
     }
 
-    /// Remove `fd` from the epollfd. Called by the scheduler after a
-    /// `FdReady` completion, so the next `wait_readable(fd)` can ADD again.
-    ///
-    /// Does NOT touch `waiters` — that's the scheduler's bookkeeping; this
-    /// is purely the kernel-side cleanup.
-    pub fn epoll_deregister(&mut self, fd: RawFd) {
+    /// Remove `fd`'s waiter iff it is still `(pid, epoch)`, DELing the fd
+    /// from the epollfd in the same critical section. Returns whether the
+    /// entry was removed (the caller then decrements `io_fd_waiters`).
+    /// `false` means the epoll thread consumed the registration first —
+    /// the fd may already carry someone else's fresh ADD; hands off.
+    pub fn cancel_waiter(&mut self, fd: RawFd, pid: Pid, epoch: u32) -> bool {
+        let mut waiters = match self.waiters.lock() {
+            Ok(g) => g,
+            Err(e) => panic!("smarm: io waiters lock poisoned (core corrupt): {e}"),
+        };
+        if waiters.get(&fd) == Some(&(pid, epoch)) {
+            waiters.remove(&fd);
             // EPOLL_CTL_DEL of an already-removed fd returns ENOENT; ignore.
             unsafe {
                 libc::epoll_ctl(self.epollfd, libc::EPOLL_CTL_DEL, fd, std::ptr::null_mut());
             }
+            true
+        } else {
+            false
+        }
     }
 }
 
@@ -354,7 +333,10 @@ impl Drop for IoThread {
         let real_tx = std::mem::replace(&mut self.tx, dead_tx);
         drop(real_tx);
 
-        // 3. Join both threads.
+        // 3. Join both threads. Safe even while the caller holds the
+        //    runtime's `io` mutex: neither thread ever takes it (they reach
+        //    the runtime through a Weak they upgrade per completion, and
+        //    the epoll thread's only lock is `waiters`).
         if let Some(h) = self.epoll_thread.take() {
             let _ = h.join();
         }
@@ -367,8 +349,6 @@ impl Drop for IoThread {
             libc::close(self.epollfd);
             libc::close(self.shutdown_read);
             libc::close(self.shutdown_write);
-            libc::close(self.wake_read);
-            libc::close(self.wake_write);
         }
     }
 }
@@ -379,36 +359,38 @@ impl Drop for IoThread {
 const SHUTDOWN_EPOLL_TOKEN: u64 = u64::MAX;
 
 // ---------------------------------------------------------------------------
-// Pool loop
+// Pool loop (producer: Blocking completions)
 // ---------------------------------------------------------------------------
 
-fn pool_loop(
-    rx: mpsc::Receiver<Request>,
-    completions: Arc<Mutex<VecDeque<Completion>>>,
-    wake_write: RawFd,
-) {
+fn pool_loop(rx: mpsc::Receiver<Request>, rt: Weak<RuntimeInner>) {
     while let Ok(Request { pid, epoch, work }) = rx.recv() {
         let result: IoResult = match panic::catch_unwind(panic::AssertUnwindSafe(work)) {
             Ok(r) => r,
             Err(payload) => Err(payload),
         };
-        match completions.lock() {
-            Ok(mut g) => g.push_back(Completion::Blocking { pid, epoch, result }),
-            Err(e) => panic!("smarm: io completions lock poisoned (core corrupt): {e}"),
+        let Some(inner) = rt.upgrade() else { return };
+        // Stash the result under the cold lock (generation-checked: an
+        // actor stopped with the op in flight discards it), decrement the
+        // in-flight count, then wake through the epoch-matched unpark. The
+        // unpark's enqueue tail wakes a parked scheduler; the actor stays
+        // `live` until it resumes and finalizes, so the decrement's
+        // ordering against the termination verdict is not load-bearing.
+        if let Some(slot) = inner.slot_at(pid) {
+            let mut cold = slot.cold.lock();
+            if slot.generation() == pid.generation() {
+                cold.pending_io_result = Some(result);
             }
-        wake_scheduler(wake_write);
+        }
+        inner.io_outstanding.fetch_sub(1, Ordering::AcqRel);
+        inner.unpark_at(pid, epoch);
     }
 }
 
 // ---------------------------------------------------------------------------
-// Epoll loop
+// Epoll loop (producer: FdReady completions)
 // ---------------------------------------------------------------------------
 
-fn epoll_loop(
-    epollfd: RawFd,
-    completions: Arc<Mutex<VecDeque<Completion>>>,
-    wake_write: RawFd,
-) {
+fn epoll_loop(epollfd: RawFd, waiters: Waiters, rt: Weak<RuntimeInner>) {
     // Buffer for epoll_wait. 64 is plenty for our scale; if a real load
     // appears that needs more, this is a one-line change.
     const MAX_EVENTS: usize = 64;
@@ -436,29 +418,41 @@ fn epoll_loop(
         }
 
         let mut shutdown_requested = false;
-        let mut pushed_any = false;
-        {
-            let mut q = match completions.lock() {
-                Ok(g) => g,
-                Err(e) => panic!("smarm: io completions lock poisoned (core corrupt): {e}"),
-            };
         for ev in events.iter().take(n as usize) {
             if ev.u64 == SHUTDOWN_EPOLL_TOKEN {
                 shutdown_requested = true;
                 continue;
             }
             let fd = ev.u64 as RawFd;
-                let evs = ev.events;
-                q.push_back(Completion::FdReady {
+            // Consume the registration: remove + DEL under the waiters
+            // lock (the ADD/DEL serialization — see module docs). A
+            // vanished entry means `cancel_waiter` beat us: the wake is
+            // already moot.
+            let entry = {
+                let mut w = match waiters.lock() {
+                    Ok(g) => g,
+                    Err(e) => {
+                        panic!("smarm: io waiters lock poisoned (core corrupt): {e}")
+                    }
+                };
+                let entry = w.remove(&fd);
+                if entry.is_some() {
+                    unsafe {
+                        libc::epoll_ctl(
+                            epollfd,
+                            libc::EPOLL_CTL_DEL,
                             fd,
-                    events: evs,
-                });
-                pushed_any = true;
+                            std::ptr::null_mut(),
+                        );
                     }
                 }
-
-        if pushed_any {
-            wake_scheduler(wake_write);
+                entry
+            };
+            if let Some((pid, epoch)) = entry {
+                let Some(inner) = rt.upgrade() else { return };
+                inner.io_fd_waiters.fetch_sub(1, Ordering::AcqRel);
+                inner.unpark_at(pid, epoch);
+            }
         }
         if shutdown_requested {
             return;
@@ -466,27 +460,8 @@ fn epoll_loop(
     }
 }
 
-/// Write one byte to the scheduler's wake pipe. Retries on EINTR; ignores
-/// EAGAIN (pipe full means there's already an outstanding wake we haven't
-/// consumed yet, which is sufficient).
-fn wake_scheduler(wake_write: RawFd) {
-    let buf: [u8; 1] = [0];
-    unsafe {
-        loop {
-            let n = libc::write(wake_write, buf.as_ptr() as *const _, 1);
-            if n < 0 {
-                let e = *libc::__errno_location();
-                if e == libc::EINTR {
-                    continue;
-                }
-            }
-            break;
-        }
-    }
-}
-
 // ---------------------------------------------------------------------------
-// Pipe helpers (unchanged from v0.2)
+// Pipe helper
 // ---------------------------------------------------------------------------
 
 fn make_pipe() -> io::Result<(RawFd, RawFd)> {
@@ -497,50 +472,3 @@ fn make_pipe() -> io::Result<(RawFd, RawFd)> {
     }
     Ok((fds[0], fds[1]))
 }
-
-/// Drain pending bytes from the wake pipe. Nonblocking (pipe is O_NONBLOCK).
-///
-/// DISCIPLINE: called only by the phase-1 drain-lock winner, immediately
-/// before `drain_completions`. Bytes are the notification channel for
-/// completions; consuming one anywhere else can strand the completion it
-/// announces (see the lost-wakeup note at the call site in `schedule_loop`).
-pub fn drain_wake_pipe(fd: RawFd) {
-    let mut buf = [0u8; 64];
-    loop {
-        let n = unsafe { libc::read(fd, buf.as_mut_ptr() as *mut _, buf.len()) };
-        if n <= 0 {
-            break;
-        }
-    }
-}
-
-/// Block on `fd` for up to `timeout`, returning when either there's data
-/// to read or the timeout elapses. `None` for `timeout` means wait forever.
-pub fn poll_wake(fd: RawFd, timeout: Option<std::time::Duration>) {
-    let timeout_ms: libc::c_int = match timeout {
-        None => -1,
-        Some(d) => {
-            let ms = d.as_millis();
-            if ms > i32::MAX as u128 {
-                i32::MAX
-            } else {
-                ms as i32
-            }
-        }
-    };
-    let mut pfd = libc::pollfd {
-        fd,
-        events: libc::POLLIN,
-        revents: 0,
-    };
-    loop {
-        let r = unsafe { libc::poll(&mut pfd as *mut _, 1, timeout_ms) };
-        if r < 0 {
-            let e = unsafe { *libc::__errno_location() };
-            if e == libc::EINTR {
-                continue;
-            }
-        }
-        break;
-    }
-}

src/lib.rs

@@ -32,6 +32,7 @@ pub mod introspect;
 #[cfg(feature = "observer")]
 pub mod observer;
 pub mod runtime;
+pub(crate) mod park;
 pub(crate) mod raw_mutex;
 pub(crate) mod slot_state;
 pub(crate) mod sync_shim;

src/park.rs

@@ -0,0 +1,994 @@
+//! Scheduler park/wake coordination layer (RFC 018).
+//!
+//! Schedulers never touch an fd to sleep: they park on a per-thread
+//! [`Parker`] and are woken through an idle-mask protocol the runtime owns
+//! outright. IO backends (epoll today, io_uring later) are *producers*
+//! behind a two-call contract — make actors runnable, then wake — which is
+//! what makes backend selection tractable (RFC 018 §step-back).
+//!
+//! Three pieces, all in [`Coordinator`]:
+//!
+//! - **Parker** (one per scheduler): permit semantics, `std::thread::park`
+//!   shaped — an unpark delivered before the park sets a permit; the next
+//!   park consumes it and returns immediately. This single property closes
+//!   the check-then-park race. Linux: `futex(2)` `FUTEX_WAIT`/`FUTEX_WAKE`
+//!   (private) with a nanosecond-precision relative `timespec` — the
+//!   `as_millis` truncation defect of the retired wake pipe is
+//!   unrepresentable here. Loom / non-Linux: `Mutex<bool>` + `Condvar`
+//!   (the loom models run against this build).
+//! - **Idle mask**: an `AtomicU64` bitmask of parked scheduler ids
+//!   (construction asserts ≤ 64 schedulers). Park protocol: set own bit,
+//!   run the caller's mandatory post-publish re-check, then wait. A
+//!   producer that published work before observing our bit has left us
+//!   work the re-check finds; one that observes the bit wakes us.
+//!
+//!   The publish/re-check pair is a store-buffer (Dekker) shape. Two sound
+//!   resolutions coexist here, chosen per call-site cost profile: the
+//!   **fence handshake** on the hot producer path
+//!   ([`Coordinator::wake_one_if_idle`]: publish work; `fence(SeqCst)`;
+//!   one *Relaxed* mask load — pairing with the consumer's `fetch_or(bit)`;
+//!   `fence(SeqCst)`; re-check inside [`Coordinator::park`]), so the
+//!   pure-compute hot path (everyone busy, mask 0) never takes the shared
+//!   mask line exclusive — the RFC's "one relaxed load" fast path, made
+//!   sound; and the **same-location-RMW read** (`fetch_or(0)`, in
+//!   [`Coordinator::wake_one`] / [`Coordinator::idle_mask`]) for the rare
+//!   paths (chain rule — at most one per wake) where reading the latest
+//!   mask by modification-order coherence is worth an RMW. The loom models
+//!   drive the fence pattern end to end (a fence-less plain-load draft
+//!   would — and did — fail model 1/2 with a lost wake, as it must).
+//! - **Timekeeper**: at most one parked scheduler holds the timer
+//!   deadline (RFC 018 §timers) so a timer expiry wakes one scheduler,
+//!   not a herd. The role is an atomic `(holder id, armed deadline)`
+//!   pair; a timer insertion with an earlier deadline wakes the holder to
+//!   re-peek. Arm / insert-check MUST be serialized by the caller (the
+//!   timers mutex in the runtime) — the atomics exist so the *busy-path
+//!   due-check* (one Relaxed load, [`Coordinator::armed_deadline_nanos`])
+//!   and the wake stay lock-free. All races are biased over-wake: a
+//!   spurious permit costs one failed pop; a missed wake would cost a
+//!   stranded actor, and is unrepresentable under the serialization rule.
+//!
+//! Every wake here is *at most one* futex round-trip and wakes *exactly
+//! one* scheduler by construction (`wake_one` CASes a bit clear before
+//! unparking its owner) — there is no shared level-triggered anything
+//! left to herd on.
+//!
+//! Standalone until the runtime swap (RFC 018 commit 2): nothing outside
+//! tests constructs a [`Coordinator`] yet.
+
+use crate::sync_shim::{fence, AtomicU64, Ordering};
+use std::time::Instant;
+
+/// Sentinel for "no timekeeper" in the holder word.
+const NO_TIMEKEEPER: u64 = u64::MAX;
+/// Sentinel for "no armed deadline" in the deadline word. Also what the
+/// busy-path due-check compares against: `now_nanos < armed` is one branch.
+pub(crate) const NO_DEADLINE: u64 = u64::MAX;
+
+/// Outcome of a [`Coordinator::park`] call.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub(crate) enum ParkResult {
+    /// A permit was consumed (wake delivered before or during the park).
+    Woken,
+    /// The deadline passed with no wake. Only possible when a deadline was
+    /// supplied (and never under loom, which has no time — see `park`).
+    TimedOut,
+    /// The post-publish re-check found work; the thread never blocked.
+    /// A permit may still be pending (a racing `wake_one` picked us after
+    /// the bit was set) — it will surface as one spurious `Woken` on a
+    /// later park. Benign: over-wake by design.
+    WorkFound,
+}
+
+// ---------------------------------------------------------------------------
+// Parker — permit-semantics thread parking
+// ---------------------------------------------------------------------------
+
+/// Linux, non-loom: futex on a state word.
+///
+/// States: EMPTY (no permit, nobody waiting), PARKED (a thread is, or is
+/// about to be, in `futex_wait`), NOTIFIED (permit pending). The classic
+/// std-parker protocol: `unpark` swaps to NOTIFIED and futex-wakes iff it
+/// displaced PARKED; `park` CASes EMPTY→PARKED, waits, and consumes
+/// NOTIFIED on every exit path.
+#[cfg(all(target_os = "linux", not(loom)))]
+mod parker {
+    use std::sync::atomic::{AtomicU32, Ordering};
+    use std::time::Instant;
+
+    const EMPTY: u32 = 0;
+    const PARKED: u32 = 1;
+    const NOTIFIED: u32 = 2;
+
+    pub(super) struct Parker {
+        state: AtomicU32,
+    }
+
+    impl Parker {
+        pub(super) fn new() -> Self {
+            Self { state: AtomicU32::new(EMPTY) }
+        }
+
+        /// Returns `true` = woken (permit consumed), `false` = timed out.
+        pub(super) fn park(&self, deadline: Option<Instant>) -> bool {
+            // Fast path: consume a pending permit without blocking.
+            if self
+                .state
+                .compare_exchange(EMPTY, PARKED, Ordering::AcqRel, Ordering::Acquire)
+                .is_err()
+            {
+                // Only NOTIFIED can be here (one thread parks at a time).
+                self.state.store(EMPTY, Ordering::Release);
+                return true;
+            }
+            loop {
+                let timeout = match deadline {
+                    None => None,
+                    Some(d) => {
+                        let now = Instant::now();
+                        if d <= now {
+                            // Deadline passed: cancel the park. The swap
+                            // races a concurrent unpark — if it delivered
+                            // NOTIFIED first, report Woken (never lose a
+                            // permit).
+                            return self.state.swap(EMPTY, Ordering::AcqRel) == NOTIFIED;
+                        }
+                        Some(d - now)
+                    }
+                };
+                futex_wait(&self.state, PARKED, timeout);
+                if self
+                    .state
+                    .compare_exchange(NOTIFIED, EMPTY, Ordering::AcqRel, Ordering::Acquire)
+                    .is_ok()
+                {
+                    return true;
+                }
+                // Spurious wake or timeout with state still PARKED: loop —
+                // the deadline check at the top decides.
+            }
+        }
+
+        pub(super) fn unpark(&self) {
+            if self.state.swap(NOTIFIED, Ordering::AcqRel) == PARKED {
+                futex_wake(&self.state, 1);
+            }
+        }
+    }
+
+    /// `FUTEX_WAIT` with a *relative* nanosecond timeout (`CLOCK_MONOTONIC`
+    /// per futex(2) for relative waits). No millisecond conversion anywhere:
+    /// the timespec carries the full sub-ms remainder (RFC 018 kills the
+    /// `as_millis` truncation structurally).
+    fn futex_wait(word: &AtomicU32, expected: u32, timeout: Option<std::time::Duration>) {
+        let ts;
+        let ts_ptr: *const libc::timespec = match timeout {
+            Some(d) => {
+                ts = libc::timespec {
+                    tv_sec: d.as_secs() as libc::time_t,
+                    tv_nsec: d.subsec_nanos() as libc::c_long,
+                };
+                &ts
+            }
+            None => std::ptr::null(),
+        };
+        // Errors (EAGAIN: word changed; ETIMEDOUT; EINTR) all mean "return
+        // and let the caller's state machine decide" — deliberately ignored.
+        unsafe {
+            libc::syscall(
+                libc::SYS_futex,
+                word.as_ptr(),
+                libc::FUTEX_WAIT | libc::FUTEX_PRIVATE_FLAG,
+                expected,
+                ts_ptr,
+            );
+        }
+    }
+
+    fn futex_wake(word: &AtomicU32, n: u32) {
+        unsafe {
+            libc::syscall(
+                libc::SYS_futex,
+                word.as_ptr(),
+                libc::FUTEX_WAKE | libc::FUTEX_PRIVATE_FLAG,
+                n,
+            );
+        }
+    }
+}
+
+/// Loom / non-Linux: `Mutex<bool>` permit + `Condvar` — loom's own model
+/// of a parker, and the portable fallback. Under loom the deadline is
+/// ignored (loom has no time); the models exercise wake paths only.
+#[cfg(any(loom, not(target_os = "linux")))]
+mod parker {
+    use crate::sync_shim::{Condvar, Mutex};
+    use std::time::Instant;
+
+    pub(super) struct Parker {
+        permit: Mutex<bool>,
+        cv: Condvar,
+    }
+
+    impl Parker {
+        pub(super) fn new() -> Self {
+            Self { permit: Mutex::new(false), cv: Condvar::new() }
+        }
+
+        /// Returns `true` = woken (permit consumed), `false` = timed out.
+        pub(super) fn park(&self, deadline: Option<Instant>) -> bool {
+            let mut permit = match self.permit.lock() {
+                Ok(g) => g,
+                Err(_) => panic!("smarm: parker permit lock poisoned (core corrupt)"),
+            };
+            loop {
+                if *permit {
+                    *permit = false;
+                    return true;
+                }
+                #[cfg(loom)]
+                {
+                    // Loom has no clock: block until a wake. Models must
+                    // deliver one (a park nobody wakes is a real deadlock
+                    // and loom reports it as such).
+                    let _ = deadline;
+                    permit = match self.cv.wait(permit) {
+                        Ok(g) => g,
+                        Err(_) => panic!("smarm: parker cv poisoned (core corrupt)"),
+                    };
+                }
+                #[cfg(not(loom))]
+                {
+                    match deadline {
+                        None => {
+                            permit = match self.cv.wait(permit) {
+                                Ok(g) => g,
+                                Err(_) => panic!("smarm: parker cv poisoned (core corrupt)"),
+                            };
+                        }
+                        Some(d) => {
+                            let now = Instant::now();
+                            if d <= now {
+                                return false;
+                            }
+                            permit = match self.cv.wait_timeout(permit, d - now) {
+                                Ok((g, _)) => g,
+                                Err(_) => panic!("smarm: parker cv poisoned (core corrupt)"),
+                            };
+                        }
+                    }
+                }
+            }
+        }
+
+        pub(super) fn unpark(&self) {
+            let mut permit = match self.permit.lock() {
+                Ok(g) => g,
+                Err(_) => panic!("smarm: parker permit lock poisoned (core corrupt)"),
+            };
+            *permit = true;
+            self.cv.notify_one();
+        }
+    }
+}
+
+use parker::Parker;
+
+// ---------------------------------------------------------------------------
+// Coordinator — idle mask + wake protocol + timekeeper
+// ---------------------------------------------------------------------------
+
+pub(crate) struct Coordinator {
+    parkers: Box<[Parker]>,
+    /// Bit `i` set = scheduler `i` is parked or committed to parking (set
+    /// before the re-check; cleared by `wake_one`'s CAS or by the parker
+    /// itself on return). AcqRel same-location-RMW handshake — see module
+    /// docs (no SeqCst needed: every producer-side read is an RMW).
+    idle: AtomicU64,
+    /// Timekeeper holder id, or `NO_TIMEKEEPER`. Written under the
+    /// caller's timer serialization (arm/disarm/insert-check); read
+    /// lock-free by the insert wake path.
+    tk_holder: AtomicU64,
+    /// Armed deadline as nanos since `origin`, or `NO_DEADLINE`. Written
+    /// only by the timekeeper arm/disarm protocol.
+    tk_armed: AtomicU64,
+    /// Earliest KNOWN timer deadline (nanos since `origin`), or
+    /// `NO_DEADLINE` — independent of whether any scheduler is parked,
+    /// which is what the timekeeper's `tk_armed` cannot give: under
+    /// saturation nobody parks and nobody arms, yet due timers must still
+    /// fire (ratified design point (a)). Maintained under the caller's
+    /// timers mutex (`note_deadline` on insert, `refresh_deadline` after a
+    /// pop/peek); read lock-free by the busy-path due-check.
+    next_deadline: AtomicU64,
+    /// Time origin for the nanos encoding.
+    origin: Instant,
+}
+
+impl Coordinator {
+    pub(crate) fn new(schedulers: usize) -> Self {
+        assert!(
+            (1..=64).contains(&schedulers),
+            "smarm: scheduler count must be 1..=64 (idle mask is one u64); got {schedulers}"
+        );
+        Self {
+            parkers: (0..schedulers).map(|_| Parker::new()).collect(),
+            idle: AtomicU64::new(0),
+            tk_holder: AtomicU64::new(NO_TIMEKEEPER),
+            tk_armed: AtomicU64::new(NO_DEADLINE),
+            next_deadline: AtomicU64::new(NO_DEADLINE),
+            origin: Instant::now(),
+        }
+    }
+
+    /// Encode a deadline for the armed snapshot / busy-path compare.
+    /// Saturating: a deadline at-or-before `origin` encodes as 0 (always
+    /// due), one beyond ~584 years as `NO_DEADLINE - 1`.
+    pub(crate) fn deadline_nanos(&self, deadline: Instant) -> u64 {
+        let nanos = deadline
+            .checked_duration_since(self.origin)
+            .map(|d| d.as_nanos())
+            .unwrap_or(0);
+        if nanos >= NO_DEADLINE as u128 {
+            NO_DEADLINE - 1
+        } else {
+            nanos as u64
+        }
+    }
+
+    /// Park scheduler `id` until a wake, the deadline, or a positive
+    /// re-check. Protocol: (1) publish own idle bit (SeqCst), (2) run
+    /// `recheck` — it MUST re-read the work source with an ordering that
+    /// pairs with the producer's publish (SeqCst load, or take the mutex
+    /// the producer publishes under); a `true` aborts the park, (3) block.
+    pub(crate) fn park(
+        &self,
+        id: usize,
+        deadline: Option<Instant>,
+        recheck: impl FnOnce() -> bool,
+    ) -> ParkResult {
+        debug_assert!(id < self.parkers.len(), "park: scheduler id out of range");
+        let bit = 1u64 << id;
+        // (1) publish. AcqRel RMW: the acquire half is the handshake — if
+        // this lands after a producer's mask RMW in modification order, we
+        // read-from it and the producer's earlier work publication is
+        // visible to the re-check below (see module docs).
+        let prev = self.idle.fetch_or(bit, Ordering::AcqRel);
+        debug_assert_eq!(prev & bit, 0, "park: idle bit already set for this id");
+        // Fence half of the producer handshake (see `wake_one_if_idle`):
+        // orders our bit-publish before the re-check's loads, so it pairs
+        // with the producer's publish→fence→mask-load — at least one side
+        // must see the other's store, whichever queue backend is in play.
+        fence(Ordering::SeqCst);
+        // (2) the mandatory post-publish re-check.
+        if recheck() {
+            self.idle.fetch_and(!bit, Ordering::AcqRel);
+            return ParkResult::WorkFound;
+        }
+        // (3) block. The permit closes the window between the re-check and
+        // the futex wait: a wake_one that picked us in that window has
+        // already CASed our bit clear and set the permit.
+        let woken = self.parkers[id].park(deadline);
+        // Clear own bit — a no-op when a waker already CASed it clear.
+        self.idle.fetch_and(!bit, Ordering::AcqRel);
+        if woken {
+            ParkResult::Woken
+        } else {
+            ParkResult::TimedOut
+        }
+    }
+
+    /// Wake exactly one parked scheduler, if any: pick the highest set idle
+    /// bit (LIFO-ish — warmest cache), CAS it clear, deliver a permit.
+    /// Empty mask = no-op (everyone is awake and will find work by
+    /// popping). Returns whether a scheduler was woken.
+    pub(crate) fn wake_one(&self) -> bool {
+        // RMW read, not a load: reads the latest mask by modification-order
+        // coherence, closing the Dekker race with a parking consumer (see
+        // module docs). The release side of the RMW is what a later-parking
+        // consumer's fetch_or acquires to make its re-check sound.
+        let mut mask = self.idle.fetch_or(0, Ordering::AcqRel);
+        loop {
+            if mask == 0 {
+                return false;
+            }
+            let id = 63 - mask.leading_zeros() as usize; // highest set bit
+            let bit = 1u64 << id;
+            // The CAS is the exactly-one guarantee: whoever clears the bit
+            // owns the wake; a racing wake_one retries on the observed value
+            // (coherence: a failed CAS can never read older than `mask`).
+            match self.idle.compare_exchange(
+                mask,
+                mask & !bit,
+                Ordering::AcqRel,
+                Ordering::Acquire,
+            ) {
+                Ok(_) => {
+                    self.parkers[id].unpark();
+                    return true;
+                }
+                Err(m) => mask = m,
+            }
+        }
+    }
+
+    /// The producer-side wake tail (`enqueue`'s fast path, RFC 018 "enqueue
+    /// wakes"). The caller has just published work (queue push); we fence,
+    /// then read the mask with ONE Relaxed load — 0 means every scheduler
+    /// is awake and the pure-compute hot path pays no RMW on the shared
+    /// mask line. Soundness is the fence handshake (module docs): our
+    /// fence orders the caller's push before the mask load; the consumer's
+    /// fence (in `park`) orders its bit-publish before its re-check — at
+    /// least one side must observe the other's store, so a consumer we
+    /// miss here is a consumer whose re-check finds the caller's work.
+    pub(crate) fn wake_one_if_idle(&self) -> bool {
+        fence(Ordering::SeqCst);
+        if self.idle.load(Ordering::Relaxed) == 0 {
+            return false;
+        }
+        self.wake_one()
+    }
+
+    /// Terminal wake (replaces the AllDone wake-pipe byte): clear the mask
+    /// and deliver a permit to *every* parker, parked or not. A permit set
+    /// on a busy scheduler costs one spurious park return — nothing at the
+    /// terminal boundary. Idempotent.
+    pub(crate) fn wake_all(&self) {
+        self.idle.store(0, Ordering::Release);
+        for p in self.parkers.iter() {
+            p.unpark();
+        }
+    }
+
+    /// Latest idle mask (RMW read — same handshake as `wake_one`). A
+    /// test-only observer: production expresses the chain rule through
+    /// `wake_one_if_idle` (fence + Relaxed load), not a mask read.
+    #[cfg(test)]
+    pub(crate) fn idle_mask(&self) -> u64 {
+        self.idle.fetch_or(0, Ordering::AcqRel)
+    }
+
+    // ----- timekeeper -----
+
+    /// Try to take the timekeeper role for scheduler `id` with `deadline`.
+    /// MUST be called under the caller's timer serialization (the timers
+    /// mutex), with `deadline` the heap minimum peeked under that same
+    /// hold — this is what makes the insert-check race-free. Returns
+    /// whether the role was taken (false = someone else holds it; park
+    /// with no deadline).
+    pub(crate) fn try_arm_timer(&self, id: usize, deadline: Instant) -> bool {
+        debug_assert!(id < self.parkers.len(), "try_arm_timer: id out of range");
+        if self
+            .tk_holder
+            .compare_exchange(NO_TIMEKEEPER, id as u64, Ordering::SeqCst, Ordering::SeqCst)
+            .is_err()
+        {
+            return false;
+        }
+        // Holder-then-deadline order: an insert-check that sees the holder
+        // with the deadline still NO_DEADLINE compares `new < MAX` = true
+        // and over-wakes — the benign direction. (Under the mandated timer
+        // serialization this interleaving cannot occur anyway.)
+        self.tk_armed.store(self.deadline_nanos(deadline), Ordering::SeqCst);
+        true
+    }
+
+    /// Release the timekeeper role (the holder, on wake, before it
+    /// re-peeks/fires). Callable without the timer serialization: a
+    /// racing insert may wake a no-longer-holder — over-wake, benign.
+    pub(crate) fn disarm_timer(&self, id: usize) {
+        debug_assert_eq!(
+            self.tk_holder.load(Ordering::SeqCst),
+            id as u64,
+            "disarm_timer by a non-holder"
+        );
+        // Deadline first: a concurrent insert-check then sees NO_DEADLINE
+        // and skips the (now pointless) wake instead of waking a stale
+        // holder id. Either order is correct; this one wastes less.
+        self.tk_armed.store(NO_DEADLINE, Ordering::SeqCst);
+        self.tk_holder.store(NO_TIMEKEEPER, Ordering::SeqCst);
+    }
+
+    /// Insert-side re-arm check: if `deadline` is earlier than the armed
+    /// snapshot, wake the timekeeper to re-peek. MUST be called under the
+    /// same timer serialization as `try_arm_timer` (see there).
+    pub(crate) fn timer_inserted(&self, deadline: Instant) {
+        if self.deadline_nanos(deadline) < self.tk_armed.load(Ordering::SeqCst) {
+            let holder = self.tk_holder.load(Ordering::SeqCst);
+            if holder != NO_TIMEKEEPER {
+                // Direct unpark, not wake_one: the wake targets the
+                // timekeeper specifically (it must re-peek the heap). Its
+                // idle bit stays set until it returns from park — a
+                // concurrent wake_one may pick it too; over-wake, benign.
+                self.parkers[holder as usize].unpark();
+            }
+        }
+    }
+
+    /// The armed-deadline snapshot (nanos since origin; `NO_DEADLINE` =
+    /// none). Test-only introspection on the timekeeper's armed value; the
+    /// busy-path due-check reads `next_deadline`, not this.
+    #[cfg(test)]
+    pub(crate) fn armed_deadline_nanos(&self) -> u64 {
+        self.tk_armed.load(Ordering::Relaxed)
+    }
+
+    // ----- earliest-deadline snapshot (busy-path due-check) -----
+
+    /// Record a newly inserted timer deadline. MUST be called under the
+    /// timers mutex (same serialization rule as `try_arm_timer`), which is
+    /// why plain compare+store suffices for the min-maintenance. Also runs
+    /// the timekeeper re-arm check (`timer_inserted`) — one call site for
+    /// both consequences of an insert.
+    pub(crate) fn note_deadline(&self, deadline: Instant) {
+        let n = self.deadline_nanos(deadline);
+        if n < self.next_deadline.load(Ordering::Relaxed) {
+            self.next_deadline.store(n, Ordering::Release);
+        }
+        self.timer_inserted(deadline);
+    }
+
+    /// Re-anchor the snapshot to the heap minimum (`None` = heap empty)
+    /// after a `pop_due` / `clear`. MUST be called under the timers mutex.
+    pub(crate) fn refresh_deadline(&self, next: Option<Instant>) {
+        let n = next.map_or(NO_DEADLINE, |d| self.deadline_nanos(d));
+        self.next_deadline.store(n, Ordering::Release);
+    }
+
+    /// Busy-path due-check: is the earliest known deadline at or past
+    /// `now`? One Relaxed load when no deadline is armed — the clock is
+    /// read only when a timer actually exists (matching the old drain
+    /// phase's is_empty guard), so the pure-compute hot path pays a load
+    /// and a branch.
+    pub(crate) fn deadline_due(&self) -> bool {
+        let n = self.next_deadline.load(Ordering::Relaxed);
+        n != NO_DEADLINE && self.deadline_nanos(Instant::now()) >= n
+    }
+
+    /// The earliest-deadline snapshot as an `Instant` (`None` = no timer
+    /// pending). Test-only: the idle path arms the timekeeper from the
+    /// timer heap's own `peek_deadline` under the timers mutex.
+    #[cfg(test)]
+    pub(crate) fn next_deadline_instant(&self) -> Option<Instant> {
+        let n = self.next_deadline.load(Ordering::Acquire);
+        if n == NO_DEADLINE {
+            None
+        } else {
+            self.origin.checked_add(std::time::Duration::from_nanos(n))
+        }
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Unit tests (std build)
+// ---------------------------------------------------------------------------
+
+#[cfg(all(test, not(loom)))]
+mod tests {
+    use super::*;
+    use std::sync::atomic::{AtomicUsize, Ordering as O};
+    use std::sync::Arc;
+    use std::time::{Duration, Instant};
+
+    #[test]
+    fn permit_before_park_returns_immediately() {
+        let c = Coordinator::new(1);
+        // Deliver the wake first (nobody parked: wake_one no-ops on the
+        // mask, so use the timekeeper-direct path? No — permit semantics
+        // are the parker's own; exercise via wake_all which permits all).
+        c.wake_all();
+        let t0 = Instant::now();
+        let r = c.park(0, None, || false);
+        assert_eq!(r, ParkResult::Woken);
+        assert!(t0.elapsed() < Duration::from_millis(100), "park blocked despite permit");
+    }
+
+    #[test]
+    fn submillisecond_deadline_is_honored() {
+        // Regression for the retired as_millis truncation: a 500µs deadline
+        // must neither busy-return instantly forever nor round to 0/∞.
+        let c = Coordinator::new(1);
+        let t0 = Instant::now();
+        let r = c.park(0, Some(t0 + Duration::from_micros(500)), || false);
+        let dt = t0.elapsed();
+        assert_eq!(r, ParkResult::TimedOut);
+        assert!(dt >= Duration::from_micros(400), "woke too early: {dt:?}");
+        assert!(dt < Duration::from_millis(50), "overslept: {dt:?}");
+    }
+
+    #[test]
+    fn recheck_true_aborts_park_and_clears_bit() {
+        let c = Coordinator::new(2);
+        let r = c.park(1, None, || true);
+        assert_eq!(r, ParkResult::WorkFound);
+        assert_eq!(c.idle_mask(), 0, "bit not cleared after WorkFound");
+    }
+
+    #[test]
+    fn recheck_observes_own_bit_published() {
+        let c = Coordinator::new(2);
+        let seen = std::cell::Cell::new(0u64);
+        let r = c.park(1, None, || {
+            seen.set(c.idle_mask());
+            true
+        });
+        assert_eq!(r, ParkResult::WorkFound);
+        assert_eq!(seen.get() & 0b10, 0b10, "bit not published before re-check");
+    }
+
+    #[test]
+    fn wake_one_wakes_exactly_one_of_n() {
+        const N: usize = 4;
+        let c = Arc::new(Coordinator::new(N));
+        let woken = Arc::new(AtomicUsize::new(0));
+        let mut ts = Vec::new();
+        for id in 0..N {
+            let c = c.clone();
+            let woken = woken.clone();
+            ts.push(std::thread::spawn(move || {
+                let r = c.park(id, None, || false);
+                assert_eq!(r, ParkResult::Woken);
+                woken.fetch_add(1, O::SeqCst);
+            }));
+        }
+        // Wait until all four are published idle.
+        let t0 = Instant::now();
+        while c.idle_mask().count_ones() != N as u32 {
+            assert!(t0.elapsed() < Duration::from_secs(5), "threads never parked");
+            std::thread::yield_now();
+        }
+        assert!(c.wake_one());
+        // Exactly one wakes; give the others a beat to (incorrectly) wake.
+        let t0 = Instant::now();
+        while woken.load(O::SeqCst) == 0 {
+            assert!(t0.elapsed() < Duration::from_secs(5), "wake_one woke nobody");
+            std::thread::yield_now();
+        }
+        std::thread::sleep(Duration::from_millis(100));
+        assert_eq!(woken.load(O::SeqCst), 1, "wake_one woke more than one");
+        assert_eq!(c.idle_mask().count_ones(), (N - 1) as u32);
+        c.wake_all();
+        for t in ts {
+            match t.join() {
+                Ok(()) => {}
+                Err(p) => std::panic::resume_unwind(p),
+            }
+        }
+        assert_eq!(woken.load(O::SeqCst), N);
+        assert_eq!(c.idle_mask(), 0);
+    }
+
+    #[test]
+    fn wake_one_prefers_highest_bit() {
+        let c = Arc::new(Coordinator::new(3));
+        let woken_id = Arc::new(AtomicUsize::new(usize::MAX));
+        let mut ts = Vec::new();
+        for id in 0..3 {
+            let c = c.clone();
+            let woken_id = woken_id.clone();
+            ts.push(std::thread::spawn(move || {
+                if c.park(id, None, || false) == ParkResult::Woken {
+                    let _ = woken_id.compare_exchange(usize::MAX, id, O::SeqCst, O::SeqCst);
+                }
+            }));
+        }
+        let t0 = Instant::now();
+        while c.idle_mask() != 0b111 {
+            assert!(t0.elapsed() < Duration::from_secs(5));
+            std::thread::yield_now();
+        }
+        assert!(c.wake_one());
+        let t0 = Instant::now();
+        while woken_id.load(O::SeqCst) == usize::MAX {
+            assert!(t0.elapsed() < Duration::from_secs(5));
+            std::thread::yield_now();
+        }
+        assert_eq!(woken_id.load(O::SeqCst), 2, "LIFO-ish: highest bit first");
+        c.wake_all();
+        for t in ts {
+            let _ = t.join();
+        }
+    }
+
+    #[test]
+    fn wake_one_on_empty_mask_is_noop() {
+        let c = Coordinator::new(2);
+        assert!(!c.wake_one());
+        assert_eq!(c.idle_mask(), 0);
+    }
+
+    #[test]
+    fn timekeeper_arm_is_exclusive_and_snapshot_readable() {
+        let c = Coordinator::new(2);
+        let d2 = Instant::now() + Duration::from_secs(10);
+        let d1 = Instant::now() + Duration::from_secs(1);
+        assert_eq!(c.armed_deadline_nanos(), NO_DEADLINE);
+        assert!(c.try_arm_timer(0, d2));
+        assert!(!c.try_arm_timer(1, d1), "second arm must fail while held");
+        assert_eq!(c.armed_deadline_nanos(), c.deadline_nanos(d2));
+        c.disarm_timer(0);
+        assert_eq!(c.armed_deadline_nanos(), NO_DEADLINE);
+        assert!(c.try_arm_timer(1, d1), "role must be re-takeable after disarm");
+        c.disarm_timer(1);
+    }
+
+    #[test]
+    fn earlier_insert_wakes_timekeeper() {
+        let c = Coordinator::new(2);
+        let far = Instant::now() + Duration::from_secs(60);
+        let near = Instant::now() + Duration::from_millis(1);
+        assert!(c.try_arm_timer(0, far));
+        // Holder not yet parked: the wake must land as a permit.
+        c.timer_inserted(near);
+        let t0 = Instant::now();
+        let r = c.park(0, Some(far), || false);
+        assert_eq!(r, ParkResult::Woken, "re-arm wake lost");
+        assert!(t0.elapsed() < Duration::from_secs(5), "slept toward the stale deadline");
+        c.disarm_timer(0);
+    }
+
+    #[test]
+    fn later_insert_does_not_wake_timekeeper() {
+        let c = Coordinator::new(2);
+        let near = Instant::now() + Duration::from_millis(20);
+        let far = Instant::now() + Duration::from_secs(60);
+        assert!(c.try_arm_timer(0, near));
+        c.timer_inserted(far); // later than armed: no wake
+        let r = c.park(0, Some(near), || false);
+        assert_eq!(r, ParkResult::TimedOut, "spurious wake for a later insert");
+        c.disarm_timer(0);
+    }
+
+    #[test]
+    #[should_panic(expected = "1..=64")]
+    fn more_than_64_schedulers_asserts() {
+        let _ = Coordinator::new(65);
+    }
+
+    #[test]
+    fn wake_one_if_idle_noop_on_empty_and_wakes_on_parked() {
+        let c = Arc::new(Coordinator::new(1));
+        assert!(!c.wake_one_if_idle(), "empty mask must be a no-op");
+        let c2 = c.clone();
+        let t = std::thread::spawn(move || {
+            assert_eq!(c2.park(0, None, || false), ParkResult::Woken);
+        });
+        let t0 = Instant::now();
+        while c.idle_mask() == 0 {
+            assert!(t0.elapsed() < Duration::from_secs(5), "never parked");
+            std::thread::yield_now();
+        }
+        assert!(c.wake_one_if_idle());
+        match t.join() {
+            Ok(()) => {}
+            Err(p) => std::panic::resume_unwind(p),
+        }
+    }
+
+    #[test]
+    fn deadline_snapshot_min_maintenance_and_due_check() {
+        let c = Coordinator::new(1);
+        assert!(!c.deadline_due(), "no deadline: never due");
+        assert_eq!(c.next_deadline_instant(), None);
+        let far = Instant::now() + Duration::from_secs(60);
+        let near = Instant::now() + Duration::from_millis(1);
+        c.note_deadline(far);
+        assert!(!c.deadline_due());
+        c.note_deadline(near); // min wins
+        assert!(c.next_deadline_instant().is_some_and(|d| d <= near));
+        c.note_deadline(far); // later insert must NOT raise the snapshot
+        assert!(c.next_deadline_instant().is_some_and(|d| d <= near));
+        std::thread::sleep(Duration::from_millis(2));
+        assert!(c.deadline_due(), "past deadline not reported due");
+        c.refresh_deadline(Some(far));
+        assert!(!c.deadline_due(), "refresh did not re-anchor");
+        c.refresh_deadline(None);
+        assert_eq!(c.next_deadline_instant(), None);
+        // A deadline at/before origin encodes as 0: always due.
+        c.note_deadline(Instant::now() - Duration::from_secs(1));
+        assert!(c.deadline_due());
+    }
+
+    #[test]
+    fn note_deadline_earlier_wakes_timekeeper_via_snapshot_path() {
+        // note_deadline must carry the timer_inserted re-arm wake too.
+        let c = Coordinator::new(2);
+        let far = Instant::now() + Duration::from_secs(60);
+        let near = Instant::now() + Duration::from_millis(1);
+        assert!(c.try_arm_timer(0, far));
+        c.note_deadline(near);
+        let r = c.park(0, Some(far), || false);
+        assert_eq!(r, ParkResult::Woken, "re-arm wake lost through note_deadline");
+        c.disarm_timer(0);
+    }
+}
+
+// ---------------------------------------------------------------------------
+// loom models — RUSTFLAGS="--cfg loom" cargo test --lib --release park
+// ---------------------------------------------------------------------------
+
+#[cfg(all(test, loom))]
+mod loom_tests {
+    use super::*;
+    use loom::sync::atomic::{AtomicU64 as LAtomicU64, Ordering as O};
+    use loom::sync::{Arc, Mutex as LMutex};
+    use loom::thread;
+    use std::time::{Duration, Instant};
+
+    /// RFC 018 loom model 1 — no lost wake.
+    /// producer{publish item; wake_one} ∥ consumer{set bit; re-check; park}:
+    /// the consumer always observes the item or a permit; it can never
+    /// sleep past a published item (loom's deadlock detector is the
+    /// assertion — a consumer parked forever fails the model).
+    #[test]
+    fn no_lost_wake() {
+        loom::model(|| {
+            let c = Arc::new(Coordinator::new(1));
+            let item = Arc::new(LAtomicU64::new(0));
+
+            let prod = {
+                let c = c.clone();
+                let item = item.clone();
+                thread::spawn(move || {
+                    // The enqueue shape: publish, then the fenced fast-path
+                    // wake tail (this is what the runtime's enqueue calls).
+                    item.store(1, O::SeqCst);
+                    c.wake_one_if_idle();
+                })
+            };
+
+            // Consumer: loop until the item is popped. Guarded load+CAS,
+            // not a blind swap — a swap writes 0 even when empty, and
+            // coherence allows that write to land after the producer's
+            // store in modification order, destroying the item (a model
+            // bug loom caught in an earlier draft of this test).
+            loop {
+                if item.load(O::SeqCst) == 1
+                    && item.compare_exchange(1, 0, O::SeqCst, O::SeqCst).is_ok()
+                {
+                    break;
+                }
+                let _ = c.park(0, None, || item.load(O::SeqCst) == 1);
+            }
+            prod.join().unwrap();
+        });
+    }
+
+    /// RFC 018 loom model 2 — chain propagation.
+    /// Two items, two sleepers, ONE producer wake: the chain rule (a woken
+    /// consumer that sees surplus work and a non-empty mask wakes again)
+    /// must get both items consumed with no further producer action.
+    #[test]
+    fn chain_propagation() {
+        loom::model(|| {
+            let c = Arc::new(Coordinator::new(2));
+            let items = Arc::new(LAtomicU64::new(0));
+            let consumed = Arc::new(LAtomicU64::new(0));
+
+            let mut hs = Vec::new();
+            for id in 0..2usize {
+                let c = c.clone();
+                let items = items.clone();
+                let consumed = consumed.clone();
+                hs.push(thread::spawn(move || loop {
+                    if consumed.load(O::SeqCst) == 2 {
+                        c.wake_all(); // release a sibling still parked
+                        break;
+                    }
+                    let cur = items.load(O::SeqCst);
+                    if cur > 0
+                        && items
+                            .compare_exchange(cur, cur - 1, O::SeqCst, O::SeqCst)
+                            .is_ok()
+                    {
+                        consumed.fetch_add(1, O::SeqCst);
+                        // THE CHAIN RULE, exactly as production expresses it
+                        // (runtime.rs schedule_loop): surplus ⇒ the fenced
+                        // fast-path wake. Models the Relaxed-load chain path,
+                        // not just the RMW one.
+                        if items.load(O::SeqCst) > 0 {
+                            c.wake_one_if_idle();
+                        }
+                        continue;
+                    }
+                    let _ = c.park(id, None, || {
+                        items.load(O::SeqCst) > 0 || consumed.load(O::SeqCst) == 2
+                    });
+                }));
+            }
+
+            // Producer (main): two items, ONE wake, via the enqueue-shaped
+            // fenced fast path.
+            items.store(2, O::SeqCst);
+            c.wake_one_if_idle();
+
+            for h in hs {
+                h.join().unwrap();
+            }
+            assert_eq!(consumed.load(O::SeqCst), 2);
+            assert_eq!(items.load(O::SeqCst), 0);
+        });
+    }
+
+    /// RFC 018 loom model 3 — timekeeper handoff.
+    /// An earlier-deadline insert racing the parking timekeeper: the
+    /// earlier deadline is always honored — either the timekeeper armed it
+    /// directly (insert landed first under the timers lock) or the insert
+    /// wakes the timekeeper to re-peek. A timekeeper sleeping toward the
+    /// stale later deadline would deadlock the model (loom has no time).
+    #[test]
+    fn timekeeper_handoff() {
+        loom::model(|| {
+            let origin = Instant::now();
+            let d_far = origin + Duration::from_secs(60);
+            let d_near = origin + Duration::from_secs(1);
+
+            let c = Arc::new(Coordinator::new(1));
+            // The timers-mutex stand-in: heap min under a lock.
+            let heap_min = Arc::new(LMutex::new(d_far));
+
+            let tk = {
+                let c = c.clone();
+                let heap_min = heap_min.clone();
+                thread::spawn(move || {
+                    // Peek + arm under the lock (the serialization rule).
+                    let armed = {
+                        let g = heap_min.lock().unwrap();
+                        let min = *g;
+                        assert!(c.try_arm_timer(0, min));
+                        min
+                    };
+                    if armed == d_far {
+                        // Insert hadn't landed: it MUST wake us. Parking
+                        // toward d_far with no wake = model deadlock.
+                        let r = c.park(0, Some(armed), || false);
+                        assert_eq!(r, ParkResult::Woken, "re-arm wake lost");
+                    }
+                    // Woken (or armed the near deadline directly): re-peek.
+                    c.disarm_timer(0);
+                    let g = heap_min.lock().unwrap();
+                    assert_eq!(*g, d_near, "earlier deadline not visible on re-peek");
+                })
+            };
+
+            // Inserter (main): publish the earlier deadline under the lock,
+            // then the insert-check.
+            {
+                let mut g = heap_min.lock().unwrap();
+                *g = d_near;
+                c.timer_inserted(d_near);
+            }
+
+            tk.join().unwrap();
+        });
+    }
+
+    /// RFC 018 loom model 4 — termination.
+    /// The AllDone verdict: producer flips done and wake_all()s; consumers
+    /// must never park forever past done (park's re-check + wake_all's
+    /// permits close every interleaving; a stuck consumer = loom deadlock).
+    #[test]
+    fn termination_no_park_past_done() {
+        loom::model(|| {
+            let c = Arc::new(Coordinator::new(2));
+            let done = Arc::new(LAtomicU64::new(0));
+
+            let mut hs = Vec::new();
+            for id in 0..2usize {
+                let c = c.clone();
+                let done = done.clone();
+                hs.push(thread::spawn(move || loop {
+                    if done.load(O::SeqCst) == 1 {
+                        break;
+                    }
+                    let _ = c.park(id, None, || done.load(O::SeqCst) == 1);
+                }));
+            }
+
+            done.store(1, O::SeqCst);
+            c.wake_all();
+
+            for h in hs {
+                h.join().unwrap();
+            }
+        });
+    }
+}

src/runtime.rs

@@ -86,21 +86,28 @@
 //! # Termination (counter-based)
 //!
 //! The old all-clear scanned the slot table under the big lock. Now:
-//! exit when `io_out == 0` (read *before* the queue lock, phase-1 ordering)
-//! and, under the queue lock, the queue is empty and `live_actors == 0`.
-//! `live_actors` is incremented in `spawn` before the enqueue and decremented
-//! at the very END of `finalize_actor`, strictly after every wakeup that
-//! finalize produces has been enqueued. The soundness crux: any enqueue
-//! targets a live (not-yet-finalized) actor, so `live == 0` implies no wakeup
-//! can still be in flight; combined with "spawner is itself live", observing
-//! `(queue empty, live == 0)` under the queue lock means no work can ever
-//! appear again.
+//! exit when `io_outstanding + io_fd_waiters == 0` (two Relaxed/Acquire
+//! atomic loads, read *before* the queue pop) and, under the queue lock,
+//! the queue is empty and `live_actors == 0`. `live_actors` is incremented
+//! in `spawn` before the enqueue and decremented at the very END of
+//! `finalize_actor`, strictly after every wakeup that finalize produces has
+//! been enqueued. The soundness crux: any enqueue targets a live
+//! (not-yet-finalized) actor, so `live == 0` implies no wakeup can still be
+//! in flight; combined with "spawner is itself live", observing
+//! `(queue empty, live == 0)` means no work can ever appear again.
 //!
-//! # Timer / IO drain (try-lock, one-winner)
+//! # Scheduler park/wake (RFC 018)
 //!
-//! Unchanged from phase 1: one winner per round drains due timers and IO
-//! completions from their own mutexes; wakeups go through the unpark
-//! protocol like everyone else's.
+//! Schedulers sleep on per-thread futex parkers via the coordination layer
+//! (`park.rs`), NOT on a shared wake pipe. IO backends are producers behind
+//! a two-call contract — make the actor runnable (`unpark_at`), whose
+//! `enqueue` tail wakes exactly one parked scheduler. The blocking pool and
+//! epoll thread each route their own completions (driver-enqueues); there
+//! is no shared completion queue, no drain lock, no one-winner drain phase.
+//! Timers fire two ways: a busy-path due-check every loop iteration (one
+//! Relaxed load of the earliest-deadline snapshot when no timer is armed),
+//! and the timekeeper — at most one parked scheduler holds the timer
+//! deadline, so an expiry wakes one scheduler, not a herd.
 
 use crate::actor::{
     clear_current_pid, is_actor_done, reset_actor_done, set_current_actor_box,
@@ -750,8 +757,21 @@ pub(crate) struct RuntimeInner {
     pub(crate) io: Mutex<Option<IoThread>>,
     /// Monotonic `MonitorId` source. Never reused.
     pub(crate) next_monitor_id: AtomicU64,
-    /// Try-lock: exactly one scheduler thread drains timers/IO per iteration.
-    drain_lock: Mutex<()>,
+    /// RFC 018: the scheduler coordination layer — per-scheduler parkers,
+    /// idle mask, wake protocol, timekeeper role, earliest-deadline
+    /// snapshot. Arc'd because `Timers` shares it (insert-side deadline
+    /// notes run under the timers mutex).
+    pub(crate) coord: Arc<crate::park::Coordinator>,
+    /// `block_on_io` requests in flight. Incremented by the submitter
+    /// BEFORE submit (underflow-proof), decremented by the pool thread on
+    /// completion. Read lock-free by the idle path's termination verdict —
+    /// the per-pop `io.lock` of the drain era is gone.
+    pub(crate) io_outstanding: AtomicU32,
+    /// Parked fd waiters. Incremented by the registrar BEFORE
+    /// `epoll_register` (rolled back on error), decremented by whoever
+    /// consumes the registration (epoll thread on readiness, canceller on
+    /// an unwound wait). Same lock-free verdict read as `io_outstanding`.
+    pub(crate) io_fd_waiters: AtomicU32,
     /// Per-thread stats, indexed by scheduler thread slot (0..N).
     pub(crate) stats: Vec<SchedulerStats>,
     /// Global counters for RFC 000 primitives.
@@ -802,6 +822,12 @@ impl RuntimeInner {
         let slots: Box<[Slot]> = (0..max_actors).map(|_| Slot::vacant()).collect();
         // Low indices on top of the stack so early spawns get low pids.
         let free: Vec<u32> = (0..max_actors as u32).rev().collect();
+        // RFC 018: the coordination layer (asserts thread_count <= 64), and
+        // the timers' hook into it — every insert under the timers mutex
+        // notes its deadline (busy-path snapshot + timekeeper re-arm).
+        let coord = Arc::new(crate::park::Coordinator::new(thread_count));
+        let mut timers = Timers::new();
+        timers.attach_coordinator(coord.clone());
         Arc::new(Self {
             run_queue: crate::run_queue::RunQueue::new(thread_count, max_actors),
             slots,
@@ -810,10 +836,12 @@ impl RuntimeInner {
             root_bits: AtomicU64::new(u64::MAX),
             root_exited: AtomicBool::new(false),
             root_swept: AtomicBool::new(false),
-            timers: Mutex::new(Timers::new()),
+            timers: Mutex::new(timers),
             io: Mutex::new(None),
             next_monitor_id: AtomicU64::new(0),
-            drain_lock: Mutex::new(()),
+            coord,
+            io_outstanding: AtomicU32::new(0),
+            io_fd_waiters: AtomicU32::new(0),
             stats,
             io_parked: AtomicU32::new(0),
             sleeping: AtomicU32::new(0),
@@ -874,6 +902,13 @@ impl RuntimeInner {
         );
         self.run_queue.push(pid);
         crate::te!(crate::trace::Event::Enqueue(pid));
+        // RFC 018 enqueue wake (fixes the silent enqueue): if a scheduler
+        // is parked, wake exactly one. The fast path when everyone is busy
+        // is a fence + one Relaxed load of an unmodified line — the
+        // pure-compute hot path pays (almost) nothing. Bias is over-wake:
+        // a spurious wake costs one futex round-trip and a failed pop; a
+        // missed wake would cost a stranded actor.
+        self.coord.wake_one_if_idle();
     }
 
     /// Make `pid` runnable if it is parked; coalesce or defer otherwise.
@@ -1076,7 +1111,16 @@ impl Runtime {
             self.inner.live_actors.load(Ordering::Acquire), 0,
             "run() called while previous run still active"
         );
-        let io_thread = match IoThread::start() {
+        // RFC 018: the IO producers reach the runtime (slot table + unpark)
+        // through a Weak, so no RuntimeInner → IoThread → RuntimeInner cycle
+        // forms. Reset the in-flight counters BEFORE the threads can touch
+        // them (a prior run left them at 0 on a clean exit; the asserts pin
+        // that).
+        debug_assert_eq!(self.inner.io_outstanding.load(Ordering::Acquire), 0);
+        debug_assert_eq!(self.inner.io_fd_waiters.load(Ordering::Acquire), 0);
+        self.inner.io_outstanding.store(0, Ordering::Release);
+        self.inner.io_fd_waiters.store(0, Ordering::Release);
+        let io_thread = match IoThread::start(Arc::downgrade(&self.inner)) {
             Ok(io) => io,
             Err(e) => panic!("failed to start IO thread: {e}"),
         };
@@ -1179,6 +1223,8 @@ impl Runtime {
         }
         self.inner.io_parked.store(0, Ordering::Relaxed);
         self.inner.sleeping.store(0, Ordering::Relaxed);
+        self.inner.io_outstanding.store(0, Ordering::Relaxed);
+        self.inner.io_fd_waiters.store(0, Ordering::Relaxed);
 
         RUNTIME.with(|r| *r.borrow_mut() = None);
 
@@ -1494,53 +1540,30 @@ fn stop_live_actors(inner: &Arc<RuntimeInner>) {
 }
 
 // ---------------------------------------------------------------------------
-// schedule_loop — runs on each scheduler OS thread
+// Timer firing — shared by the busy-path due-check and the timekeeper
 // ---------------------------------------------------------------------------
 
-fn schedule_loop(inner: &Arc<RuntimeInner>, slot_idx: usize) {
-    crate::preempt::configure_preempt(inner.alloc_interval, inner.timeslice_cycles);
-    let stats = &inner.stats[slot_idx];
-
-    loop {
-        // ----------------------------------------------------------------
-        // 1. Try to win the drain lock (timers + IO). One winner per round;
-        //    losers skip immediately and proceed to step 2.
-        // ----------------------------------------------------------------
-        if let Ok(_drain_guard) = inner.drain_lock.try_lock() {
-            // Timers and IO live behind their own mutexes (phase 1), so the
-            // pure-yield / pure-compute hot path never contends a global lock
-            // just to discover there is nothing to drain. The clock is read
-            // only when the timer heap is non-empty.
-            let due = {
-                let mut t = match inner.timers.lock() {
-                    Ok(t) => t,
+/// Pop and dispatch every due timer. `pop_due` re-anchors the
+/// earliest-deadline snapshot under the timers mutex before returning, so
+/// a caller that raced a concurrent insert simply comes back on the next
+/// due-check. Dispatch runs with the timers lock released.
+fn fire_due_timers(inner: &Arc<RuntimeInner>, try_only: bool) {
+    let due = if try_only {
+        // Busy path: if another scheduler is already in the timers mutex
+        // (firing, inserting, or peeking) skip — the snapshot stays due
+        // until someone actually pops, so the check re-fires next loop.
+        match inner.timers.try_lock() {
+            Ok(mut t) => t.pop_due(std::time::Instant::now()),
+            Err(std::sync::TryLockError::WouldBlock) => return,
+            Err(std::sync::TryLockError::Poisoned(e)) => {
+                panic!("smarm: timers lock poisoned (core corrupt): {e}")
+            }
+        }
+    } else {
+        match inner.timers.lock() {
+            Ok(mut t) => t.pop_due(std::time::Instant::now()),
             Err(e) => panic!("smarm: timers lock poisoned (core corrupt): {e}"),
-                };
-                if t.is_empty() {
-                    Vec::new()
-                } else {
-                    t.pop_due(std::time::Instant::now())
         }
-            };
-            let completions = match inner.io.lock() {
-                Ok(mut io) => io
-                    .as_mut()
-                    .map(|io| {
-                        // Consume wake-pipe bytes ONLY here, under the drain
-                        // lock and strictly before draining completions.
-                        // Producers push their completion before writing the
-                        // byte, so every byte consumed here has its completion
-                        // visible to the drain below. Consuming bytes anywhere
-                        // else — in particular after an idle poll, outside the
-                        // lock — loses wakeups: a try_lock loser can eat the
-                        // byte for a completion the winner never saw, leaving
-                        // it stranded (and its EPOLLONESHOT fd disarmed) until
-                        // an unrelated timer forces another drain pass.
-                        crate::io::drain_wake_pipe(io.wake_fd());
-                        io.drain_completions()
-                    })
-                    .unwrap_or_default(),
-                Err(e) => panic!("smarm: io lock poisoned (core corrupt): {e}"),
     };
     for entry in due {
         match entry.reason {
@@ -1549,9 +1572,7 @@ fn schedule_loop(inner: &Arc<RuntimeInner>, slot_idx: usize) {
             // actor is between `timers.insert_sleep` and
             // `park_current`; RunningNotified makes the upcoming park
             // re-queue), or gone (no-op).
-                    crate::timer::Reason::Sleep { epoch } => {
-                        inner.unpark_at(entry.pid, epoch)
-                    }
+            crate::timer::Reason::Sleep { epoch } => inner.unpark_at(entry.pid, epoch),
             crate::timer::Reason::WaitTimeout { target, epoch } => {
                 // The callback may call unpark_at itself.
                 target.on_timeout(entry.pid, epoch);
@@ -1566,57 +1587,26 @@ fn schedule_loop(inner: &Arc<RuntimeInner>, slot_idx: usize) {
             crate::timer::Reason::Send { fire } => fire(),
         }
     }
+}
 
-            for completion in completions {
-                match completion {
-                    crate::io::Completion::Blocking { pid, epoch, result } => {
-                        match inner.io.lock() {
-                            Ok(mut io) => {
-                                if let Some(io) = io.as_mut() {
-                                    io.outstanding = io.outstanding.saturating_sub(1);
+// ---------------------------------------------------------------------------
+// schedule_loop — runs on each scheduler OS thread
+// ---------------------------------------------------------------------------
+
+fn schedule_loop(inner: &Arc<RuntimeInner>, slot_idx: usize) {
+    crate::preempt::configure_preempt(inner.alloc_interval, inner.timeslice_cycles);
+    let stats = &inner.stats[slot_idx];
+
+    loop {
+        // ----------------------------------------------------------------
+        // 1. Busy-path timer due-check (RFC 018 design point (a)): under
+        //    saturation nobody parks, so no timekeeper exists — due timers
+        //    must still fire. One Relaxed load + branch when no timer is
+        //    armed; the clock is read only when one is.
+        // ----------------------------------------------------------------
+        if inner.coord.deadline_due() {
+            fire_due_timers(inner, true);
         }
-                            }
-                            Err(e) => {
-                                panic!("smarm: io lock poisoned (core corrupt): {e}")
-                            }
-                        }
-                        // Stash the result under the cold lock, then unpark.
-                        // The protocol also covers the submit→park window
-                        // (RunningNotified), which the old code missed for
-                        // Blocking completions — a latent lost wakeup.
-                        if let Some(slot) = inner.slot_at(pid) {
-                            {
-                                let mut cold = slot.cold.lock();
-                                if slot.generation() == pid.generation() {
-                                    cold.pending_io_result = Some(result);
-                                } else {
-                                    // Actor died (stopped) with the op in
-                                    // flight; discard the result.
-                                }
-                            }
-                            inner.unpark_at(pid, epoch);
-                        }
-                    }
-                    crate::io::Completion::FdReady { fd, events: _ } => {
-                        // Resolve the parked pid under the io lock, then wake
-                        // through the protocol. Lock order: io before all.
-                        let parked = match inner.io.lock() {
-                            Ok(mut io) => io.as_mut().and_then(|io| {
-                                let entry = io.waiters.remove(&fd);
-                                io.epoll_deregister(fd);
-                                entry
-                            }),
-                            Err(e) => {
-                                panic!("smarm: io lock poisoned (core corrupt): {e}")
-                            }
-                        };
-                        if let Some((pid, epoch)) = parked {
-                            inner.unpark_at(pid, epoch);
-                        }
-                    }
-                }
-            }
-        } // drain_guard drops here
 
         // ----------------------------------------------------------------
         // 2. Pop a runnable pid. Pop order (RFC 005): wake slot first, then
@@ -1625,7 +1615,7 @@ fn schedule_loop(inner: &Arc<RuntimeInner>, slot_idx: usize) {
         // ----------------------------------------------------------------
         enum Pop {
             Got(Pid),
-            Idle { io_outstanding: u32, wake_fd: Option<std::os::fd::RawFd> },
+            Idle,
             AllDone,
             /// Root has exited and nothing is runnable: stop the parked-forever
             /// remainder, then re-pop. Fires at most once per run.
@@ -1650,19 +1640,12 @@ fn schedule_loop(inner: &Arc<RuntimeInner>, slot_idx: usize) {
             crate::te!(crate::trace::Event::SlotPop(pid));
             pid
         } else {
-            // Read IO liveness BEFORE the queue lock (phase-1 ordering: a
-            // completion resurrects an actor only via the drain path, whose
-            // enqueue would be visible under the queue lock we take next).
-            let (io_out, io_fd) = {
-                let io = match inner.io.lock() {
-                    Ok(io) => io,
-                    Err(e) => panic!("smarm: io lock poisoned (core corrupt): {e}"),
-                };
-                match io.as_ref() {
-                    Some(io) => (io.outstanding + io.waiters.len() as u32, Some(io.wake_fd())),
-                    None => (0, None),
-                }
-            };
+            // Read IO liveness BEFORE the queue pop — two atomic loads now
+            // (RFC 018), not a per-pop `io.lock`: a completion resurrects
+            // an actor via the producer's own unpark→enqueue, whose entry
+            // would be visible to the pop below.
+            let io_out = inner.io_outstanding.load(Ordering::Acquire)
+                + inner.io_fd_waiters.load(Ordering::Acquire);
 
             stats.run_queue_len.store(inner.run_queue.len(), Ordering::Relaxed);
             let pop = match inner.run_queue.pop() {
@@ -1694,7 +1677,7 @@ fn schedule_loop(inner: &Arc<RuntimeInner>, slot_idx: usize) {
                         // the idle wait below on the next pass.
                         Pop::RootDrain
                     } else {
-                        Pop::Idle { io_outstanding: io_out, wake_fd: io_fd }
+                        Pop::Idle
                     }
                 }
             };
@@ -1709,22 +1692,15 @@ fn schedule_loop(inner: &Arc<RuntimeInner>, slot_idx: usize) {
                         Ok(mut timers) => timers.clear(),
                         Err(e) => panic!("smarm: timers lock poisoned (core corrupt): {e}"),
                     }
-                    // Terminal wake: a sibling scheduler may be blocked in its
-                    // idle wait on a snapshot that is now terminally stale — an
-                    // orphaned long deadline (it would sleep it out in full) or
-                    // a stale `io_outstanding > 0` from a stop-cancelled waiter
-                    // (it would block in poll(-1) forever; cancellation produces
-                    // no completion, so nothing else writes the wake pipe).
-                    // One byte wakes every poller; each re-runs the verdict,
-                    // reaches AllDone itself, and re-wakes — idempotent.
-                    match inner.io.lock() {
-                        Ok(io) => {
-                            if let Some(io) = io.as_ref() {
-                                io.wake();
-                            }
-                        }
-                        Err(e) => panic!("smarm: io lock poisoned (core corrupt): {e}"),
-                    }
+                    // Terminal wake (replaces the wake-pipe byte): a sibling
+                    // may be parked on a snapshot that is now terminally
+                    // stale — an orphaned long deadline, or a stale
+                    // `io_fd_waiters > 0` from a stop-cancelled waiter
+                    // (cancellation produces no completion, so nothing else
+                    // will ever wake it). `wake_all` permits every parker;
+                    // each sibling re-runs the verdict, reaches AllDone
+                    // itself, and re-wakes — idempotent.
+                    inner.coord.wake_all();
                     return;
                 }
                 Pop::RootDrain => {
@@ -1734,40 +1710,51 @@ fn schedule_loop(inner: &Arc<RuntimeInner>, slot_idx: usize) {
                     stop_live_actors(inner);
                     continue;
                 }
-                Pop::Idle { io_outstanding, wake_fd } => {
-                    // Something is still in flight. Sleep on the appropriate
-                    // source to avoid hammering the queue mutex; retry on wake.
-                    let next_deadline = match inner.timers.lock() {
-                        Ok(timers) => timers.peek_deadline(),
-                        Err(e) => panic!("smarm: timers lock poisoned (core corrupt): {e}"),
+                Pop::Idle => {
+                    // Something is still in flight. Park on our own futex
+                    // until a producer wakes us (enqueue tail), a deadline
+                    // passes, or the re-check finds the world changed.
+                    //
+                    // Timekeeper (RFC 018): at most one parked scheduler
+                    // holds the timer deadline — the first idler to arm it
+                    // parks with a timeout, the rest park indefinitely, so a
+                    // timer expiry wakes one scheduler, not a herd. Peek and
+                    // arm under the timers mutex (the serialization that
+                    // makes the insert-side re-arm race-free).
+                    let tk_deadline = {
+                        let timers = match inner.timers.lock() {
+                            Ok(t) => t,
+                            Err(e) => {
+                                panic!("smarm: timers lock poisoned (core corrupt): {e}")
+                            }
                         };
-                    match (next_deadline, wake_fd) {
-                        (Some(deadline), fd_opt) => {
-                            let now = std::time::Instant::now();
-                            if deadline > now {
-                                let timeout = deadline - now;
-                                match fd_opt {
-                                    Some(fd) => {
-                                        // Wake only; the byte (if any) is
-                                        // consumed by the next drain-lock
-                                        // winner in phase 1. Level-triggered
-                                        // poll means an unconsumed byte makes
-                                        // this return immediately, so a loser
-                                        // spins briefly until the winner
-                                        // releases — never sleeps through it.
-                                        crate::io::poll_wake(fd, Some(timeout));
-                                    }
-                                    None => thread::sleep(timeout),
-                                }
-                            }
-                        }
-                        (None, Some(fd)) if io_outstanding > 0 => {
-                            // See above: no byte consumption outside phase 1.
-                            crate::io::poll_wake(fd, None);
-                        }
-                        _ => {
-                            thread::sleep(std::time::Duration::from_micros(100));
-                        }
+                        timers
+                            .peek_deadline()
+                            .filter(|d| inner.coord.try_arm_timer(slot_idx, *d))
+                    };
+                    // The mandatory post-publish re-check: a producer that
+                    // enqueued (or a verdict input that flipped) before it
+                    // could see our idle bit has left us the evidence.
+                    let _ = inner.coord.park(slot_idx, tk_deadline, || {
+                        !inner.run_queue.is_empty()
+                            || (inner.live_actors.load(Ordering::Acquire) == 0
+                                && inner.io_outstanding.load(Ordering::Acquire) == 0
+                                && inner.io_fd_waiters.load(Ordering::Acquire) == 0)
+                            || (inner.root_exited.load(Ordering::Acquire)
+                                && !inner.root_swept.load(Ordering::Acquire))
+                            || inner.coord.deadline_due()
+                    });
+                    if tk_deadline.is_some() {
+                        // Hand the role back BEFORE firing: pop_due can run
+                        // `Send` thunks that insert new timers, and the
+                        // insert-side re-arm check must see either no
+                        // timekeeper (skip) or a real parked one — never us,
+                        // awake and about to re-peek anyway.
+                        inner.coord.disarm_timer(slot_idx);
+                        // Woken for the deadline, for work, or to re-peek
+                        // after an earlier insert — fire whatever is due;
+                        // the next idle pass re-arms with the new minimum.
+                        fire_due_timers(inner, false);
                     }
                     continue;
                 }
@@ -1780,6 +1767,21 @@ fn schedule_loop(inner: &Arc<RuntimeInner>, slot_idx: usize) {
         //    by the at-most-once-enqueued invariant nothing else can have
         //    changed the state of a queued actor.
         // ----------------------------------------------------------------
+
+        // RFC 018 chain rule: we just took one runnable; if more remain and
+        // a sibling is parked, wake exactly one so the surplus runs in
+        // PARALLEL rather than serially behind us (without this the surplus
+        // is not stranded — we re-pop it after resuming — but it waits out
+        // our whole timeslice while an idle core sits available). Cheap: the
+        // queue-length check is queue-local, and `wake_one_if_idle` is a
+        // fence + one Relaxed mask load when nobody is parked. A Relaxed
+        // miss here is safe — the enqueue that created the surplus already
+        // issued its own wake (RFC 018 no-lost-wake); this only sharpens
+        // parallelism latency.
+        if !inner.run_queue.is_empty() {
+            inner.coord.wake_one_if_idle();
+        }
+
         let slot = match inner.slot_at(pid) {
             Some(s) => s,
             None => continue, // can't happen for real pids; defensive

src/scheduler.rs

@@ -72,6 +72,7 @@ use crate::runtime::{
     self, RuntimeInner, YieldIntent, RUNTIME,
 };
 use crate::supervisor::Signal;
+use std::sync::atomic::Ordering;
 use std::sync::Arc;
 
 // ---------------------------------------------------------------------------
@@ -753,7 +754,14 @@ where
                 Err(e) => panic!("smarm: io lock poisoned (core corrupt): {e}"),
             };
             match io.as_mut() {
-                Some(io) => io.submit(me, epoch, work),
+                Some(io) => {
+                    // RFC 018: count the op in flight BEFORE submit — the
+                    // pool decrements on completion, and an increment that
+                    // trailed the completion would underflow. Under the io
+                    // lock, so ordered against the same-lock submit.
+                    inner.io_outstanding.fetch_add(1, Ordering::AcqRel);
+                    io.submit(me, epoch, work);
+                }
                 None => panic!("io thread not started"),
             }
         });
@@ -813,7 +821,17 @@ fn wait_fd(fd: std::os::fd::RawFd, readable: bool, writable: bool) -> std::io::R
             Err(e) => panic!("smarm: io lock poisoned (core corrupt): {e}"),
         };
         match io.as_mut() {
-            Some(io) => io.epoll_register(fd, me, epoch, readable, writable),
+            Some(io) => {
+                // RFC 018: count the waiter BEFORE the ADD (mirror of
+                // submit); roll back if the registration fails so a
+                // rejected wait leaves the verdict counters clean.
+                inner.io_fd_waiters.fetch_add(1, Ordering::AcqRel);
+                let r = io.epoll_register(fd, me, epoch, readable, writable);
+                if r.is_err() {
+                    inner.io_fd_waiters.fetch_sub(1, Ordering::AcqRel);
+                }
+                r
+            }
             None => panic!("io thread not started"),
         }
     })?;
@@ -838,9 +856,12 @@ fn wait_fd(fd: std::os::fd::RawFd, readable: bool, writable: bool) -> std::io::R
                     Err(e) => panic!("smarm: io lock poisoned (core corrupt): {e}"),
                 };
                 if let Some(io) = io.as_mut() {
-                    if io.waiters.get(&self.fd) == Some(&(self.me, self.epoch)) {
-                        io.waiters.remove(&self.fd);
-                        io.epoll_deregister(self.fd);
+                    // `cancel_waiter` removes + DELs iff still ours, all
+                    // under the waiters lock (the ADD/DEL serialization);
+                    // decrement only when we actually removed it — a
+                    // FdReady that consumed it already did the decrement.
+                    if io.cancel_waiter(self.fd, self.me, self.epoch) {
+                        inner.io_fd_waiters.fetch_sub(1, Ordering::AcqRel);
                     }
                 }
             });
@@ -908,7 +929,14 @@ impl crate::channel::Selectable for FdArm {
             };
             match io.as_mut() {
                 Some(io) => {
-                    io.epoll_register(self.fd, pid, epoch, self.readable, self.writable)
+                    inner.io_fd_waiters.fetch_add(1, Ordering::AcqRel);
+                    let r = io.epoll_register(
+                        self.fd, pid, epoch, self.readable, self.writable,
+                    );
+                    if r.is_err() {
+                        inner.io_fd_waiters.fetch_sub(1, Ordering::AcqRel);
+                    }
+                    r
                 }
                 None => panic!("io thread not started"),
             }
@@ -936,9 +964,8 @@ impl crate::channel::Selectable for FdArm {
                 Err(e) => panic!("smarm: io lock poisoned (core corrupt): {e}"),
             };
             if let Some(io) = io.as_mut() {
-                if io.waiters.get(&self.fd) == Some(&(pid, epoch)) {
-                    io.waiters.remove(&self.fd);
-                    io.epoll_deregister(self.fd);
+                if io.cancel_waiter(self.fd, pid, epoch) {
+                    inner.io_fd_waiters.fetch_sub(1, Ordering::AcqRel);
                 }
             }
         });

src/sync_shim.rs

@@ -6,10 +6,19 @@
 //! Build the loom models with: `RUSTFLAGS="--cfg loom" cargo test --lib --release`
 
 #[cfg(loom)]
-pub(crate) use loom::sync::atomic::{AtomicU64, AtomicUsize, Ordering};
+pub(crate) use loom::sync::atomic::{fence, AtomicU64, AtomicUsize, Ordering};
 
 #[cfg(not(loom))]
-pub(crate) use std::sync::atomic::{AtomicU64, AtomicUsize, Ordering};
+pub(crate) use std::sync::atomic::{fence, AtomicU64, AtomicUsize, Ordering};
+
+// park.rs condvar-parker (loom + non-Linux builds only; the Linux non-loom
+// build parks on a futex and never touches these — gating them identically
+// keeps the default build free of unused imports).
+#[cfg(loom)]
+pub(crate) use loom::sync::{Condvar, Mutex};
+
+#[cfg(all(not(loom), not(target_os = "linux")))]
+pub(crate) use std::sync::{Condvar, Mutex};
 
 /// `UnsafeCell` with loom's `with`/`with_mut` access API; pass-through cost
 /// is zero in normal builds (`#[inline]`, newtype over std's cell).

src/timer.rs

@@ -141,6 +141,14 @@ impl PartialOrd for Entry {
 
 #[derive(Default)]
 pub struct Timers {
+    /// RFC 018: the scheduler coordination layer. Attached once at
+    /// `RuntimeInner::new`; every insert notes its deadline (min-maintained
+    /// snapshot for the busy-path due-check + the timekeeper re-arm wake)
+    /// and every pop/clear re-anchors the snapshot to the heap minimum.
+    /// All calls happen under the timers mutex — the serialization the
+    /// coordinator's timer protocol mandates. `None` only in unit tests
+    /// that construct a bare `Timers`.
+    coord: Option<std::sync::Arc<crate::park::Coordinator>>,
     /// Reverse-wrapped so the smallest deadline is at the top.
     heap: BinaryHeap<Reverse<Entry>>,
     /// Monotonic counter for the tiebreaker `seq` field (and the `TimerId` of a
@@ -157,7 +165,18 @@ pub struct Timers {
 
 impl Timers {
     pub fn new() -> Self {
-        Self { heap: BinaryHeap::new(), next_seq: 0, armed: std::collections::HashSet::new() }
+        Self {
+            coord: None,
+            heap: BinaryHeap::new(),
+            next_seq: 0,
+            armed: std::collections::HashSet::new(),
+        }
+    }
+
+    /// Attach the scheduler coordination layer (RFC 018). Called once, at
+    /// runtime construction, before any scheduler thread exists.
+    pub(crate) fn attach_coordinator(&mut self, c: std::sync::Arc<crate::park::Coordinator>) {
+        self.coord = Some(c);
     }
 
     /// Insert a `Sleep` timer. Convenience for the common case.
@@ -242,6 +261,13 @@ impl Timers {
             #[cfg(feature = "smarm-causal")]
             wall,
         }));
+        // RFC 018: publish the (possibly new-minimum) deadline to the
+        // busy-path snapshot and wake the timekeeper if it is parked
+        // toward a later one. We hold the timers mutex — the mandated
+        // serialization for both.
+        if let Some(c) = &self.coord {
+            c.note_deadline(deadline);
+        }
         seq
     }
 
@@ -255,6 +281,9 @@ impl Timers {
     pub fn clear(&mut self) {
         self.heap.clear();
         self.armed.clear();
+        if let Some(c) = &self.coord {
+            c.refresh_deadline(None);
+        }
     }
 
     /// Soonest pending deadline, or `None` if the heap is empty.
@@ -324,6 +353,13 @@ impl Timers {
             }
             out.push(entry);
         }
+        // RFC 018: re-anchor the busy-path snapshot to the new heap minimum
+        // (still under the timers mutex). A causal-shift re-queue above went
+        // through `heap.push` directly, so this peek is the one place the
+        // snapshot is guaranteed to catch up.
+        if let Some(c) = &self.coord {
+            c.refresh_deadline(self.peek_deadline());
+        }
         out
     }
 }

tests/park_wake.rs

@@ -0,0 +1,69 @@
+//! RFC 018 scheduler park/wake — observable-behavior guards.
+//!
+//! These pin the two timer-latency properties the park/wake swap must
+//! preserve or introduce:
+//!
+//! - `sleep_fires_under_saturation`: due timers fire even when every
+//!   scheduler is busy (nobody parked ⇒ no timekeeper) — the busy-path
+//!   due-check, ratified design point (a). The old drain phase gave this
+//!   for free (timers drained every loop iteration); the new design must
+//!   not lose it.
+//! - `submillisecond_sleep_is_prompt`: a sub-ms sleep completes promptly.
+//!   Under the old wake pipe, `poll_wake`'s `as_millis` truncation turned
+//!   sub-ms deadlines into 0ms busy-polls (correct wall time, pathological
+//!   CPU); under park/wake the futex timespec carries full nanosecond
+//!   precision.
+
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::Arc;
+use std::time::{Duration, Instant};
+
+#[test]
+fn sleep_fires_under_saturation() {
+    let rt = smarm::runtime::init(smarm::runtime::Config::exact(4));
+    rt.run(|| {
+        let stop = Arc::new(AtomicBool::new(false));
+        let mut spinners = Vec::new();
+        // 8 spinners over 4 schedulers: the run queue never empties, so no
+        // scheduler ever parks and no timekeeper exists. Only the busy-path
+        // due-check can fire the sleeper's timer before the spinners quit.
+        for _ in 0..8 {
+            let stop = stop.clone();
+            spinners.push(smarm::spawn(move || {
+                let t0 = Instant::now();
+                while !stop.load(Ordering::Relaxed) && t0.elapsed() < Duration::from_secs(5) {
+                    smarm::yield_now();
+                }
+            }));
+        }
+        let t0 = Instant::now();
+        smarm::sleep(Duration::from_millis(10));
+        let dt = t0.elapsed();
+        stop.store(true, Ordering::Relaxed);
+        for s in spinners {
+            let _ = s.join();
+        }
+        assert!(
+            dt < Duration::from_millis(500),
+            "10ms sleep took {dt:?} under scheduler saturation — busy-path \
+             timer firing is broken (timekeeper-only firing stalls under load)"
+        );
+    });
+}
+
+#[test]
+fn submillisecond_sleep_is_prompt() {
+    let rt = smarm::runtime::init(smarm::runtime::Config::exact(2));
+    rt.run(|| {
+        // Warm one iteration, then measure.
+        smarm::sleep(Duration::from_micros(500));
+        let t0 = Instant::now();
+        smarm::sleep(Duration::from_micros(500));
+        let dt = t0.elapsed();
+        assert!(dt >= Duration::from_micros(400), "woke early: {dt:?}");
+        assert!(
+            dt < Duration::from_millis(100),
+            "500µs sleep took {dt:?} — sub-ms deadline handling is broken"
+        );
+    });
+}