feat: race over Futures via context.io.race (PLAN-IO-UNIFY Phase 4)

Re-home the proven first-wins race from sched.race(*Task) onto *Future handles
+ the Io protocol; the old Task-based race is REPLACED (ufcs overload-by-receiver
is rejected, and only 1821 used it).

- Protocol: add Io.current_park() -> ParkToken — the running fiber as a token,
  captured WITHOUT parking — so race can register the SAME coordinator across N
  futures' park slots, then park once via suspend_raw; any completion readies it.
  Scheduler returns {self.current} (bails outside a fiber); CBlockingIo returns
  {null} (race never parks there — futures are born .ready).
- race :: ufcs (io: Io, futures: $T) -> RaceResult(T), kept in sched.sx (it needs
  meta.sx's make_enum/make_variant; pulling that into the io.sx prelude part-file
  would cycle). Winner scan -> register/park/deregister -> make_variant the winner
  -> Phase-3 cancel each still-.pending loser (no join). RaceResult reused
  unchanged (*Future(R) projects field 0 'value' -> R).
- TRUE-cancel: parked losers stop at their next suspend (timers evicted by cancel's
  wake), so race returns at WINNER-time, not slowest-loser-time.
- Adversarial review fixes: (1) an all-failing/all-cancelling racer set no longer
  deadlock-aborts the scheduler — race bails loudly ('all futures settled without
  a winner') when nothing is .ready and nothing is still .pending; (2) only
  .pending losers are cancelled, so a loser that already .failed keeps its real
  outcome label instead of being stomped to .canceled.

Re-point 1821 to context.io.async + context.io.race (winner a=111, losers
.canceled, completion log only 'task 1 @ 10ms', final clock 10ms — was 30 under
the old cooperative join). New 1826 locks the failing-loser case. Byte-identical
on aarch64-macOS + aarch64-linux. Suite 853/0; .ir churn is the current_park
vtable method.

current/PLAN-IO-UNIFY.md

@@ -36,6 +36,30 @@ installed via `push Context { io = xx scheduler } { … s.run(); }` — exactly
 just with the scheduler now reachable as `context.io`.
 
 ## Status (2026-06-28)
+- **Phase 4 — `race` over Futures via `context.io.race`. DONE.** Re-homed the
+  proven first-wins race from `sched.race(*Task)` onto `*Future` handles + the
+  `Io` protocol; the old Task-based `race` is REPLACED (ufcs overload-by-receiver
+  is rejected — "duplicate top-level decl" — and only 1821 used it).
+  - **Protocol affordance:** added `Io.current_park() -> ParkToken` (the running
+    fiber as a token, captured WITHOUT parking) so race can register the SAME
+    coordinator across N futures' `park` slots, then park once via `suspend_raw`;
+    any completion `ready`s it. Scheduler returns `{self.current}` (bails outside
+    a fiber); CBlockingIo returns `{null}` (race never parks there — futures born
+    `.ready`). The await comment already anticipated this fan-in.
+  - **race** (`ufcs (io: Io, futures: $T) -> RaceResult(T)`, in sched.sx — it
+    needs meta.sx's `make_enum`/`make_variant`, and pulling that into the io.sx
+    prelude part-file would cycle): winner scan → register+park → deregister →
+    `make_variant` the winner → Phase-3 `cancel` each loser (NO join). `RaceResult`
+    reused unchanged (`*Future(R)` projects field 0 `value` → R).
+  - **Winner-time return:** with true cancellation the parked losers stop at their
+    next suspend (their timers evicted by cancel's wake), so race returns at the
+    winner's virtual time, not the slowest loser's. 1821 re-pointed to
+    `context.io.async` + `context.io.race`: `winner a=111`, losers `.canceled`,
+    completion log ONLY `task 1 @ 10ms`, final clock `10ms` (was 30 under the old
+    cooperative join). Byte-identical on aarch64-macOS + aarch64-linux. Suite
+    853/0; `.ir` churn (current_park vtable method) regenerated, only 1821 stdout
+    changed otherwise.
+
 - **Phase 3 — TRUE cancellation via `suspend_raw -> !`. DONE.** A cancelled async
   worker now abandons its body at its next suspend instead of running to
   completion. Pieces:

examples/concurrency/1821-concurrency-fiber-race.sx

@@ -1,21 +1,19 @@
-// Stream B2/A1 — structured first-wins `race` over the M:1 fiber scheduler.
+// Stream B2 — structured first-wins `race` over `context.io` (PLAN-IO-UNIFY
+// Phase 4). `context.io.race(.(a = fa, b = fb, c = fc))` takes a named tuple of
+// already-spawned `*Future(..)` handles (from `context.io.async`), SUSPENDS the
+// calling fiber until the FIRST is `.ready`, and returns a comptime-SYNTHESIZED
+// tagged-union (`RaceResult`) mirroring the tuple's labels — variant NAME = the
+// tuple label, payload = that future's result type. Here the three workers return
+// DIFFERENT types (i64 / bool / f64), so the minted union is
+// `enum { a: i64; b: bool; c: f64 }` and the winner is matched by label.
 //
-// `s.race((a: ta, b: tb, c: tc))` takes a named tuple of already-spawned
+// TRUE cancellation (Phase 3): the workers sleep 10/20/30 ms (deterministic
-// `*Task(..)` handles, SUSPENDS the calling fiber until the FIRST task is ready,
+// virtual clock), so `a` wins at t=10. The losers `b`/`c` are parked mid-`sleep`
-// and returns a comptime-SYNTHESIZED tagged-union (`RaceResult`) mirroring the
+// when cancelled; their next `suspend_raw` raises `Canceled` and unwinds the body,
-// tuple's labels — variant NAME = the tuple label, payload = that task's result
+// so their POST-SLEEP `rec(...)` NEVER runs and `race` returns at WINNER-time. The
-// type. Here the three tasks return DIFFERENT types (i64 / bool / f64), so the
+// completion log therefore shows ONLY `a @ 10ms`, and the final virtual clock is
-// minted union is `enum { a: i64; b: bool; c: f64 }` and the winner is matched by
+// 10 — NOT 30 (the old cooperative-join behaviour that let losers run to their
-// label. After picking the winner, `race` CANCELS and JOINS every loser, so no
+// natural end). The losers end `.canceled` with their work stopped.
-// loser fiber outlives the call (structured concurrency).
-//
-// Deterministic by virtual time (like 1817 — no real clock): the tasks sleep
-// 10/20/30 ms, so `a` (shortest) wins at t=10. Cancellation is COOPERATIVE (M:1,
-// no preemption): the losers were already parked mid-`sleep` when cancelled, so
-// they cannot be preempted — `race` joins them, letting each run to its natural
-// end (its value discarded) before returning. The completion log therefore shows
-// all three finishing (a@10 winner, b@20, c@30 joined) and the final virtual
-// clock is 30. Each loser's `canceled` flag is set and its worker `finished`.
 //
 // aarch64-pinned (the scheduler's per-arch asm + per-OS mmap/event constants):
 // runs end-to-end on a matching host (macOS + linux), ir-only on a mismatch.
@@ -31,25 +29,28 @@ main :: () -> i64 {
     ps := @s; pl := @lg;
 
     // The coordinator runs as a fiber so `race` has a `current` to park.
-    s.spawn(() => {
+    push .{ io = xx s } {
-        // Three async tasks with DIFFERENT result types and sleep durations.
+        ps.spawn(() => {
-        a := ps.go(() -> i64  => { ps.sleep(10); rec(pl, 1, ps.now_ms()); 111 });
+            // Three async workers, DIFFERENT result types and sleep durations.
-        b := ps.go(() -> bool => { ps.sleep(20); rec(pl, 2, ps.now_ms()); true });
+            a := context.io.async(() -> (i64,  !) => { try context.io.sleep(10); rec(pl, 1, context.io.now_ms()); 111 });
-        c := ps.go(() -> f64  => { ps.sleep(30); rec(pl, 3, ps.now_ms()); 2.5 });
+            b := context.io.async(() -> (bool, !) => { try context.io.sleep(20); rec(pl, 2, context.io.now_ms()); true });
+            c := context.io.async(() -> (f64,  !) => { try context.io.sleep(30); rec(pl, 3, context.io.now_ms()); 2.5 });
 
-        // Race them. `a` (sleep 10) wins; `b` and `c` are cancelled + joined.
+            // Race them. `a` (sleep 10) wins; `b` and `c` are cancelled — their
-        winner := ps.race(.(a = a, b = b, c = c));
+            // post-sleep work never runs (true cancellation).
+            winner := context.io.race(.(a = a, b = b, c = c));
             if winner == {
                 case .a: (v) { print("winner: a (i64)  = {}\n", v); }
                 case .b: (v) { print("winner: b (bool) = {}\n", v); }
                 case .c: (v) { print("winner: c (f64)  = {}\n", v); }
             }
 
-        // The losers were cancelled (flag set) and joined (worker finished).
+            // The losers were cancelled; their work was stopped at the suspend.
-        print("loser b: canceled={} finished={}\n", b.canceled, b.finished);
+            print("loser b: canceled={}\n", b.state == .canceled);
-        print("loser c: canceled={} finished={}\n", c.canceled, c.finished);
+            print("loser c: canceled={}\n", c.state == .canceled);
         });
-    s.run();
+        ps.run();
+    }
 
     print("completion order (id @ virtual-ms):\n");
     i := 0;

examples/concurrency/1826-concurrency-fiber-race-failing-loser.sx

@@ -0,0 +1,35 @@
+// Stream B2 — `context.io.race` tolerates a FAILING racer (PLAN-IO-UNIFY Phase 4).
+// A `race` is first-SUCCESS-wins: a racer that ends `.failed` is simply not a
+// winner candidate; as long as ANOTHER racer succeeds, `race` returns that winner.
+// Here `a` raises at t=5 and `b` succeeds (42) at t=10, so `b` wins. The failed
+// racer keeps its real outcome label (`.failed`) — `race` only cancels still-
+// in-flight (`.pending`) losers, so it never stomps `a`'s `.failed` to `.canceled`.
+//
+// (Regression: an all-FAILING racer set instead bails loudly — "race — all
+// futures settled without a winner" — rather than dead-locking the scheduler.)
+//
+// aarch64-pinned (the scheduler's per-arch asm): runs end-to-end on a matching
+// host (macOS + linux), ir-only on a mismatch.
+#import "modules/std.sx";
+sched :: #import "modules/std/sched.sx";
+
+main :: () -> i64 {
+    s := sched.Scheduler.init();
+    ps := @s;
+    push .{ io = xx s } {
+        ps.spawn(() => {
+            a := context.io.async(() -> (i64, !) => { try context.io.sleep(5);  raise error.Boom; });
+            b := context.io.async(() -> (i64, !) => { try context.io.sleep(10); 42 });
+            winner := context.io.race(.(a = a, b = b));
+            if winner == {
+                case .a: (v) { print("winner: a = {}\n", v); }
+                case .b: (v) { print("winner: b = {}\n", v); }
+            }
+            // The failing loser keeps its real outcome — not stomped to .canceled.
+            print("a: failed={} canceled={}\n", a.state == .failed, a.state == .canceled);
+        });
+        ps.run();
+    }
+    print("final clock: {}ms\n", s.now_ms());
+    return 0;
+}

examples/concurrency/expected/1821-concurrency-fiber-race.stdout

@@ -1,8 +1,6 @@
 winner: a (i64)  = 111
-loser b: canceled=1 finished=1
+loser b: canceled=true
-loser c: canceled=1 finished=1
+loser c: canceled=true
 completion order (id @ virtual-ms):
   task 1 @ 10ms
-  task 2 @ 20ms
+final virtual clock: 10ms
-  task 3 @ 30ms
-final virtual clock: 30ms

examples/concurrency/expected/1826-concurrency-fiber-race-failing-loser.build

@@ -0,0 +1 @@
+{ "target": "macos" }

examples/concurrency/expected/1826-concurrency-fiber-race-failing-loser.exit

@@ -0,0 +1 @@
+0

examples/concurrency/expected/1826-concurrency-fiber-race-failing-loser.stderr

@@ -0,0 +1 @@
+

examples/concurrency/expected/1826-concurrency-fiber-race-failing-loser.stdout

@@ -0,0 +1,3 @@
+winner: b = 42
+a: failed=true canceled=false
+final clock: 10ms

library/modules/std/core.sx

@@ -128,6 +128,13 @@ Io :: protocol #inline {
     poll :: (self: *Self, deadline_ms: i64) -> i64;
     now_ms :: (self: *Self) -> i64;
     arm_timer :: (self: *Self, deadline_ms: i64, park: ParkToken) -> *void;
+    // `current_park()` — a `ParkToken` identifying the CURRENTLY-running execution
+    // context, so a fan-in waiter (`race`) can register the SAME awaiter across
+    // several futures' `park` slots before parking once. A suspending impl
+    // returns `{ handle = <current fiber> }`; the blocking impl has no fiber and
+    // returns `{ handle = null }` (race never parks there — its futures are born
+    // `.ready`). Unlike `suspend_raw`, this captures the awaiter WITHOUT parking.
+    current_park :: (self: *Self) -> ParkToken;
 }
 
 // --- Context ---

library/modules/std/io.sx

@@ -73,6 +73,12 @@ impl Io for CBlockingIo {
     arm_timer :: (self: *CBlockingIo, deadline_ms: i64, park: ParkToken) -> *void {
         return null;
     }
+    // No fibers in the blocking model — there is no current execution context to
+    // register as a fan-in waiter. `race`'s futures are born `.ready` here, so it
+    // finds a winner without ever parking; this null token is never consulted.
+    current_park :: (self: *CBlockingIo) -> ParkToken {
+        return .{ handle = null };
+    }
 }
 
 // --- Future($R): the handle to an async task's eventual result ---

library/modules/std/sched.sx

@@ -765,6 +765,20 @@ impl Io for Scheduler {
         self.timers.append(t, self.own_allocator);
         return null;
     }
+
+    // The running fiber as a `ParkToken`, for fan-in registration (`race`): the
+    // caller stamps this handle into several futures' `park` slots so ANY of
+    // their completions `ready`s it, then parks once via `suspend_raw`. MUST be
+    // called from inside a fiber (race parks `current`); a null current would
+    // register a `null` waiter no completion can wake — bail loudly, mirroring
+    // `suspend_self` / `sleep` / `arm_timer`.
+    current_park :: (self: *Scheduler) -> ParkToken {
+        if self.current == null {
+            print("sched: current_park() called outside a fiber (no running fiber)\n");
+            abort();
+        }
+        return .{ handle = xx self.current };
+    }
 }
 
 // --- the context switch (naked) + first-entry trampoline -------------------
@@ -1154,112 +1168,120 @@ cancel :: ufcs (t: *Task($R)) {
     t.state = .canceled;
 }
 
-// --- B2/A1: structured first-wins `race` over the M:1 Task layer -------------
+// --- B2/A1: structured first-wins `race` over `context.io` Futures -----------
 //
-// `race((a: ta, b: tb, …))` starts from N already-spawned `*Task(..)` handles,
+// `context.io.race((a: fa, b: fb, …))` starts from N already-spawned `*Future(..)`
-// returns when the FIRST completes, and STRUCTURALLY cancels + joins the losers
+// handles (from `context.io.async`), returns when the FIRST is `.ready`, and
-// before returning — no loser fiber outlives the call. The result is a
+// CANCELS every loser before returning — with Phase-3 TRUE cancellation each loser
-// comptime-synthesized tagged-union (`RaceResult`) mirroring the input tuple's
+// stops at its next suspend, so `race` returns at WINNER-time, not slowest-loser-
-// labels: each variant's NAME is the tuple label, its payload is that task's
+// time. The result is a comptime-synthesized tagged-union (`RaceResult`) mirroring
-// result type. The tuple must be NAMED (`(a: ta, b: tb)`); a positional-tuple
+// the input tuple's labels: each variant's NAME is the tuple label, its payload is
-// form (`._0`/`._1` variants) is future work — `field_name` yields "" for an
+// that future's result type. The tuple must be NAMED (`(a: fa, b: fb)`); a
-// unnamed element, which `make_enum` rejects as a duplicate-name collision.
+// positional-tuple form (`._0`/`._1`) is future work — `field_name` yields "" for
+// an unnamed element, which `make_enum` rejects as a duplicate-name collision.
 //
-//     fa := s.go(() -> A => read_a(conn));   // *Task(A)
+// It runs over the `Io` PROTOCOL (`current_park`/`suspend_raw`/`ready`), so it is
-//     fb := s.go(() -> B => read_b(conn));   // *Task(B)
+// colorblind: under the fiber scheduler it really parks/wakes; under the blocking
-//     winner := s.race((a: fa, b: fb));      // RaceResult = enum { a: A; b: B }
+// `CBlockingIo` every future is born `.ready`, so the winner scan returns
-//     if winner == { case .a: (v) {…}  case .b: (v) {…} }   // loser cancelled+joined
+// immediately and it never parks. Lives here (not io.sx) because the `RaceResult`
+// synthesis needs the metatype WRITE side (`make_enum`/`make_variant`), and
+// meta.sx imports only std.sx — pulling it into the io.sx prelude part-file would
+// cycle.
+//
+//     fa := context.io.async(() -> (A, !) => read_a(conn));   // *Future(A)
+//     fb := context.io.async(() -> (B, !) => read_b(conn));   // *Future(B)
+//     winner := context.io.race(.(a = fa, b = fb));           // enum { a: A; b: B }
+//     if winner == { case .a: (v) {…}  case .b: (v) {…} }     // losers cancelled
 
-// Synthesize the race RESULT type for a named tuple `$T` of `*Task(..)` handles.
+// Synthesize the race RESULT type for a named tuple `$T` of `*Future(..)` handles.
-// `*Task(R)` projects to its result `R` via `field_type(pointee(field_type(T, i)), 0)`:
+// `*Future(R)` projects to its result `R` via `field_type(pointee(field_type(T, i)), 0)`:
-// `field_type(T, i)` = `*Task(R)`, `pointee` strips the pointer to `Task(R)`, and
+// `field_type(T, i)` = `*Future(R)`, `pointee` strips the pointer to `Future(R)`,
-// field 0 of `Task` is `value: R`. One nominal type per distinct `T` (type-fn
+// and field 0 of `Future` is `value: R` (the `Value :: R` type member is not a
-// identity), so the decl, every `make_variant(RaceResult(T), …)` call, and the
+// data field). One nominal type per distinct `T` (type-fn identity), so the decl,
-// `-> RaceResult(T)` return all name the SAME union. (The 0649 composition shape,
+// every `make_variant(RaceResult(T), …)` call, and the `-> RaceResult(T)` return
-// with `Task` in place of the stand-in `Box`.)
+// all name the SAME union.
 RaceResult :: ($T: Type) -> Type {
     vs : [field_count(T)]EnumVariant = ---;
     inline for 0..field_count(T) (i) {
         vs[i] = EnumVariant.{
             name    = field_name(T, i),                          // tuple label → variant name
-            payload = field_type(pointee(field_type(T, i)), 0),  // *Task(R) -> Task(R) -> R
+            payload = field_type(pointee(field_type(T, i)), 0),  // *Future(R) -> Future(R) -> R
         };
     }
     return make_enum("RaceResult", vs[0..field_count(T)]);
 }
 
-// Structured first-wins race. Suspends the calling fiber until the FIRST task is
+// Structured first-wins race over the `Io` protocol. Suspends the calling fiber
-// `.ready`, builds a `RaceResult(T)` carrying that winner's value, then CANCELS
+// until the FIRST future is `.ready`, builds a `RaceResult(T)` carrying that
-// and JOINS every loser before returning.
+// winner's value, then CANCELS every loser and returns immediately.
 //
-// MUST be called from inside a fiber (there must be a `current` to park), like
+// MUST be called from inside a fiber under a suspending `Io` (there must be a
-// `wait`/`sleep`; a null `current` bails loudly rather than dereferencing null.
+// `current` to park) — `current_park` bails loudly on a null current. Under
+// `CBlockingIo` the futures are already `.ready`, so the winner scan returns
+// without ever calling `current_park`/`suspend_raw`.
 //
-// COOPERATIVE-CANCEL SEMANTIC (M:1, no preemption): a loser already past its
+// TRUE-CANCEL SEMANTIC (Phase 3): each loser was parked mid-suspend when cancelled;
-// work's first line cannot be preempted — `cancel` sets its flag and the JOIN
+// `cancel(f)` flips its sticky flag and wakes its worker fiber, whose next
-// waits for the worker to reach its natural end (the value is discarded). A loser
+// `suspend_raw` raises `Canceled` and unwinds the body — its post-suspend work
-// that had not yet started skips its work entirely (`go`'s `if t.canceled == 0`
+// never runs. `race` does NOT join the losers (they unwind on their own next turn),
-// guard). Either way `race` returns only once every loser's worker has `finished`,
+// so it returns at winner-time. The winner's value is taken from `f.value`.
-// so no loser fiber is still live past the call (structured concurrency).
+race :: ufcs (io: Io, futures: $T) -> RaceResult(T) {
-race :: ufcs (self: *Scheduler, tasks: $T) -> RaceResult(T) {
+    // Phase 1 — first winner. Scan for an already-`.ready` future (lowest index
-    cur := self.current;
+    // wins on a same-tick tie → deterministic). If none, register THIS coordinator
-    if cur == null {
+    // (`current_park`) as the waiter on every still-`.pending` future's `park`
-        print("sched: race() called outside a fiber (no running fiber)\n");
+    // slot and park once via `suspend_raw`; any completion `ready`s us. On wake
-        abort();
+    // DEREGISTER from ALL of them (clear our handle): a later loser completion must
-    }
+    // never `ready` a coordinator that has since moved on (the spurious/lost-wakeup
-
+    // hazard `wake` guards). A spurious wake with nothing ready re-registers + re-parks.
-    // Phase 1 — first winner. Scan for an already-`.ready` task (lowest index
-    // wins on a same-tick tie → deterministic). If none, register the caller as
-    // the waiter on every still-`.pending` task and park. On wake DEREGISTER from
-    // ALL of them: a later loser completion must never wake the caller again — by
-    // the time it fires the caller may be running or parked on a different join,
-    // and a stale waiter-wake would be a spurious/lost wakeup (the queue-corruption
-    // hazard `wake` guards). A spurious wake with nothing ready re-registers and
-    // re-parks.
     winner_idx : i64 = -1;
     while winner_idx < 0 {
         inline for 0..field_count(T) (i) {
-            if winner_idx < 0 and tasks[i].state == .ready { winner_idx = i; }
+            if winner_idx < 0 and futures[i].state == .ready { winner_idx = i; }
         }
         if winner_idx >= 0 { break; }
+        me := io.current_park();
+        any_pending := false;
         inline for 0..field_count(T) (i) {
-            if tasks[i].state == .pending { tasks[i].waiter = xx cur; }
+            if futures[i].state == .pending { futures[i].park.handle = me.handle; any_pending = true; }
         }
-        self.suspend_self();
+        // No `.ready` winner and nothing still `.pending` → every racer settled
+        // `.failed`/`.canceled` with no success. `race` is first-SUCCESS-wins and
+        // its `RaceResult` carries only success values, so there is no winner to
+        // return. Parking here would deadlock (no future can ever `ready` us);
+        // bail loudly with a specific message instead. (A recoverable all-fail —
+        // a failable `race` that raises — is a deliberate future refinement.)
+        if !any_pending {
+            print("sched: race — all futures settled without a winner (all failed/canceled)\n");
+            abort();
+        }
+        // The coordinator is the user's fiber (no cancel flag), so `suspend_raw`
+        // never raises here; the `catch {}` just discards the `!` for the type.
+        pk : ParkToken = .{ handle = null };
+        io.suspend_raw(@pk) catch {};
         inline for 0..field_count(T) (i) {
-            if tasks[i].waiter == (xx cur) { tasks[i].waiter = null; }
+            if futures[i].park.handle == me.handle { futures[i].park.handle = null; }
         }
     }
 
-    // Phase 2 — build the winner variant. `tasks[i].value` carries the CONCRETE
+    // Phase 2 — build the winner variant. `futures[i].value` carries the CONCRETE
     // result type of variant `i` (comptime-cursor tuple indexing), so the matching
-    // unrolled arm constructs `RaceResult(T)`'s i-th variant directly — no nested
+    // unrolled arm constructs `RaceResult(T)`'s i-th variant directly. `i` is the
-    // `inline if` to recover the payload type. `i` is the comptime cursor; the
+    // comptime cursor; the runtime `if i == winner_idx` selects the one arm.
-    // runtime `if i == winner_idx` selects the one arm that fires.
     result : RaceResult(T) = ---;
     inline for 0..field_count(T) (i) {
         if i == winner_idx {
-            result = make_variant(RaceResult(T), i, tasks[i].value);
+            result = make_variant(RaceResult(T), i, futures[i].value);
         }
     }
 
-    // Phase 3 — cancel + JOIN every loser, one at a time. `cancel` sets the flag;
+    // Phase 3 — cancel every still-IN-FLIGHT loser. With true cancellation a
-    // the join then ensures the loser's worker fiber has `finished` (not merely
+    // parked loser's next `suspend_raw` raises `Canceled` and unwinds its body;
-    // been flagged): if it has not, register as ITS sole waiter and park until the
+    // its `park` was cleared above, so its completion `ready`s nobody. No join —
-    // worker's tail wakes us (the worker sets `finished = 1` and wakes its waiter
+    // `race` returns now. Only `.pending` losers are cancelled: a loser that
-    // whether it ran the work or skipped it on an early cancel). Checking
+    // already settled (`.ready`/`.failed`) is done — cancelling it would do
-    // `finished` BEFORE parking avoids a lost wakeup (mirrors `wait` checking
+    // nothing useful and would stomp its real outcome label to `.canceled`.
-    // `.ready`). Only the loser being joined has a registered waiter, so no other
-    // task's completion can wake us mid-join.
     inline for 0..field_count(T) (i) {
-        if i != winner_idx {
+        if i != winner_idx and futures[i].state == .pending { futures[i].cancel(); }
-            tasks[i].cancel();
-            if tasks[i].finished == 0 {
-                tasks[i].waiter = xx cur;
-                self.suspend_self();
-                if tasks[i].waiter == (xx cur) { tasks[i].waiter = null; }
-            }
-        }
     }
 
     return result;