@scriptc/cli-darwin-x64 0.0.0 → 0.2.0

package/dist/lib/runtime-sources/src/scr_async.c

@@ -0,0 +1,3920 @@
+/* Async runtime: stackful fibers, promises, and a dependency-free event
+ * loop (microtask queue + timer min-heap).
+ *
+ * Model (see docs/ir.md):
+ * - An async function's body is ordinary compiled C, run on its own fiber
+ *   (dedicated ucontext stack). Calling it runs the body EAGERLY until
+ *   the first suspension (JS's synchronous-prefix rule), then control
+ *   returns to the spawner with a +1 promise.
+ * - `await` on a pending promise parks the fiber on the promise's waiter
+ *   list and switches back to whoever resumed it. Fulfillment moves waiters
+ *   to the microtask queue (microtasks run before timers, like Node).
+ * - Each fiber carries its own exception cell (scr_exc_swap_cell) so a
+ *   pending exception can't leak across concurrent execution contexts. A
+ *   throw escaping an async body rejects its promise; awaiting a rejected
+ *   promise re-throws into the awaiter.
+ * - Loop exhaustion with still-suspended fibers exits normally (Node's
+ *   behavior); those stacks are deliberately not unwound, and the RC audit
+ *   downgrades to a note in that case.
+ */
+#define _XOPEN_SOURCE 700
+/* Anonymous mappings alongside the XSI ucontext API on Darwin and glibc. */
+#ifndef _DARWIN_C_SOURCE
+#define _DARWIN_C_SOURCE 1
+#endif
+#ifndef _DEFAULT_SOURCE
+#define _DEFAULT_SOURCE 1
+#endif
+#include "scr_runtime.h"
+
+#include <errno.h>
+#include <fcntl.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <time.h>
+#ifdef _WIN32
+/* Windows arm: fibers come from the Win32 Fibers API (CreateFiberEx /
+ * SwitchToFiber — the direct ucontext analog: cooperatively scheduled,
+ * same thread, own stack); the idle sleep is nanosleep (mingw-w64 ships
+ * it, over Sleep). poll(2) has no Windows arm — see the sleep seam in
+ * scr_loop_run: the poller-backed units (net/dgram/watch) are not built
+ * for win32 targets (native-toolchain.ts gates them), so their fds never appear; the
+ * events unit DOES cross-compile (scr_events.c's win32 arm) and is
+ * served by a capped nanosleep — dispatch at the next turn's top, the
+ * cap bounding signal/stdin latency instead of a pollable wake fd. */
+#include <windows.h>
+#elif defined(__wasi__)
+#include <poll.h>
+#else
+#include <poll.h>
+#include <pthread.h>
+#include <signal.h>
+#include <sys/mman.h>
+#include <ucontext.h>
+#include <unistd.h>
+#endif
+
+#ifdef __wasi__
+/* WASI Preview 1 has no process-spawn capability. The child unit is not
+ * linked for this target, but the generic loop keeps its quiescence hooks;
+ * their empty implementation makes the absence explicit at link time
+ * without weakening async/timer support. */
+bool scr_children_pending(void) { return false; }
+bool scr_children_ready(void) { return false; }
+bool scr_children_reffed_pending(void) { return false; }
+bool scr_children_failed_pending(void) { return false; }
+void scr_children_poll(void) {}
+int scr_children_wake_fd(void) { return -1; }
+bool scr_children_wait(double max_wait_ms) {
+  (void)max_wait_ms;
+  return false;
+}
+void scr_children_teardown(void) {}
+#endif
+
+#if defined(__has_feature)
+#if __has_feature(address_sanitizer)
+#define SCR_ASAN_FIBERS 1
+void __sanitizer_start_switch_fiber(void **fake_stack_save, const void *bottom, size_t size);
+void __sanitizer_finish_switch_fiber(void *fake_stack_save, const void **bottom_old, size_t *size_old);
+#endif
+#endif
+
+/* ASan inflates stack frames severalfold (unoptimized locals + redzones),
+ * so the sanitized lane gets proportionally bigger fiber stacks — same
+ * headroom DISCIPLINE, instrumented scale. Keep the island's stack budget
+ * (scr_island.c) at half of whichever size is active. 8MB under ASan is
+ * the measured need: an engine call costs 64–96KB there, and a real
+ * embedded graph entered FROM A FIBER (a commander action awaiting
+ * generateText — zod parses inside promise chains) nests dozens of engine
+ * frames; native POSIX stacks are mapped and committed lazily, so idle
+ * fibers pay address space, not RSS. */
+#ifdef SCR_ASAN_FIBERS
+#define SCR_FIBER_STACK (8 * 1024 * 1024)
+#else
+#define SCR_FIBER_STACK (256 * 1024)
+#endif
+
+#if !defined(_WIN32) && !defined(__wasi__)
+/* Keep stacks out of malloc's size classes. Reuse a bounded number during
+ * a burst of promise/generator work, then unmap the spares before the loop
+ * sleeps. This avoids both heap fragmentation and per-call mmap overhead.
+ * The sanitizer lane unmaps every finished stack immediately. */
+#ifndef SCR_ASAN_FIBERS
+#define SCR_FIBER_SPARES 4
+static SCR_TL void *scr_fiber_spares[SCR_FIBER_SPARES];
+static SCR_TL size_t scr_fiber_nspares;
+static SCR_TL bool scr_fiber_cleanup_registered;
+#endif
+
+static void scr_fiber_stacks_clear(void) {
+#ifndef SCR_ASAN_FIBERS
+  while (scr_fiber_nspares > 0) {
+    (void)munmap(scr_fiber_spares[--scr_fiber_nspares], SCR_FIBER_STACK);
+    scr_fiber_spares[scr_fiber_nspares] = NULL;
+  }
+#endif
+}
+
+static void *scr_fiber_stack_new(void) {
+#ifndef SCR_ASAN_FIBERS
+  if (scr_fiber_nspares > 0) {
+    void *stack = scr_fiber_spares[--scr_fiber_nspares];
+    scr_fiber_spares[scr_fiber_nspares] = NULL;
+    return stack;
+  }
+  if (!scr_fiber_cleanup_registered) {
+    scr_fiber_cleanup_registered = true;
+    scr_atexit(scr_fiber_stacks_clear);
+  }
+#endif
+  void *stack = mmap(NULL, SCR_FIBER_STACK, PROT_READ | PROT_WRITE,
+                    MAP_PRIVATE | MAP_ANON, -1, 0);
+  if (stack == MAP_FAILED) scr_trap("scriptc: out of memory\n");
+  return stack;
+}
+
+static void scr_fiber_stack_free(void *stack) {
+  if (!stack) return;
+#ifndef SCR_ASAN_FIBERS
+  if (scr_fiber_nspares < SCR_FIBER_SPARES) {
+    scr_fiber_spares[scr_fiber_nspares++] = stack;
+    return;
+  }
+#endif
+  (void)munmap(stack, SCR_FIBER_STACK);
+}
+#endif
+
+/* ── promises ─────────────────────────────────────────────────────────── */
+
+enum { SCR_PROM_PENDING = 0, SCR_PROM_FULFILLED = 1, SCR_PROM_REJECTED = 2 };
+
+typedef struct ScrFiber ScrFiber;
+
+struct ScrPromise {
+  size_t rc;
+  int state;
+  bool top_level_handled; /* uncaughtException handled entry evaluation */
+  /* Payload (fulfillment value or rejection reason), ScrExcCell-style.
+   * trace_fn is non-NULL iff the REF payload type carries a cycle header
+   * (a promise settled with a cycle-capable value can be a cycle member —
+   * e.g. a closure that captures a box holding this very promise). */
+  ScrExcKind payload_kind; /* NONE = void fulfillment */
+  double f64;
+  bool b;
+  void *payload;
+  void *(*retain_fn)(void *);
+  void (*release_fn)(void *);
+  ScrTraceFn trace_fn;
+  /* Fibers parked on this promise. */
+  ScrFiber **waiters;
+  size_t nwaiters, waiters_cap;
+  /* Combinator callbacks (Promise.race / Promise.all): at settle,
+   * fulfilled promises run adapt(dst, this) — an emitted adapter
+   * converting the payload into the destination's inner type — and
+   * rejections copy raw (reasons are dynamically tagged). Each entry owns
+   * a reference on its dst. Promise.all entries carry the shared
+   * countdown state (`all` non-NULL, `all_idx` the entry's INPUT index);
+   * race entries leave both zeroed and keep the adapt dispatch. */
+  struct ScrPromiseCbWaiter {
+    void (*adapt)(ScrPromise *dst, ScrPromise *src);
+    ScrPromise *dst;
+    struct ScrAllState *all;
+    size_t all_idx;
+  } *cbs;
+  size_t ncbs, cbs_cap;
+  /* Unhandled-rejection tracking: set when rejected, cleared on await. */
+  bool rejection_observed;
+  /* Set when the checkpoint report delivered THIS promise to
+   * 'unhandledRejection' listeners — a handler attached after that is
+   * Node's 'rejectionHandled' moment (scr_prom_observe below). */
+  bool reported_unhandled;
+};
+
+#ifdef SCR_RC_AUDIT
+static long scr_live_promises = 0;
+long scr_promise_live_count(void) { return scr_live_promises; }
+#endif
+
+/* ── Promise.all shared state ─────────────────────────────────────────
+ * One per Promise.all call: the values array filled per INPUT index as
+ * entries fulfill, the countdown of fulfillments still missing, and the
+ * per-element-kind store helper. The state holds +1 on `values` (NULL for
+ * void-element all) and is itself refcounted by the entries that still
+ * point at it (parked cb waiters plus the builder while it subscribes);
+ * the RESULT promise is NOT held here — each parked entry's cb.dst is the
+ * retained result, so promise teardown paths need no all-specific
+ * destination handling. */
+typedef struct ScrAllState {
+  size_t rc;
+  size_t remaining;
+  ScrArr *values;
+  void (*store)(ScrArr *a, double i, ScrPromise *src);
+} ScrAllState;
+
+static void scr_promise_all_state_release(ScrAllState *st) {
+  if (--st->rc == 0) {
+    if (st->values) scr_arr_release(st->values);
+    free(st);
+  }
+}
+
+static void scr_oom(void) {
+  fputs("scriptc: out of memory\n", stderr);
+  abort();
+}
+
+static void scr_promise_trace(void *o, ScrTraceVisit visit, void *ctx) {
+  ScrPromise *p = (ScrPromise *)o;
+  /* Waiters are fibers (not refcounted objects) — the settled payload and
+   * the combinator destinations are the strong references this promise
+   * owns (destination promises are cycle-headered). */
+  if ((p->payload_kind == SCR_EXC_REF || p->payload_kind == SCR_EXC_OBJ ||
+       p->payload_kind == SCR_EXC_PRIMITIVE_REF) &&
+      p->trace_fn) {
+    visit(p->payload, ctx);
+  }
+  for (size_t i = 0; i < p->ncbs; i++) visit(p->cbs[i].dst, ctx);
+}
+
+/* Teardown for the collector: release the payload only when the trace does
+ * NOT visit it (the complement — see the contract in scr_runtime.h). The
+ * cb destinations are always visited, so only the array frees here. */
+static void scr_promise_gcfree(void *o) {
+  ScrPromise *p = (ScrPromise *)o;
+  if (p->payload_kind == SCR_EXC_STR) {
+    scr_str_release((ScrStr *)p->payload);
+  } else if ((p->payload_kind == SCR_EXC_REF || p->payload_kind == SCR_EXC_OBJ ||
+              p->payload_kind == SCR_EXC_PRIMITIVE_REF) &&
+             p->payload && !p->trace_fn) {
+    p->release_fn(p->payload);
+  }
+  /* Parked Promise.all states are NOT traced children (their values array
+   * is reachable only through them) — release like the untraced payload
+   * above; the cb destinations stay untouched (trace visits them). */
+  for (size_t i = 0; i < p->ncbs; i++) {
+    if (p->cbs[i].all) scr_promise_all_state_release(p->cbs[i].all);
+  }
+  free(p->waiters);
+  free(p->cbs);
+#ifdef SCR_RC_AUDIT
+  scr_live_promises--;
+#endif
+  scr_cyc_free(p);
+}
+
+ScrPromise *scr_promise_new(void) {
+  ScrPromise *p = scr_cyc_alloc(sizeof *p, &scr_promise_trace, &scr_promise_gcfree);
+  p->rc = 1;
+#ifdef SCR_RC_AUDIT
+  scr_live_promises++;
+#endif
+  return p;
+}
+
+/* The 'rejectionHandled' hook (scr_async_dyn.c installs it at listener
+ * registration — the scr_urj_deliver_fn pattern, so listener-free
+ * binaries keep their size class): called when a promise the checkpoint
+ * report already delivered as unhandled gains a handler. */
+void (*scr_rjh_notify_fn)(ScrPromise *p) = NULL;
+
+/* Every handler attach funnels here: mark the rejection observed, and
+ * fire Node's 'rejectionHandled' when the attach arrived AFTER the
+ * report delivered this promise to 'unhandledRejection' listeners (the
+ * model's one late-handling window — earlier handling keeps the promise
+ * out of the report entirely). The flag clears on the first attach, so
+ * one report fires at most one 'rejectionHandled' — Node's pairing. */
+static void scr_prom_observe(ScrPromise *p) {
+  p->rejection_observed = true;
+  if (p->reported_unhandled) {
+    p->reported_unhandled = false;
+    if (scr_rjh_notify_fn != NULL) scr_rjh_notify_fn(p);
+  }
+}
+
+/* The attach-time handled mark (scr_async_dyn.c's dyn then/catch with a
+ * rejection handler, plus the module loader's ownership of every module
+ * evaluation promise): Node marks a promise handled at ATTACH, including
+ * while it is still pending, not when a later reaction runs. The pending
+ * mark is essential for a module dependency whose importer aborts while
+ * evaluating a later sibling — the loader still owns that dependency's
+ * eventual rejection, so it must never surface as an unrelated unhandled
+ * rejection. */
+void scr_promise_mark_handled(ScrPromise *p) {
+  scr_prom_observe(p);
+}
+
+ScrPromise *scr_promise_retain(ScrPromise *p) {
+  if (p && p->rc != SIZE_MAX) {
+    p->rc++;
+    scr_cyc_mark_live(p);
+  }
+  return p;
+}
+
+static void scr_promise_release_payload(ScrPromise *p) {
+  if (p->payload_kind == SCR_EXC_STR) scr_str_release((ScrStr *)p->payload);
+  else if ((p->payload_kind == SCR_EXC_REF || p->payload_kind == SCR_EXC_OBJ ||
+            p->payload_kind == SCR_EXC_PRIMITIVE_REF) &&
+           p->payload) p->release_fn(p->payload);
+  p->payload_kind = SCR_EXC_NONE;
+  p->payload = NULL;
+}
+
+void scr_promise_release(ScrPromise *p) {
+  if (!p || p->rc == SIZE_MAX) return;
+  if (--p->rc == 0) {
+    scr_cyc_on_dead(p);
+    scr_promise_release_payload(p);
+    free(p->waiters);
+    for (size_t i = 0; i < p->ncbs; i++) {
+      scr_promise_release(p->cbs[i].dst);
+      if (p->cbs[i].all) scr_promise_all_state_release(p->cbs[i].all);
+    }
+    free(p->cbs);
+#ifdef SCR_RC_AUDIT
+    scr_live_promises--;
+#endif
+    scr_cyc_free(p);
+  } else {
+    scr_cyc_on_release(p); /* possible cycle root; may collect — p is done */
+  }
+}
+
+void *scr_promise_retain_v(void *p) { return scr_promise_retain((ScrPromise *)p); }
+void scr_promise_release_v(void *p) { scr_promise_release((ScrPromise *)p); }
+void scr_promise_trace_v(void *p, ScrTraceVisit visit, void *ctx) {
+  scr_promise_trace(p, visit, ctx);
+}
+
+/* ── fibers and the scheduler ─────────────────────────────────────────── */
+
+/* One saved execution context. POSIX: a ucontext_t (swapcontext saves the
+ * outgoing context INTO the `from` slot). Windows: the fiber HANDLE —
+ * SwitchToFiber saves the outgoing state inside the fiber object itself,
+ * so the slot only needs to name the destination; `from` is unused. */
+#ifdef _WIN32
+typedef void *ScrCtx;
+#elif defined(__wasi__)
+typedef unsigned char ScrCtx;
+#else
+typedef ucontext_t ScrCtx;
+#endif
+
+struct ScrFiber {
+  ScrCtx ctx;
+  /* wasm32 uses LLVM's stackless switched-coroutine frame instead of a
+   * native saved stack. The compiler emits these two generic wrappers in
+   * every WASI module. */
+  void *coro;
+  ScrCtx *return_to; /* whoever resumed us last (spawner or the loop) */
+  char *stack;       /* POSIX only; Windows fibers own their stack (NULL) */
+  ScrPromise *promise; /* the promise this fiber settles (owned +1); NULL
+                        * for generator fibers (gen non-NULL instead) */
+  ScrGen *gen;         /* the generator this fiber runs (borrowed — the
+                        * ScrGen owns the fiber, never the reverse) */
+  ScrExcCell exc;
+  /* AsyncLocalStorage context (owned; NULL = empty). Inherited from the
+   * SPAWNER at spawn (Node's init-time capture), swapped active with the
+   * fiber (the exc-cell pattern) so run()'s window rides awaits. */
+  ScrAlsCtx *als;
+  bool done;
+  /* Trampoline args: the spawn wrapper stores a pointer to a stack-local
+   * argpack; the trampoline copies it out before the spawner resumes. */
+  void *argpack;
+  void (*entry)(ScrFiber *self, void *argpack);
+  /* queueMicrotask entries: a stackless pseudo-fiber — the READY queue's
+   * FIFO IS the microtask queue, so a closure rides it as a fiber-shaped
+   * envelope and scr_resume_fiber runs it on the main stack (a throw is
+   * an UNCAUGHT exception, like Node's queueMicrotask, never a
+   * rejection). Owned; NULL on real fibers. */
+  ScrClosure *micro_cb;
+  /* Async-generator settlement jobs: stackless READY-queue envelopes.
+   * The retained generator is pumped on the main stack. */
+  ScrGen *gen_job;
+#ifdef SCR_ASAN_FIBERS
+  void *fake_stack;
+#endif
+};
+
+/* ── AsyncLocalStorage (node:async_hooks) ─────────────────────────────
+ * Stores are process-lived ids; the CONTEXT is an immutable refcounted
+ * snapshot of (store id → dyn value) entries. One ACTIVE SLOT pointer
+ * (the exc-cell pattern): main owns a static slot, every fiber owns its
+ * own field, scr_switch repoints the active slot — so run()'s window
+ * rides the fiber across awaits, and a spawned fiber INHERITS the
+ * spawner's snapshot (Node's init-time capture through AsyncResource).
+ * Snapshots are immutable: enter builds a fresh one (entries retained),
+ * restore swaps the previous back — timer capture retains pointers. */
+
+/* The ScrAlsCtx layout and the API over it live in scr_runtime.h /
+ * scr_async_dyn.c (gated — the size-class stance); this always-linked
+ * core keeps only what the fiber machinery itself touches: the active
+ * slot, the snapshot RC pair, and the switch/spawn/destroy wiring. */
+static ScrAlsCtx *scr_als_main_slot = NULL;
+ScrAlsCtx **scr_als_active = &scr_als_main_slot;
+
+ScrAlsCtx *scr_als_ctx_retain(ScrAlsCtx *c) {
+  if (c) c->rc++;
+  return c;
+}
+
+void scr_als_ctx_release(ScrAlsCtx *c) {
+  if (!c || --c->rc != 0) return;
+  for (size_t i = 0; i < c->len; i++) scr_dyn_release(c->entries[i].value);
+  free(c);
+}
+
+
+
+static ScrCtx scr_loop_ctx; /* the main stack (scheduler home) */
+static ScrFiber *scr_current = NULL;
+static long scr_fibers_live = 0;
+static long scr_fibers_abandoned = 0;
+
+long scr_abandoned_fiber_count(void) { return scr_fibers_abandoned; }
+
+/* True while executing on an async fiber (vs the main stack). The island
+ * sizes its engine stack budget per stack: fibers are small and fixed,
+ * the main stack is the process's megabytes (scr_island.c isl_entry). */
+bool scr_on_fiber(void) { return scr_current != NULL; }
+void *scr_fiber_self(void) { return scr_current; }
+
+#ifdef __wasi__
+extern void scr_wasi_coro_resume(void *handle);
+extern void scr_wasi_coro_destroy(void *handle);
+
+void scr_wasi_coro_started(void *handle) {
+  if (scr_current == NULL) {
+    fputs("scriptc: internal error: coroutine started outside a fiber\n", stderr);
+    abort();
+  }
+  scr_current->coro = handle;
+}
+#endif
+
+/* Microtask queue: ready fibers, FIFO. */
+static ScrFiber **scr_ready = NULL;
+static size_t scr_ready_head = 0, scr_ready_len = 0, scr_ready_cap = 0;
+
+static void scr_ready_push(ScrFiber *f) {
+  if (scr_ready_head + scr_ready_len == scr_ready_cap) {
+    if (scr_ready_len > 0) {
+      memmove(scr_ready, scr_ready + scr_ready_head, scr_ready_len * sizeof *scr_ready);
+    }
+    scr_ready_head = 0;
+    if (scr_ready_len == scr_ready_cap) {
+      scr_ready_cap = scr_ready_cap ? scr_ready_cap * 2 : 16;
+      scr_ready = realloc(scr_ready, scr_ready_cap * sizeof *scr_ready);
+      if (!scr_ready) scr_oom();
+    }
+  }
+  scr_ready[scr_ready_head + scr_ready_len++] = f;
+}
+
+static void scr_async_gen_job(ScrGen *g);
+static void scr_async_gen_resume_ready(ScrGen *g);
+
+/* Timer min-heap, FIFO tiebreak via sequence numbers. `id` is nonzero only
+ * for setInterval entries (the clearInterval handle; setTimeout has no
+ * clear surface, so plain timeouts never need one) and `repeat_ms` is the
+ * interval's period — a fired interval re-enters the heap with a fresh
+ * deadline and seq, so ties against later timers stay FIFO like Node's. */
+typedef struct {
+  double deadline_ms;
+  unsigned long seq;
+  ScrClosure *cb; /* owned */
+  double repeat_ms;
+  unsigned long id;
+  bool reffed;     /* keeps the loop alive; unref() clears it (Node semantics) */
+  double delay_ms; /* the original delay — refresh() re-arms to now + this */
+} ScrTimer;
+
+static ScrTimer *scr_timers = NULL;
+static size_t scr_ntimers = 0, scr_timers_cap = 0;
+static unsigned long scr_timer_seq = 0;
+/* REF'd armed timers — the loop's timer-liveness count (an unref'd timer
+ * sits in the heap and still FIRES if the loop runs for other reasons, but
+ * does not by itself keep the loop alive). Kept in sync by push/pop/remove
+ * and by ref/unref. */
+static size_t scr_reffed_timers = 0;
+
+/* Exported: the island's timer machinery (scr_web.c) shares this clock so
+ * its deadlines are comparable with the loop's. */
+double scr_now_ms(void) {
+  struct timespec ts;
+  clock_gettime(CLOCK_MONOTONIC, &ts);
+  return (double)ts.tv_sec * 1000.0 + (double)ts.tv_nsec / 1e6;
+}
+
+static bool scr_timer_before(const ScrTimer *a, const ScrTimer *b) {
+  if (a->deadline_ms != b->deadline_ms) return a->deadline_ms < b->deadline_ms;
+  return a->seq < b->seq;
+}
+
+static void scr_timer_push(ScrTimer t) {
+  if (scr_ntimers == scr_timers_cap) {
+    scr_timers_cap = scr_timers_cap ? scr_timers_cap * 2 : 16;
+    scr_timers = realloc(scr_timers, scr_timers_cap * sizeof *scr_timers);
+    if (!scr_timers) scr_oom();
+  }
+  if (t.reffed) scr_reffed_timers++;
+  size_t i = scr_ntimers++;
+  scr_timers[i] = t;
+  while (i > 0) {
+    size_t parent = (i - 1) / 2;
+    if (!scr_timer_before(&scr_timers[i], &scr_timers[parent])) break;
+    ScrTimer tmp = scr_timers[i];
+    scr_timers[i] = scr_timers[parent];
+    scr_timers[parent] = tmp;
+    i = parent;
+  }
+}
+
+static ScrTimer scr_timer_pop(void) {
+  ScrTimer top = scr_timers[0];
+  if (top.reffed && scr_reffed_timers > 0) scr_reffed_timers--;
+  scr_timers[0] = scr_timers[--scr_ntimers];
+  size_t i = 0;
+  for (;;) {
+    size_t l = 2 * i + 1, r = 2 * i + 2, min = i;
+    if (l < scr_ntimers && scr_timer_before(&scr_timers[l], &scr_timers[min])) min = l;
+    if (r < scr_ntimers && scr_timer_before(&scr_timers[r], &scr_timers[min])) min = r;
+    if (min == i) break;
+    ScrTimer tmp = scr_timers[i];
+    scr_timers[i] = scr_timers[min];
+    scr_timers[min] = tmp;
+    i = min;
+  }
+  return top;
+}
+
+/* Node's delay coercion (lib/internal/timers.js): NaN/negative/sub-ms
+ * clamp to 1, the delay TRUNCATES to integer milliseconds (setTimeout(cb,
+ * 1.8) and setTimeout(cb, 1.1) both land in the 1ms bucket, so a batch of
+ * fractional delays fires in REGISTRATION order — test-timers-non-integer-
+ * delay's contract), and anything past TIMEOUT_MAX (2^31-1) becomes 1. */
+static double scr_timer_coerce_ms(double ms) {
+  if (!(ms >= 1)) return 1;
+  if (ms > 2147483647.0) return 1; /* past TIMEOUT_MAX (also +Infinity) */
+  return (double)(unsigned long long)ms; /* trunc: positive, finite, in range */
+}
+
+void scr_set_timeout(ScrClosure *cb, double ms) {
+  ms = scr_timer_coerce_ms(ms);
+  ScrTimer t = {scr_now_ms() + ms, scr_timer_seq++, cb /* ownership moves in */, 0, 0, true, ms};
+  scr_timer_push(t);
+}
+
+/* ── setInterval / clearInterval ─────────────────────────────────────
+ * Intervals ride the same heap: the entry carries its period and a
+ * nonzero id (the compiled handle — Node's Timeout object reduced to the
+ * number the fallback declarations promise). A live interval keeps the
+ * loop alive exactly because it sits in the heap; clearInterval removes
+ * the entry EAGERLY (a lazily-cancelled entry would hold the loop open
+ * until its deadline — an interval cleared at t=0 with a one-hour period
+ * must let the program exit NOW, like Node). Clearing the interval whose
+ * callback is currently running (the self-clearing spinner pattern) is
+ * handled by the firing pair below: the entry is out of the heap while
+ * its callback runs, so the run loop re-pushes it only if the callback
+ * did not clear it. */
+
+static unsigned long scr_interval_next_id = 1;
+static unsigned long scr_firing_id = 0; /* interval currently running */
+static bool scr_firing_cleared = false;
+static bool scr_firing_reffed = true;  /* the running interval's ref state */
+static bool scr_firing_refresh = false; /* refresh() called mid-callback */
+
+/* Removes heap entry i (swap-with-last, then sift whichever way the moved
+ * entry needs). */
+static void scr_timer_remove_at(size_t i) {
+  if (scr_timers[i].reffed && scr_reffed_timers > 0) scr_reffed_timers--;
+  scr_timers[i] = scr_timers[--scr_ntimers];
+  if (i >= scr_ntimers) return;
+  /* Sift up if the moved entry beats its parent, else sift down. */
+  while (i > 0) {
+    size_t parent = (i - 1) / 2;
+    if (!scr_timer_before(&scr_timers[i], &scr_timers[parent])) break;
+    ScrTimer tmp = scr_timers[i];
+    scr_timers[i] = scr_timers[parent];
+    scr_timers[parent] = tmp;
+    i = parent;
+  }
+  for (;;) {
+    size_t l = 2 * i + 1, r = 2 * i + 2, min = i;
+    if (l < scr_ntimers && scr_timer_before(&scr_timers[l], &scr_timers[min])) min = l;
+    if (r < scr_ntimers && scr_timer_before(&scr_timers[r], &scr_timers[min])) min = r;
+    if (min == i) break;
+    ScrTimer tmp = scr_timers[i];
+    scr_timers[i] = scr_timers[min];
+    scr_timers[min] = tmp;
+    i = min;
+  }
+}
+
+double scr_set_interval(ScrClosure *cb, double ms) {
+  ms = scr_timer_coerce_ms(ms); /* Node's clamp-and-trunc, like setTimeout */
+  unsigned long id = scr_interval_next_id++;
+  ScrTimer t = {scr_now_ms() + ms, scr_timer_seq++, cb /* ownership moves in */, ms, id, true, ms};
+  scr_timer_push(t);
+  return (double)id;
+}
+
+/* A one-shot timer WITH a clear handle — the island's setTimeout (its
+ * clearTimeout must cancel, unlike the static surface's clear-less
+ * setTimeout). Rides the interval id space so scr_clear_interval serves
+ * both; repeat_ms 0 keeps it one-shot at the firing site. */
+double scr_set_timeout_handle(ScrClosure *cb, double ms) {
+  ms = scr_timer_coerce_ms(ms);
+  unsigned long id = scr_interval_next_id++;
+  ScrTimer t = {scr_now_ms() + ms, scr_timer_seq++, cb /* ownership moves in */, 0, id, true, ms};
+  scr_timer_push(t);
+  return (double)id;
+}
+
+/* process.env-style timer ref bookkeeping: unref() drops a timer from the
+ * loop's liveness count (it still fires if the loop runs on), ref() puts
+ * it back. Node-exact for the static timer surface; NaN/0/absent handles
+ * are tolerated (never a handle). The one-shot setTimeout with NO clear
+ * handle (id 0) cannot be unref'd by id — the compiler routes .unref()
+ * only over handle-returning timers, so id 0 never reaches here. */
+static ScrTimer *scr_timer_find(unsigned long id) {
+  for (size_t i = 0; i < scr_ntimers; i++) {
+    if (scr_timers[i].id == id) return &scr_timers[i];
+  }
+  return NULL;
+}
+
+void scr_timer_unref(double handle) {
+  if (!(handle >= 1)) return;
+  unsigned long id = (unsigned long)handle;
+  if (id == scr_firing_id) {
+    /* Firing interval unref'ing itself: the re-arm carries the flag. */
+    scr_firing_reffed = false;
+    return;
+  }
+  ScrTimer *t = scr_timer_find(id);
+  if (t && t->reffed) {
+    t->reffed = false;
+    if (scr_reffed_timers > 0) scr_reffed_timers--;
+  }
+}
+
+void scr_timer_ref(double handle) {
+  if (!(handle >= 1)) return;
+  unsigned long id = (unsigned long)handle;
+  if (id == scr_firing_id) {
+    scr_firing_reffed = true;
+    return;
+  }
+  ScrTimer *t = scr_timer_find(id);
+  if (t && !t->reffed) {
+    t->reffed = true;
+    scr_reffed_timers++;
+  }
+}
+
+/* refresh(): re-arm to now + the original delay — Node's Timeout.refresh.
+ * A HEAP entry (armed timeout or interval between ticks) re-enters the
+ * heap with a fresh deadline and seq (FIFO against later same-deadline
+ * timers, like a brand-new timer); the timer whose callback is RUNNING
+ * sets the firing flag and the loop re-arms the one-shot after the
+ * callback returns (the interval re-arm already does). A one-shot that
+ * fired on an earlier turn is gone — its closure released — so refresh
+ * of a dead handle is a tolerated no-op (SEMANTICS: Node re-activates
+ * even a fired Timeout; the compiled handle cannot). */
+void scr_timer_refresh(double handle) {
+  if (!(handle >= 1)) return;
+  unsigned long id = (unsigned long)handle;
+  if (id == scr_firing_id) {
+    scr_firing_refresh = true;
+    return;
+  }
+  for (size_t i = 0; i < scr_ntimers; i++) {
+    if (scr_timers[i].id == id) {
+      ScrTimer t = scr_timers[i];
+      scr_timer_remove_at(i);
+      t.deadline_ms = scr_now_ms() + t.delay_ms;
+      t.seq = scr_timer_seq++;
+      scr_timer_push(t);
+      return;
+    }
+  }
+}
+
+/* hasRef(): true iff the handle names a live, still-reffed timer. */
+bool scr_timer_has_ref(double handle) {
+  if (!(handle >= 1)) return false;
+  unsigned long id = (unsigned long)handle;
+  if (id == scr_firing_id) return scr_firing_reffed;
+  ScrTimer *t = scr_timer_find(id);
+  return t != NULL && t->reffed;
+}
+
+/* ── process.getActiveResourcesInfo (the loop's own bookkeeping) ──────
+ * 'Timeout' per armed heap timer — plus the firing, uncleared one (Node
+ * counts a Timeout as active while its callback runs, until clearTimeout
+ * drops it) — and 'Immediate' per queued, unfired immediate (a FIRED
+ * immediate no longer counts, Node's current answer inside the
+ * callback). Resource kinds this runtime does not model as loop handles
+ * (TCP wraps, FS requests, ...) are absent — SEMANTICS.md names the
+ * divergence. Result +1. */
+static size_t scr_pending_immediates; /* defined below with the queue */
+ScrArr *scr_active_resources(void) {
+  ScrArr *arr = scr_arr_new(SCR_ELEM_STR, 8);
+  size_t timeouts = scr_ntimers + ((scr_firing_id != 0 && !scr_firing_cleared) ? 1 : 0);
+  for (size_t i = 0; i < timeouts; i++) scr_arr_push_ref(arr, scr_str_new("Timeout", 7));
+  for (size_t i = 0; i < scr_pending_immediates; i++) scr_arr_push_ref(arr, scr_str_new("Immediate", 9));
+  return arr;
+}
+
+/* ── process.nextTick (the user tick queue) ───────────────────────────
+ * A FIFO of user callbacks drained BEFORE promise jobs at every loop
+ * checkpoint, to joint exhaustion with them — Node's tick-then-microtask
+ * order. Pending ticks are always-ready work: the loop neither sleeps
+ * nor exits while any exist. Ticks scheduled from 'exit' listeners (or
+ * left queued on the uncaught paths) never run — the teardown releases
+ * them, like Node dropping the queue at exit. */
+typedef struct ScrNtick {
+  ScrClosure *cb;             /* owned; NULL for a raw C-hook entry */
+  ScrFsRenameFn error_cb;     /* process-write success adapter, or NULL */
+  void (*raw)(void);          /* hook when cb == NULL (stream markers) */
+  struct ScrNtick *next;
+} ScrNtick;
+
+static ScrNtick *scr_nt_head = NULL;
+static ScrNtick *scr_nt_tail = NULL;
+
+/* Public: releases every queued tick without running it — the loop-exit
+ * teardown below AND the exit-listener runner (scr_events.c) both call
+ * it, because 'exit' listeners run AFTER the loop's teardown and may
+ * enqueue ticks that must never run (Node) yet must not leak. */
+
+void scr_next_tick(ScrClosure *cb /*moves*/) {
+  ScrNtick *t = calloc(1, sizeof *t);
+  if (!t) scr_oom();
+  t->cb = cb;
+  if (scr_nt_tail) scr_nt_tail->next = t;
+  else scr_nt_head = t;
+  scr_nt_tail = t;
+}
+
+/* stdout/stderr write completions are Node nextTicks too, but their
+ * callback receives the success `null` argument. Reuse the error-first
+ * adapter ABI also used by fs.rename: emitted adapters construct the
+ * program's Error | null union, and NULL selects its null arm. */
+void scr_process_write_callback(ScrClosure *cb /*moves*/, ScrFsRenameFn fn) {
+  ScrNtick *t = calloc(1, sizeof *t);
+  if (!t) scr_oom();
+  t->cb = cb;
+  t->error_cb = fn;
+  if (scr_nt_tail) scr_nt_tail->next = t;
+  else scr_nt_head = t;
+  scr_nt_tail = t;
+}
+
+/* A RAW C-hook tick: the stream unit enqueues one marker per deferred
+ * stream emission so those emissions interleave with user nextTicks in
+ * true FIFO order — in Node they ARE nextTicks (resume_, emitReadable_,
+ * endReadableNT, ...). The hook dispatches exactly one stream tick;
+ * teardown just drops markers (the stream queue owns its entries). */
+void scr_next_tick_raw(void (*fn)(void)) {
+  ScrNtick *t = calloc(1, sizeof *t);
+  if (!t) scr_oom();
+  t->raw = fn;
+  if (scr_nt_tail) scr_nt_tail->next = t;
+  else scr_nt_head = t;
+  scr_nt_tail = t;
+}
+
+void scr_nticks_teardown(void) {
+  while (scr_nt_head != NULL) {
+    ScrNtick *t = scr_nt_head;
+    scr_nt_head = t->next;
+    if (t->cb) scr_closure_release(t->cb);
+    free(t);
+  }
+  scr_nt_tail = NULL;
+}
+
+/* Releases every armed timer — the island teardown calls this before the
+ * engine dies so closures holding engine callbacks free first and the
+ * counting allocator's zero-live audit holds (the loop only exits with
+ * entries still armed on the uncaught/unhandled paths). */
+static void scr_immediates_teardown(void);
+void scr_timers_teardown(void) {
+  for (size_t i = 0; i < scr_ntimers; i++) scr_closure_release(scr_timers[i].cb);
+  scr_ntimers = 0;
+  scr_reffed_timers = 0;
+  scr_immediates_teardown();
+  scr_nticks_teardown();
+}
+
+void scr_clear_interval(double handle) {
+  if (!(handle >= 1)) return; /* NaN/0/negative: never a handle; Node tolerates */
+  unsigned long id = (unsigned long)handle;
+  if (id == scr_firing_id) {
+    scr_firing_cleared = true; /* the run loop drops the callback */
+    return;
+  }
+  for (size_t i = 0; i < scr_ntimers; i++) {
+    if (scr_timers[i].id == id) {
+      scr_closure_release(scr_timers[i].cb);
+      scr_timer_remove_at(i);
+      return;
+    }
+  }
+}
+
+/* ── immediates (Node's check phase) ──────────────────────────────────
+ * setImmediate callbacks run once per loop turn, AFTER due timers, in
+ * FIFO order — and immediates queued while the phase runs wait for the
+ * NEXT turn (the phase snapshots its end index). The queue is an
+ * append-only array with a head cursor; clearImmediate NULLs the slot in
+ * place (order and indices stay stable mid-phase). Ids ride their own
+ * space — clearTimeout of an Immediate is a no-op, like Node. */
+typedef struct {
+  unsigned long id;
+  ScrClosure *cb; /* owned; NULL once fired or cleared */
+  bool reffed;    /* keeps the loop alive; unref() clears it (Node semantics) */
+} ScrImmediate;
+
+static ScrImmediate *scr_immediates = NULL;
+static size_t scr_nimmediates = 0, scr_immediates_head = 0, scr_immediates_cap = 0;
+static unsigned long scr_immediate_seq = 0;
+/* Pending (queued, uncleared) immediates and the reffed subset — the
+ * loop's no-sleep signal and liveness count respectively. */
+static size_t scr_pending_immediates = 0;
+static size_t scr_reffed_immediates = 0;
+
+double scr_set_immediate(ScrClosure *cb) {
+  if (scr_nimmediates == scr_immediates_cap) {
+    scr_immediates_cap = scr_immediates_cap ? scr_immediates_cap * 2 : 16;
+    scr_immediates = realloc(scr_immediates, scr_immediates_cap * sizeof *scr_immediates);
+    if (!scr_immediates) scr_oom();
+  }
+  unsigned long id = ++scr_immediate_seq;
+  scr_immediates[scr_nimmediates].id = id;
+  scr_immediates[scr_nimmediates].cb = cb; /* ownership moves in */
+  scr_immediates[scr_nimmediates].reffed = true;
+  scr_nimmediates++;
+  scr_pending_immediates++;
+  scr_reffed_immediates++;
+  return (double)id;
+}
+
+static ScrImmediate *scr_immediate_find(unsigned long id) {
+  for (size_t i = scr_immediates_head; i < scr_nimmediates; i++) {
+    if (scr_immediates[i].id == id && scr_immediates[i].cb != NULL) return &scr_immediates[i];
+  }
+  return NULL;
+}
+
+void scr_clear_immediate(double handle) {
+  if (!(handle >= 1)) return; /* NaN/0/negative: never a handle; Node tolerates */
+  ScrImmediate *im = scr_immediate_find((unsigned long)handle);
+  if (im == NULL) return; /* already fired/cleared: no-op, like Node */
+  scr_closure_release(im->cb);
+  im->cb = NULL;
+  if (scr_pending_immediates > 0) scr_pending_immediates--;
+  if (im->reffed) {
+    im->reffed = false;
+    if (scr_reffed_immediates > 0) scr_reffed_immediates--;
+  }
+}
+
+/* Immediate.unref()/ref()/hasRef(): the Timeout trio's exact story over
+ * the immediate queue. A fired or cleared handle is a tolerated no-op
+ * (Node's Immediate methods on a dead handle do nothing). */
+void scr_immediate_unref(double handle) {
+  if (!(handle >= 1)) return;
+  ScrImmediate *im = scr_immediate_find((unsigned long)handle);
+  if (im && im->reffed) {
+    im->reffed = false;
+    if (scr_reffed_immediates > 0) scr_reffed_immediates--;
+  }
+}
+
+void scr_immediate_ref(double handle) {
+  if (!(handle >= 1)) return;
+  ScrImmediate *im = scr_immediate_find((unsigned long)handle);
+  if (im && !im->reffed) {
+    im->reffed = true;
+    scr_reffed_immediates++;
+  }
+}
+
+bool scr_immediate_has_ref(double handle) {
+  if (!(handle >= 1)) return false;
+  ScrImmediate *im = scr_immediate_find((unsigned long)handle);
+  return im != NULL && im->reffed;
+}
+
+/* The teardown sweep's immediate arm (called from scr_timers_teardown):
+ * unref'd immediates left queued at normal loop exit never fire (Node),
+ * so their closures release here or the RC audit counts them as leaks. */
+static void scr_immediates_teardown(void) {
+  for (size_t i = scr_immediates_head; i < scr_nimmediates; i++) {
+    if (scr_immediates[i].cb != NULL) scr_closure_release(scr_immediates[i].cb);
+  }
+  scr_nimmediates = 0;
+  scr_immediates_head = 0;
+  scr_pending_immediates = 0;
+  scr_reffed_immediates = 0;
+}
+
+/* Unhandled rejections: rejected promises retained here until observed. */
+static ScrPromise **scr_maybe_unhandled = NULL;
+static size_t scr_nunhandled = 0, scr_unhandled_cap = 0;
+
+static void scr_track_rejection(ScrPromise *p) {
+  if (scr_nunhandled == scr_unhandled_cap) {
+    scr_unhandled_cap = scr_unhandled_cap ? scr_unhandled_cap * 2 : 8;
+    scr_maybe_unhandled = realloc(scr_maybe_unhandled, scr_unhandled_cap * sizeof *scr_maybe_unhandled);
+    if (!scr_maybe_unhandled) scr_oom();
+  }
+  scr_maybe_unhandled[scr_nunhandled++] = scr_promise_retain(p);
+}
+
+/* ── context switching (ASan-annotated) ───────────────────────────────── */
+
+#ifdef SCR_ASAN_FIBERS
+static void *scr_main_fake_stack; /* fake-stack slot for the main context */
+#endif
+
+#ifdef _WIN32
+/* The current context's handle, converting the main thread to a fiber on
+ * first need (SwitchToFiber requires the caller BE a fiber; the runtime is
+ * single-threaded, so one lazy conversion covers the process). Also the
+ * loop-home keeper: whenever main is the caller this IS scr_loop_ctx's
+ * value — assigned at the conversion so parked fibers can always switch
+ * back to it. */
+static void *scr_win_self(void) {
+  static bool converted = false;
+  if (!converted) {
+    scr_loop_ctx = ConvertThreadToFiber(NULL);
+    if (scr_loop_ctx == NULL) {
+      fputs("scriptc: ConvertThreadToFiber failed\n", stderr);
+      abort();
+    }
+    converted = true;
+  }
+  return GetCurrentFiber();
+}
+#endif
+
+static void scr_switch(ScrCtx *from, ScrCtx *to, ScrFiber *to_fiber) {
+#ifdef SCR_ASAN_FIBERS
+  ScrFiber *from_fiber = scr_current;
+#endif
+  scr_current = to_fiber;
+  scr_exc_swap_cell(to_fiber ? &to_fiber->exc : NULL);
+  scr_als_active = to_fiber ? &to_fiber->als : &scr_als_main_slot;
+#ifdef _WIN32
+  /* SwitchToFiber snapshots the outgoing fiber's state inside its own
+   * fiber object — `from` has nothing to record. */
+  (void)from;
+  SwitchToFiber(*to);
+#elif defined(__wasi__)
+  (void)from;
+  (void)to;
+  scr_trap("scriptc: internal error: native stack switch reached on WASI\n");
+#elif defined(SCR_ASAN_FIBERS)
+  void **save = from_fiber ? &from_fiber->fake_stack : &scr_main_fake_stack;
+  const void *bottom = to_fiber ? to_fiber->stack : NULL;
+  size_t size = to_fiber ? SCR_FIBER_STACK : 0;
+  __sanitizer_start_switch_fiber(save, bottom, size);
+  swapcontext(from, to);
+  const void *old_bottom;
+  size_t old_size;
+  __sanitizer_finish_switch_fiber(
+      scr_current ? scr_current->fake_stack : scr_main_fake_stack, &old_bottom, &old_size);
+#else
+  swapcontext(from, to);
+#endif
+}
+
+static void scr_promise_settle_wake(ScrPromise *p);
+
+/* Copies src's settlement into a still-pending dst (payload RETAINED —
+ * src can settle other destinations and still be awaited) and wakes dst's
+ * waiters. A rejection consumed this way counts as HANDLED on src (a
+ * combinator attached a handler, like Node's race). The Promise.race
+ * same-inner-type adapter and the rejection path both land here. */
+static void scr_promise_settle_from(ScrPromise *dst, ScrPromise *src) {
+  if (dst->state == SCR_PROM_PENDING) {
+    dst->state = src->state;
+    dst->payload_kind = src->payload_kind;
+    dst->f64 = src->f64;
+    dst->b = src->b;
+    dst->retain_fn = src->retain_fn;
+    dst->release_fn = src->release_fn;
+    dst->trace_fn = src->trace_fn;
+    dst->payload = NULL;
+    if (src->payload_kind == SCR_EXC_STR && src->payload) {
+      dst->payload = scr_str_retain((ScrStr *)src->payload);
+    } else if ((src->payload_kind == SCR_EXC_REF || src->payload_kind == SCR_EXC_OBJ ||
+                src->payload_kind == SCR_EXC_PRIMITIVE_REF) &&
+               src->payload) {
+      dst->payload = src->retain_fn(src->payload);
+    }
+    scr_promise_settle_wake(dst);
+  }
+  if (src->state == SCR_PROM_REJECTED) scr_prom_observe(src);
+}
+
+/* The emitted same-type Promise.race adapter (see raceAdapterFor). */
+void scr_promise_adapt_copy(ScrPromise *dst, ScrPromise *src) {
+  scr_promise_settle_from(dst, src);
+}
+
+/* One Promise.all entry settling: a fulfillment stores its payload into
+ * the values array at the entry's INPUT index (input order regardless of
+ * settlement order) and the LAST missing fulfillment fulfills the result
+ * with the array (void-element all fulfills void); a rejection copies raw
+ * into the result — the first one in SETTLEMENT order wins through the
+ * destination's own pending check, and later ones still count as HANDLED
+ * on their entries (settle_from marks them observed), exactly Node's
+ * subscribe-to-everything behavior. */
+static void scr_promise_all_settle(ScrAllState *st, ScrPromise *result, size_t idx,
+                                    ScrPromise *src) {
+  if (src->state == SCR_PROM_FULFILLED) {
+    if (st->store) st->store(st->values, (double)idx, src);
+    if (--st->remaining == 0 && result->state == SCR_PROM_PENDING) {
+      if (st->values) {
+        scr_promise_fulfill_ref(result, scr_arr_retain(st->values), scr_arr_retain_v,
+                                 scr_arr_release_v,
+                                 st->values->elem_trace ? scr_arr_trace_v : NULL);
+      } else {
+        scr_promise_fulfill_void(result);
+      }
+    }
+  } else {
+    scr_promise_settle_from(result, src);
+  }
+}
+
+/* Settle helpers wake waiters into the microtask queue and run the
+ * combinator callbacks: adapters for fulfillments (they convert the
+ * payload to the destination's inner type), the raw copy for rejections
+ * (reasons are dynamically tagged). first-settle-wins rides the
+ * destination's own pending check. Promise.all entries dispatch through
+ * their shared countdown state instead of the adapt pair. */
+static void scr_promise_settle_wake(ScrPromise *p) {
+  for (size_t i = 0; i < p->nwaiters; i++) scr_ready_push(p->waiters[i]);
+  p->nwaiters = 0;
+  for (size_t i = 0; i < p->ncbs; i++) {
+    if (p->cbs[i].all) {
+      scr_promise_all_settle(p->cbs[i].all, p->cbs[i].dst, p->cbs[i].all_idx, p);
+      scr_promise_all_state_release(p->cbs[i].all);
+    } else if (p->state == SCR_PROM_FULFILLED) {
+      p->cbs[i].adapt(p->cbs[i].dst, p);
+    } else {
+      scr_promise_settle_from(p->cbs[i].dst, p);
+    }
+    scr_promise_release(p->cbs[i].dst);
+  }
+  p->ncbs = 0;
+  if (p->state == SCR_PROM_REJECTED) scr_track_rejection(p);
+}
+
+/* ── Promise.race ─────────────────────────────────────────────────────
+ * The compiler emits: a fresh result promise, then one race_add per
+ * entry. A settled entry settles the result immediately (first add
+ * wins); pending entries park a callback waiter that fires inside the
+ * entry's settle. Fibers awaiting the result still wake through the
+ * microtask queue. Losing entries keep their own settlements (payloads
+ * are retained, not moved). */
+
+void scr_promise_race_add(ScrPromise *race, ScrPromise *in,
+                           void (*adapt)(ScrPromise *dst, ScrPromise *src)) {
+  if (in->state != SCR_PROM_PENDING) {
+    if (in->state == SCR_PROM_FULFILLED) adapt(race, in);
+    else scr_promise_settle_from(race, in);
+    return;
+  }
+  if (in->ncbs == in->cbs_cap) {
+    in->cbs_cap = in->cbs_cap ? in->cbs_cap * 2 : 4;
+    in->cbs = realloc(in->cbs, in->cbs_cap * sizeof *in->cbs);
+    if (!in->cbs) scr_oom();
+  }
+  in->cbs[in->ncbs].adapt = adapt;
+  in->cbs[in->ncbs].dst = scr_promise_retain(race);
+  in->cbs[in->ncbs].all = NULL;
+  in->cbs[in->ncbs].all_idx = 0;
+  in->ncbs++;
+}
+
+/* ── Promise.all ──────────────────────────────────────────────────────
+ * BORROWS the entries array `ps` and the pre-capacity values array
+ * `values` (NULL for void elements), retains what it keeps, and returns
+ * the result promise +1. `values` arrives EMPTY with cap >= ps->len (the
+ * emitted construction passes the length) and is pre-sized here: every
+ * slot's zero bits are the per-kind placeholder (0.0 / false / NULL),
+ * legal to overwrite through the store helpers and to release. Entries
+ * already settled settle inline (a rejection settles the result NOW —
+ * first rejection in settlement order — and later rejected entries are
+ * still marked handled); pending entries park a countdown waiter. The
+ * empty array fulfills immediately (awaiters take the settled-await
+ * microtask hop, like Node's one-tick resolve). */
+ScrPromise *scr_promise_all(ScrArr *ps, ScrArr *values,
+                             void (*store)(ScrArr *a, double i, ScrPromise *src)) {
+  size_t n = ps->len;
+  ScrPromise *result = scr_promise_new();
+  if (values && n > 0) {
+    /* Pre-size as holes. Promise settlement writes the present slots through
+     * the normal setters; this also remains safe when n is sparse-sized and
+     * the array intentionally has no dense allocation. */
+    scr_arr_set_len(values, (double)n);
+  }
+  ScrAllState *st = malloc(sizeof *st);
+  if (!st) scr_oom();
+  st->rc = 1; /* the builder's reference, dropped at the end */
+  st->remaining = n;
+  st->values = values ? scr_arr_retain(values) : NULL;
+  st->store = store;
+  for (size_t i = 0; i < n; i++) {
+    ScrPromise *in = (ScrPromise *)scr_arr_get_ref(ps, (double)i); /* +1 */
+    if (in->state != SCR_PROM_PENDING) {
+      scr_promise_all_settle(st, result, i, in);
+    } else {
+      if (in->ncbs == in->cbs_cap) {
+        in->cbs_cap = in->cbs_cap ? in->cbs_cap * 2 : 4;
+        in->cbs = realloc(in->cbs, in->cbs_cap * sizeof *in->cbs);
+        if (!in->cbs) scr_oom();
+      }
+      in->cbs[in->ncbs].adapt = NULL;
+      in->cbs[in->ncbs].dst = scr_promise_retain(result);
+      in->cbs[in->ncbs].all = st;
+      in->cbs[in->ncbs].all_idx = i;
+      in->ncbs++;
+      st->rc++;
+    }
+    scr_promise_release(in);
+  }
+  /* n == 0, or every entry was already fulfilled inline. */
+  if (st->remaining == 0 && result->state == SCR_PROM_PENDING) {
+    if (st->values) {
+      scr_promise_fulfill_ref(result, scr_arr_retain(st->values), scr_arr_retain_v,
+                               scr_arr_release_v,
+                               st->values->elem_trace ? scr_arr_trace_v : NULL);
+    } else {
+      scr_promise_fulfill_void(result);
+    }
+  }
+  scr_promise_all_state_release(st);
+  return result;
+}
+
+/* Per-element-kind store helpers for the emitted Promise.all: write the
+ * entry's fulfillment payload into the values array at its input index.
+ * The payload accessors return retained/by-value (losing a reference is
+ * impossible: set_* takes ownership and releases the zero placeholder). */
+void scr_promise_all_store_f64(ScrArr *a, double i, ScrPromise *src) {
+  scr_arr_set_f64(a, i, scr_promise_payload_f64(src));
+}
+void scr_promise_all_store_bool(ScrArr *a, double i, ScrPromise *src) {
+  scr_arr_set_bool(a, i, scr_promise_payload_bool(src));
+}
+void scr_promise_all_store_str(ScrArr *a, double i, ScrPromise *src) {
+  scr_arr_set_ref(a, i, scr_promise_payload_str(src));
+}
+void scr_promise_all_store_ref(ScrArr *a, double i, ScrPromise *src) {
+  scr_arr_set_ref(a, i, scr_promise_payload_ref(src));
+}
+
+/* Fulfillment payload accessors for the emitted race adapters — all
+ * retained/by-value (the source keeps its settlement). */
+double scr_promise_payload_f64(ScrPromise *p) { return p->f64; }
+bool scr_promise_payload_bool(ScrPromise *p) { return p->b; }
+ScrStr *scr_promise_payload_str(ScrPromise *p) {
+  return scr_str_retain((ScrStr *)p->payload);
+}
+/* Thin views for the gated dyn-async TU (scr_async_dyn.c): the payload
+ * KIND, whether a REF payload is a dyn value (the dyn adapters), and the
+ * settled-await primitive (park/hop + rejection re-throw; true =
+ * fulfilled). */
+static bool scr_await_settled(ScrPromise *p); /* defined with the await family */
+int scr_promise_payload_kind(const ScrPromise *p) { return (int)p->payload_kind; }
+bool scr_promise_payload_is_dyn(const ScrPromise *p) { return p->retain_fn == scr_dyn_retain_v; }
+double scr_promise_payload_num(const ScrPromise *p) { return p->f64; }
+bool scr_promise_payload_flag(const ScrPromise *p) { return p->b; }
+bool scr_promise_await_settled(ScrPromise *p) { return scr_await_settled(p); }
+void scr_promise_payload_release(const ScrPromise *p, void *v) { p->release_fn(v); }
+
+void *scr_promise_payload_ref(ScrPromise *p) {
+  return p->payload ? p->retain_fn(p->payload) : NULL;
+}
+
+/* MOVES a pending exception out of `cell` into `p` as its rejection payload
+ * (the cell resets to NONE). Shared by fiber completion and the new-Promise
+ * executor runners; callers wake waiters themselves. */
+static void scr_promise_reject_from_cell(ScrPromise *p, ScrExcCell *cell) {
+  p->state = SCR_PROM_REJECTED;
+  p->payload_kind = cell->kind;
+  p->f64 = cell->f64;
+  p->b = cell->b;
+  p->payload = cell->payload;
+  p->retain_fn = cell->retain_fn;
+  p->release_fn = cell->release_fn;
+  p->trace_fn = cell->trace_fn;
+  cell->kind = SCR_EXC_NONE;
+  cell->payload = NULL;
+  cell->trace_fn = NULL;
+}
+
+static void scr_fiber_finish(ScrFiber *self) {
+  if (self->gen != NULL) {
+    /* Generator completion: the emitted trampoline already stored the
+     * completion value (or consumed the GENRET sentinel); a real body
+     * exception stays pending in this fiber's cell — the consumer-side
+     * resume moves it into the resumer's cell. No promise to settle. */
+    self->done = true;
+    return;
+  }
+  ScrPromise *p = self->promise;
+  if (scr_exc_pending()) {
+    /* The body's exception escaped: the promise rejects with the payload
+     * (moved from the fiber's cell into the promise). */
+    scr_promise_reject_from_cell(p, &self->exc);
+  }
+  /* Fulfillment payload was stored by the trampoline before finishing. */
+  scr_promise_settle_wake(p);
+  self->done = true;
+}
+
+#ifdef _WIN32
+static void CALLBACK scr_trampoline(void *param) {
+  (void)param;
+#else
+static void scr_trampoline(void) {
+#endif
+#ifdef SCR_ASAN_FIBERS
+  const void *ob;
+  size_t os;
+  __sanitizer_finish_switch_fiber(scr_current->fake_stack, &ob, &os);
+#endif
+  ScrFiber *self = scr_current;
+  self->entry(self, self->argpack);
+  scr_fiber_finish(self);
+  /* Dead fiber: hop back to whoever resumed us. The loop frees us. */
+  scr_switch(&self->ctx, self->return_to, NULL);
+  /* unreachable */
+}
+
+#if !defined(_WIN32) && !defined(__wasi__)
+/* Apple Silicon's makecontext clears all of uc_stack before installing the
+ * initial registers. Passing the whole mapping eagerly commits 256 KiB
+ * (8 MiB under ASan) per call, defeating the lazy allocation above. Its
+ * zero-argument startup frame only needs the top of the downward-growing
+ * stack. Initialize that window, then publish the FULL stack bounds before
+ * any execution or sanitizer switch. The stack top and saved SP are the
+ * same as when initializing the full mapping; ordinary calls can still
+ * use all of SCR_FIBER_STACK.
+ * See libplatform/src/ucontext/generic/makecontext.c's arm64 implementation. */
+static void scr_fiber_context_init(ScrFiber *f) {
+  f->stack = scr_fiber_stack_new();
+  if (getcontext(&f->ctx) != 0) scr_trap("scriptc: getcontext failed\n");
+  f->ctx.uc_stack.ss_sp = f->stack;
+  f->ctx.uc_stack.ss_size = SCR_FIBER_STACK;
+  f->ctx.uc_stack.ss_flags = 0;
+  f->ctx.uc_link = NULL;
+#if defined(__APPLE__) && defined(__aarch64__)
+  const size_t bootstrap = 16 * 1024;
+  f->ctx.uc_stack.ss_sp = f->stack + SCR_FIBER_STACK - bootstrap;
+  f->ctx.uc_stack.ss_size = bootstrap;
+#endif
+  makecontext(&f->ctx, scr_trampoline, 0);
+  f->ctx.uc_stack.ss_sp = f->stack;
+  f->ctx.uc_stack.ss_size = SCR_FIBER_STACK;
+}
+#endif
+
+/* Frees a finished fiber's execution resources (the promise release and
+ * bookkeeping stay at the call sites). Windows: DeleteFiber tears down the
+ * fiber object and its stack — legal here because a finished fiber has
+ * switched away and can never be current again. */
+static void scr_fiber_destroy(ScrFiber *f) {
+#ifdef _WIN32
+  DeleteFiber(f->ctx);
+#elif defined(__wasi__)
+  if (f->coro != NULL) scr_wasi_coro_destroy(f->coro);
+#endif
+  scr_als_ctx_release(f->als);
+#if !defined(_WIN32) && !defined(__wasi__)
+  scr_fiber_stack_free(f->stack);
+#else
+  free(f->stack);
+#endif
+  free(f);
+}
+
+/* Spawns and EAGERLY runs an async body until its first suspension (JS's
+ * synchronous-prefix rule). Returns the fiber's promise, +1. */
+ScrPromise *scr_async_spawn(void (*entry)(ScrFiber *, void *), void *argpack) {
+  ScrFiber *f = calloc(1, sizeof *f);
+  if (!f) scr_oom();
+  f->promise = scr_promise_new();
+  f->entry = entry;
+  f->argpack = argpack;
+  /* AsyncLocalStorage: the child runs in the SPAWNER's context (Node's
+   * init-time capture); snapshots are immutable, so a retain suffices. */
+  f->als = scr_als_ctx_retain(*scr_als_active);
+  scr_fibers_live++;
+
+#ifdef _WIN32
+  /* Commit one page initially; Windows commits further pages as the stack
+   * grows. Keep the executable's stack reserve, as with CreateFiber,
+   * without committing SCR_FIBER_STACK for every idle fiber. */
+  ScrCtx here = scr_win_self();
+  f->ctx = CreateFiberEx(4096, 0, 0, scr_trampoline, NULL);
+  if (f->ctx == NULL) scr_oom();
+#elif defined(__wasi__)
+  ScrFiber *spawner = scr_current;
+  scr_current = f;
+  scr_exc_swap_cell(&f->exc);
+  scr_als_active = &f->als;
+  entry(f, argpack); /* eager prefix; returns at suspend/final suspend */
+  scr_current = spawner;
+  scr_exc_swap_cell(spawner ? &spawner->exc : NULL);
+  scr_als_active = spawner ? &spawner->als : &scr_als_main_slot;
+  ScrPromise *result = scr_promise_retain(f->promise);
+  if (f->done) {
+    scr_promise_release(f->promise);
+    scr_fiber_destroy(f);
+    scr_fibers_live--;
+  }
+  return result;
+#else
+  scr_fiber_context_init(f);
+
+  ucontext_t here;
+#endif
+#ifndef __wasi__
+  f->return_to = &here;
+  ScrFiber *spawner = scr_current;
+  scr_switch(&here, &f->ctx, f);
+  /* back: the fiber suspended or finished. The switch back targeted NULL
+   * (the child doesn't know who spawned it), which pointed both the current
+   * fiber AND the exception machinery at main — restore the spawner's cell
+   * too, or a fiber that eagerly spawned another keeps throwing/catching
+   * against main's cell and completes with its own cell empty (a phantom
+   * `undefined` rejection while the real payload leaks). */
+  scr_current = spawner;
+  scr_exc_swap_cell(spawner ? &spawner->exc : NULL);
+  scr_als_active = spawner ? &spawner->als : &scr_als_main_slot;
+  ScrPromise *result = scr_promise_retain(f->promise);
+  if (f->done) {
+    scr_promise_release(f->promise);
+    scr_fiber_destroy(f);
+    scr_fibers_live--;
+  }
+  return result;
+#endif
+}
+
+/* A runtime-authored C continuation whose first operation awaits one known
+ * dependency. Native fibers can park normally. wasm32 has no resumable C
+ * stack, so queue the fiber only after the dependency settles; the entry's
+ * await then extracts the result without attempting a stack switch. The
+ * READY queue supplies the settled-await microtask hop in both cases. */
+ScrPromise *scr_async_spawn_after(ScrPromise *dependency,
+                                  void (*entry)(ScrFiber *, void *),
+                                  void *argpack) {
+#ifndef __wasi__
+  (void)dependency;
+  return scr_async_spawn(entry, argpack);
+#else
+  ScrFiber *f = calloc(1, sizeof *f);
+  if (!f) scr_oom();
+  f->promise = scr_promise_new();
+  f->entry = entry;
+  f->argpack = argpack;
+  f->als = scr_als_ctx_retain(*scr_als_active);
+  scr_fibers_live++;
+
+  if (dependency->state == SCR_PROM_PENDING) {
+    if (dependency->nwaiters == dependency->waiters_cap) {
+      dependency->waiters_cap = dependency->waiters_cap ? dependency->waiters_cap * 2 : 4;
+      dependency->waiters = realloc(
+          dependency->waiters, dependency->waiters_cap * sizeof *dependency->waiters);
+      if (!dependency->waiters) scr_oom();
+    }
+    dependency->waiters[dependency->nwaiters++] = f;
+  } else {
+    scr_ready_push(f);
+  }
+  return scr_promise_retain(f->promise);
+#endif
+}
+
+/* Parks the current fiber on `p` until it settles. Only fibers await. */
+static void scr_await_park(ScrPromise *p) {
+  ScrFiber *self = scr_current;
+  if (!self) {
+    fputs("scriptc: internal error: await outside an async function\n", stderr);
+    abort();
+  }
+  if (p->nwaiters == p->waiters_cap) {
+    p->waiters_cap = p->waiters_cap ? p->waiters_cap * 2 : 4;
+    p->waiters = realloc(p->waiters, p->waiters_cap * sizeof *p->waiters);
+    if (!p->waiters) scr_oom();
+  }
+  p->waiters[p->nwaiters++] = self;
+  scr_switch(&self->ctx, self->return_to, NULL);
+  /* Resumed by the loop: return_to must now point at the loop's context. */
+}
+
+#ifdef __wasi__
+/* The emitted coroutine suspends immediately after these preparation
+ * calls. Settled promises still queue one ready turn; pending promises own
+ * the continuation through their waiter list. */
+void scr_wasi_await_prepare(ScrFiber *self, ScrPromise *p) {
+  if (p->state == SCR_PROM_PENDING) {
+    if (p->nwaiters == p->waiters_cap) {
+      p->waiters_cap = p->waiters_cap ? p->waiters_cap * 2 : 4;
+      p->waiters = realloc(p->waiters, p->waiters_cap * sizeof *p->waiters);
+      if (!p->waiters) scr_oom();
+    }
+    p->waiters[p->nwaiters++] = self;
+  } else {
+    scr_ready_push(self);
+  }
+}
+
+void scr_wasi_await_hop_prepare(ScrFiber *self) { scr_ready_push(self); }
+
+bool scr_wasi_module_await_prepare(ScrFiber *self, ScrPromise *p) {
+  if (p->state != SCR_PROM_PENDING) return false;
+  scr_wasi_await_prepare(self, p);
+  return true;
+}
+
+void scr_wasi_async_finish(ScrFiber *self) { scr_fiber_finish(self); }
+
+void scr_wasi_gen_finish(ScrFiber *self) {
+  if (scr_exc_genret_pending()) {
+    scr_exc_clear();
+    scr_gen_ret_to_out(self->gen);
+  }
+  scr_fiber_finish(self);
+}
+#endif
+
+/* One microtask hop: park the current fiber on the READY queue itself and
+ * yield. JS's `await` ALWAYS suspends — even on an already-settled promise
+ * the continuation runs as a microtask — so without this hop an awaiter of
+ * a settled promise would continue synchronously inside the spawner,
+ * observably earlier than Node (corpus 1428 pins the ordering). */
+static void scr_await_yield(void) {
+  ScrFiber *self = scr_current;
+  if (!self) {
+    fputs("scriptc: internal error: await outside an async function\n", stderr);
+    abort();
+  }
+  scr_ready_push(self);
+  scr_switch(&self->ctx, self->return_to, NULL);
+}
+
+/* The emitted promise-or-absent await's unit arm (`await u` where u holds
+ * undefined/null): JS awaits non-thenables through exactly one microtask
+ * turn — the same hop a settled promise takes. */
+void scr_await_hop(void) { scr_await_yield(); }
+
+/* Copies a rejection into the active execution context's exception cell.
+ * The payload is RETAINED, not moved — a promise can be awaited more than
+ * once, and the executable's top-level completion probe consumes the same
+ * settlement from the main stack after the loop drains. */
+static void scr_promise_rethrow(ScrPromise *p) {
+  switch (p->payload_kind) {
+  case SCR_EXC_F64: scr_throw_f64(p->f64); break;
+  case SCR_EXC_BOOL: scr_throw_bool(p->b); break;
+  case SCR_EXC_STR: scr_throw_str(scr_str_retain((ScrStr *)p->payload)); break;
+  case SCR_EXC_REF: scr_throw_ref(p->retain_fn(p->payload), p->retain_fn, p->release_fn, p->trace_fn); break;
+  case SCR_EXC_PRIMITIVE_REF: scr_throw_primitive_ref(p->retain_fn(p->payload), p->retain_fn, p->release_fn, p->trace_fn); break;
+  case SCR_EXC_OBJ: scr_throw_obj(p->retain_fn(p->payload), p->retain_fn, p->release_fn, p->trace_fn); break;
+  case SCR_EXC_NONE:
+  case SCR_EXC_GENRET: /* unreachable: the sentinel never settles a promise */
+    scr_throw_str(scr_str_new("undefined", 9));
+    break;
+  }
+}
+
+/* Await result extraction. Rejection re-throws into the awaiter. */
+static bool scr_await_settled(ScrPromise *p) {
+#ifdef __wasi__
+  if (p->state == SCR_PROM_PENDING) {
+    fputs("scriptc: internal error: resumed before awaited promise settled\n", stderr);
+    abort();
+  }
+#else
+  if (p->state != SCR_PROM_PENDING) scr_await_yield();
+  while (p->state == SCR_PROM_PENDING) scr_await_park(p);
+#endif
+  scr_prom_observe(p);
+  if (p->state == SCR_PROM_REJECTED) {
+    scr_promise_rethrow(p);
+    return false;
+  }
+  return true;
+}
+
+double scr_await_f64(ScrPromise *p) {
+  return scr_await_settled(p) ? p->f64 : 0;
+}
+bool scr_await_bool(ScrPromise *p) {
+  return scr_await_settled(p) ? p->b : false;
+}
+void *scr_await_ref(ScrPromise *p) {
+  return scr_await_settled(p) && p->payload ? p->retain_fn(p->payload) : NULL;
+}
+void scr_await_void(ScrPromise *p) { scr_await_settled(p); }
+
+/* ECMAScript's INTERNAL module-dependency wait. Unlike a user-authored
+ * await, an already-completed dependency does not introduce a promise-job
+ * hop: the evaluator continues synchronously into the importer. A pending
+ * dependency still parks this module fiber, and rejection propagates into
+ * it exactly like an ordinary await. */
+void scr_module_await(ScrPromise *p) {
+#ifndef __wasi__
+  while (p->state == SCR_PROM_PENDING) scr_await_park(p);
+#endif
+  scr_prom_observe(p);
+  if (p->state == SCR_PROM_REJECTED) scr_promise_rethrow(p);
+}
+
+int scr_promise_finish_top_level(ScrPromise *p) {
+  if (p->state == SCR_PROM_PENDING) {
+    /* ECMAScript module evaluation is still pending, but Node's ref'd
+     * event-loop work is exhausted. Node exits with its dedicated
+     * unsettled-top-level-await status. */
+    scr_exit_code_note(13);
+    return 13;
+  }
+  scr_prom_observe(p);
+  return p->state == SCR_PROM_REJECTED && !p->top_level_handled ? 1 : 0;
+}
+
+void scr_promise_rethrow_top_level(ScrPromise *p) {
+  if (p->state == SCR_PROM_REJECTED) scr_promise_rethrow(p);
+}
+
+
+
+
+
+/* Fulfillment storers, called by the compiled trampolines / resolve thunks. */
+void scr_promise_fulfill_f64(ScrPromise *p, double v) {
+  if (p->state != SCR_PROM_PENDING) return; /* first settle wins */
+  p->state = SCR_PROM_FULFILLED;
+  p->payload_kind = SCR_EXC_F64;
+  p->f64 = v;
+  scr_promise_settle_wake(p);
+}
+void scr_promise_fulfill_bool(ScrPromise *p, bool v) {
+  if (p->state != SCR_PROM_PENDING) return;
+  p->state = SCR_PROM_FULFILLED;
+  p->payload_kind = SCR_EXC_BOOL;
+  p->b = v;
+  scr_promise_settle_wake(p);
+}
+void scr_promise_fulfill_ref(ScrPromise *p, void *v, void *(*retain)(void *), void (*release)(void *), ScrTraceFn trace) {
+  if (p->state != SCR_PROM_PENDING) {
+    if (release) release(v); /* value was +1 for us; drop it */
+    return;
+  }
+  p->state = SCR_PROM_FULFILLED;
+  p->payload_kind = v ? SCR_EXC_REF : SCR_EXC_NONE;
+  p->payload = v;
+  p->retain_fn = retain;
+  p->release_fn = release;
+  p->trace_fn = trace;
+  scr_promise_settle_wake(p);
+}
+void scr_promise_fulfill_void(ScrPromise *p) {
+  if (p->state != SCR_PROM_PENDING) return;
+  p->state = SCR_PROM_FULFILLED;
+  p->payload_kind = SCR_EXC_NONE;
+  scr_promise_settle_wake(p);
+}
+
+/* String payloads ride the STR kind so awaits type them precisely. */
+void scr_promise_fulfill_str(ScrPromise *p, ScrStr *v) {
+  if (p->state != SCR_PROM_PENDING) {
+    scr_str_release(v);
+    return;
+  }
+  p->state = SCR_PROM_FULFILLED;
+  p->payload_kind = SCR_EXC_STR;
+  p->payload = v; /* ownership moves in */
+  p->retain_fn = (void *(*)(void *))scr_str_retain;
+  p->release_fn = (void (*)(void *))scr_str_release;
+  scr_promise_settle_wake(p);
+}
+ScrStr *scr_await_str(ScrPromise *p) {
+  return scr_await_settled(p) && p->payload ? scr_str_retain((ScrStr *)p->payload) : NULL;
+}
+
+/* REJECTS `p` with the exception pending in the ACTIVE cell (moved out —
+ * the cell resets, so the caller returns to the engine/loop with a clean
+ * cell) and wakes waiters; the rejection enters the unhandled ledger like
+ * any other. The island → static promise bridge's rejection half: its
+ * settle callback converts the engine reason into a pending exception
+ * (the exception bridge's one conversion), then lands it here. A no-op
+ * with nothing pending; an already-settled `p` just clears the cell
+ * (first settle wins, like resolve-then-throw). */
+void scr_promise_reject_pending(ScrPromise *p) {
+  if (!scr_exc_pending()) return;
+  if (p->state == SCR_PROM_PENDING) {
+    scr_promise_reject_from_cell(p, scr_exc_current_cell());
+    scr_promise_settle_wake(p);
+  } else {
+    scr_exc_clear();
+  }
+}
+
+/* ── settled-promise minting (the fs/promises bridge) ─────────────────
+ * The promise-returning stdlib functions run their syscall SYNCHRONOUSLY
+ * (the documented non-interleaving divergence, SEMANTICS.md) and wrap the
+ * outcome here: a pending exception in the active cell becomes the
+ * promise's REJECTION (moved out of the cell — the caller's pending check
+ * then sees a clean cell, exactly like Node where fs/promises failures
+ * reject instead of throwing), anything else fulfills with the payload.
+ * The rejected case still enters the unhandled-rejection ledger, so a
+ * never-awaited failure reports at exit like any other rejection. */
+
+static ScrPromise *scr_promise_settled_common(void) {
+  if (!scr_exc_pending()) return NULL;
+  ScrPromise *p = scr_promise_new();
+  scr_promise_reject_from_cell(p, scr_exc_current_cell());
+  scr_promise_settle_wake(p); /* no waiters yet; enters the ledger */
+  return p;
+}
+
+ScrPromise *scr_promise_settled_str(ScrStr *v) {
+  ScrPromise *rejected = scr_promise_settled_common();
+  if (rejected) {
+    scr_str_release(v); /* NULL-tolerant; the op returned a dummy */
+    return rejected;
+  }
+  ScrPromise *p = scr_promise_new();
+  scr_promise_fulfill_str(p, v); /* moves in */
+  return p;
+}
+
+ScrPromise *scr_promise_settled_void(void) {
+  ScrPromise *rejected = scr_promise_settled_common();
+  if (rejected) return rejected;
+  ScrPromise *p = scr_promise_new();
+  scr_promise_fulfill_void(p);
+  return p;
+}
+
+ScrPromise *scr_promise_settled_ref(void *v, void *(*retain)(void *), void (*release)(void *),
+                                     ScrTraceFn trace) {
+  ScrPromise *rejected = scr_promise_settled_common();
+  if (rejected) {
+    if (v && release) release(v);
+    return rejected;
+  }
+  ScrPromise *p = scr_promise_new();
+  scr_promise_fulfill_ref(p, v, retain, release, trace); /* moves in */
+  return p;
+}
+
+/* ── fs/promises ─────────────────────────────────────────────────────
+ * The promise forms run the SAME sync operations and mint an already-
+ * settled promise (scr_promise_settled_*): success fulfills, a pending
+ * exception moves in as the rejection — catchable at the await, like
+ * Node. DOCUMENTED DIVERGENCE (SEMANTICS.md): the syscall blocks the
+ * event loop, so I/O never interleaves with timers or other fibers —
+ * observable only in concurrent code, not in the sequential await
+ * chains CLIs are made of. */
+
+ScrPromise *scr_fsp_read_file(ScrStr *path) {
+  return scr_promise_settled_str(scr_fs_read_file(path));
+}
+
+ScrPromise *scr_fsp_write_file(ScrStr *path, ScrStr *data) {
+  scr_fs_write_file(path, data);
+  return scr_promise_settled_void();
+}
+
+ScrPromise *scr_fsp_write_file_mode(ScrStr *path, ScrStr *data, double mode) {
+  scr_fs_write_file_mode(path, data, mode);
+  return scr_promise_settled_void();
+}
+
+ScrPromise *scr_fsp_mkdir(ScrStr *path) {
+  scr_fs_mkdir(path);
+  return scr_promise_settled_void();
+}
+
+ScrPromise *scr_fsp_mkdir_mode(ScrStr *path, double mode) {
+  scr_fs_mkdir_mode(path, mode);
+  return scr_promise_settled_void();
+}
+
+ScrPromise *scr_fsp_mkdir_recursive(ScrStr *path) {
+  scr_fs_mkdir_recursive(path);
+  return scr_promise_settled_void();
+}
+
+ScrPromise *scr_fsp_mkdir_recursive_mode(ScrStr *path, double mode) {
+  scr_fs_mkdir_recursive_mode(path, mode);
+  return scr_promise_settled_void();
+}
+
+ScrPromise *scr_fsp_unlink(ScrStr *path) {
+  scr_fs_unlink(path);
+  return scr_promise_settled_void();
+}
+
+ScrPromise *scr_fsp_chmod(ScrStr *path, double mode) {
+  scr_fs_chmod(path, mode);
+  return scr_promise_settled_void();
+}
+
+ScrPromise *scr_fsp_readdir(ScrStr *path) {
+  ScrArr *names = scr_fs_readdir(path);
+  return scr_promise_settled_ref(names, &scr_arr_retain_v, &scr_arr_release_v, NULL);
+}
+
+ScrPromise *scr_fsp_rm(ScrStr *path) {
+  scr_fs_rm(path);
+  return scr_promise_settled_void();
+}
+
+ScrPromise *scr_fsp_stat(ScrStr *path) {
+  ScrStats *st = scr_fs_stat(path);
+  return scr_promise_settled_ref(st, &scr_stats_retain_v, &scr_stats_release_v, NULL);
+}
+
+ScrPromise *scr_fsp_realpath(ScrStr *path) {
+  return scr_promise_settled_str(scr_fs_realpath_promise(path));
+}
+
+ScrPromise *scr_fsp_lstat(ScrStr *path) {
+  ScrStats *st = scr_fs_lstat(path);
+  return scr_promise_settled_ref(st, &scr_stats_retain_v, &scr_stats_release_v, NULL);
+}
+
+ScrPromise *scr_fsp_rename(ScrStr *oldpath, ScrStr *newpath) {
+  scr_fs_rename(oldpath, newpath);
+  return scr_promise_settled_void();
+}
+
+/* Shared native work queue, initially introduced by fs.rename and also used
+ * by node:zlib's callback codecs. Work starts immediately on a native worker,
+ * matching libuv's key contract while JS/main is synchronously occupied.
+ * Error/Buffer construction, callback invocation, and every RC mutation stay
+ * on the main runtime thread when a later loop turn observes completion.
+ *
+ * A compact operation queue is also the liveness handle: an outstanding
+ * callback-style request keeps the loop alive. Four persistent workers mirror
+ * libuv's default pool size without creating one OS thread per request. A
+ * platform mutex publishes each result before the loop removes it from the
+ * completion queue. */
+#if !defined(SCR_LIB) && !defined(__wasi__)
+typedef struct ScrFsRenameOp {
+  bool generic;
+  bool ran;
+  void *payload;
+  ScrWorkFn work;
+  ScrWorkFn after;
+  ScrWorkFn destroy;
+  ScrStr *oldpath;
+  ScrStr *newpath;
+  ScrClosure *cb;
+  ScrFsRenameFn fn;
+  int error;
+  struct ScrFsRenameOp *next;
+} ScrFsRenameOp;
+
+static ScrFsRenameOp *scr_fs_rename_work = NULL;
+static ScrFsRenameOp **scr_fs_rename_work_tail = &scr_fs_rename_work;
+static ScrFsRenameOp *scr_fs_rename_done = NULL;
+static ScrFsRenameOp **scr_fs_rename_done_tail = &scr_fs_rename_done;
+static size_t scr_fs_rename_pending_count = 0;
+static size_t scr_fs_rename_worker_count = 0;
+static bool scr_fs_rename_stopping = false;
+static bool scr_fs_rename_shutdown_registered = false;
+
+#ifdef _WIN32
+static HANDLE scr_fs_rename_workers[4];
+static SRWLOCK scr_fs_rename_lock = SRWLOCK_INIT;
+static CONDITION_VARIABLE scr_fs_rename_cond = CONDITION_VARIABLE_INIT;
+static void scr_fs_rename_lock_enter(void) { AcquireSRWLockExclusive(&scr_fs_rename_lock); }
+static void scr_fs_rename_lock_leave(void) { ReleaseSRWLockExclusive(&scr_fs_rename_lock); }
+static void scr_fs_rename_wait(void) {
+  (void)SleepConditionVariableSRW(&scr_fs_rename_cond, &scr_fs_rename_lock, INFINITE, 0);
+}
+static void scr_fs_rename_signal(void) { WakeConditionVariable(&scr_fs_rename_cond); }
+static void scr_fs_rename_broadcast(void) { WakeAllConditionVariable(&scr_fs_rename_cond); }
+#else
+static pthread_t scr_fs_rename_workers[4];
+static pthread_mutex_t scr_fs_rename_lock = PTHREAD_MUTEX_INITIALIZER;
+static pthread_cond_t scr_fs_rename_cond = PTHREAD_COND_INITIALIZER;
+static void scr_fs_rename_lock_enter(void) { (void)pthread_mutex_lock(&scr_fs_rename_lock); }
+static void scr_fs_rename_lock_leave(void) { (void)pthread_mutex_unlock(&scr_fs_rename_lock); }
+static void scr_fs_rename_wait(void) {
+  (void)pthread_cond_wait(&scr_fs_rename_cond, &scr_fs_rename_lock);
+}
+static void scr_fs_rename_signal(void) { (void)pthread_cond_signal(&scr_fs_rename_cond); }
+static void scr_fs_rename_broadcast(void) { (void)pthread_cond_broadcast(&scr_fs_rename_cond); }
+#endif
+
+static void scr_fs_rename_op_release(ScrFsRenameOp *op) {
+  if (op->generic) {
+    op->destroy(op->payload);
+    free(op);
+    return;
+  }
+  scr_str_release(op->oldpath);
+  scr_str_release(op->newpath);
+  scr_closure_release(op->cb);
+  free(op);
+}
+
+static void scr_fs_rename_worker_loop(void) {
+  for (;;) {
+    scr_fs_rename_lock_enter();
+    while (scr_fs_rename_work == NULL && !scr_fs_rename_stopping) scr_fs_rename_wait();
+    if (scr_fs_rename_stopping) {
+      scr_fs_rename_lock_leave();
+      return;
+    }
+    ScrFsRenameOp *op = scr_fs_rename_work;
+    scr_fs_rename_work = op->next;
+    if (scr_fs_rename_work == NULL) scr_fs_rename_work_tail = &scr_fs_rename_work;
+    scr_fs_rename_lock_leave();
+
+    if (op->generic) op->work(op->payload);
+    else op->error = scr_fs_rename_raw(op->oldpath, op->newpath);
+    op->ran = true;
+
+    scr_fs_rename_lock_enter();
+    op->next = NULL;
+    *scr_fs_rename_done_tail = op;
+    scr_fs_rename_done_tail = &op->next;
+    scr_fs_rename_lock_leave();
+  }
+}
+
+#ifdef _WIN32
+static DWORD WINAPI scr_fs_rename_worker(LPVOID payload) {
+  (void)payload;
+  scr_fs_rename_worker_loop();
+  return 0;
+}
+#else
+static void *scr_fs_rename_worker(void *payload) {
+  (void)payload;
+  scr_fs_rename_worker_loop();
+  return NULL;
+}
+#endif
+
+/* Runs before the ordinary runtime atexit cleanup/audit (it is registered
+ * lazily by the first rename request, after scr_init/scr_lib_init). Fatal
+ * top-level or callback exceptions can stop the loop with requests still in
+ * flight; join the workers before releasing their retained path/callback
+ * payloads so sanitized builds keep the program's real exit verdict. */
+static void scr_fs_renames_shutdown(void) {
+  scr_fs_rename_lock_enter();
+  scr_fs_rename_stopping = true;
+  scr_fs_rename_broadcast();
+  scr_fs_rename_lock_leave();
+
+  for (size_t i = 0; i < scr_fs_rename_worker_count; i++) {
+#ifdef _WIN32
+    (void)WaitForSingleObject(scr_fs_rename_workers[i], INFINITE);
+    CloseHandle(scr_fs_rename_workers[i]);
+#else
+    (void)pthread_join(scr_fs_rename_workers[i], NULL);
+#endif
+  }
+
+  scr_fs_rename_lock_enter();
+  ScrFsRenameOp *work = scr_fs_rename_work;
+  ScrFsRenameOp *done = scr_fs_rename_done;
+  scr_fs_rename_work = NULL;
+  scr_fs_rename_work_tail = &scr_fs_rename_work;
+  scr_fs_rename_done = NULL;
+  scr_fs_rename_done_tail = &scr_fs_rename_done;
+  scr_fs_rename_pending_count = 0;
+  scr_fs_rename_lock_leave();
+  while (work != NULL) {
+    ScrFsRenameOp *next = work->next;
+    scr_fs_rename_op_release(work);
+    work = next;
+  }
+  while (done != NULL) {
+    ScrFsRenameOp *next = done->next;
+    scr_fs_rename_op_release(done);
+    done = next;
+  }
+}
+
+static bool scr_fs_renames_pending(void) { return scr_fs_rename_pending_count != 0; }
+
+static bool scr_fs_renames_dispatch(void) {
+  scr_fs_rename_lock_enter();
+  ScrFsRenameOp *op = scr_fs_rename_done;
+  if (op != NULL) {
+    scr_fs_rename_done = op->next;
+    if (scr_fs_rename_done == NULL) scr_fs_rename_done_tail = &scr_fs_rename_done;
+  }
+  scr_fs_rename_lock_leave();
+  if (op == NULL) return false;
+  scr_fs_rename_pending_count--;
+  if (op->generic) {
+    if (!op->ran) op->work(op->payload);
+    op->after(op->payload);
+    scr_fs_rename_op_release(op);
+    return true;
+  }
+  ScrCaught *caught = NULL;
+  ScrError *err = NULL;
+  if (op->error != 0) {
+    scr_fs_rename_error(op->error, op->oldpath, op->newpath);
+    caught = scr_exc_take();
+    if (caught && caught->kind == SCR_EXC_OBJ && scr_error_is(caught->payload)) {
+      err = (ScrError *)caught->payload;
+    }
+  }
+  op->fn(op->cb, err);
+  scr_caught_release(caught);
+  scr_fs_rename_op_release(op);
+  return true;
+}
+
+static void scr_work_submit_op(ScrFsRenameOp *op) {
+  if (!scr_fs_rename_shutdown_registered) {
+    if (atexit(scr_fs_renames_shutdown) != 0) {
+      scr_trap("scriptc: could not register worker cleanup\n");
+    }
+    scr_fs_rename_shutdown_registered = true;
+  }
+  scr_fs_rename_pending_count++;
+  scr_fs_rename_lock_enter();
+  *scr_fs_rename_work_tail = op;
+  scr_fs_rename_work_tail = &op->next;
+  int create_error = 0;
+  if (scr_fs_rename_worker_count < 4) {
+#ifdef _WIN32
+    HANDLE worker = CreateThread(NULL, 0, scr_fs_rename_worker, NULL, 0, NULL);
+    if (worker != NULL) {
+      scr_fs_rename_workers[scr_fs_rename_worker_count] = worker;
+      scr_fs_rename_worker_count++;
+    } else {
+      create_error = EAGAIN;
+    }
+#else
+    pthread_t worker;
+    create_error = pthread_create(&worker, NULL, scr_fs_rename_worker, NULL);
+    if (create_error == 0) {
+      scr_fs_rename_workers[scr_fs_rename_worker_count] = worker;
+      scr_fs_rename_worker_count++;
+    }
+#endif
+  }
+  if (create_error != 0 && scr_fs_rename_worker_count == 0) {
+    /* Submission failure remains asynchronous. Rename preserves its
+     * resource error; generic work falls back to the main loop turn. */
+    while (scr_fs_rename_work != NULL) {
+      ScrFsRenameOp *failed = scr_fs_rename_work;
+      scr_fs_rename_work = failed->next;
+      failed->error = create_error;
+      failed->ran = !failed->generic;
+      failed->next = NULL;
+      *scr_fs_rename_done_tail = failed;
+      scr_fs_rename_done_tail = &failed->next;
+    }
+    scr_fs_rename_work_tail = &scr_fs_rename_work;
+  } else {
+    scr_fs_rename_signal();
+  }
+  scr_fs_rename_lock_leave();
+}
+#elif defined(__wasi__)
+/* WASI Preview 1 has no threads, but rename itself is available. Queue the
+ * request and perform it on the next loop turn so callback timing remains
+ * asynchronous and the request remains a liveness handle. */
+typedef struct ScrFsRenameOp {
+  bool generic;
+  void *payload;
+  ScrWorkFn work;
+  ScrWorkFn after;
+  ScrWorkFn destroy;
+  ScrStr *oldpath;
+  ScrStr *newpath;
+  ScrClosure *cb;
+  ScrFsRenameFn fn;
+  struct ScrFsRenameOp *next;
+} ScrFsRenameOp;
+
+static ScrFsRenameOp *scr_fs_rename_done = NULL;
+static ScrFsRenameOp **scr_fs_rename_done_tail = &scr_fs_rename_done;
+static size_t scr_fs_rename_pending_count = 0;
+
+static bool scr_fs_renames_pending(void) { return scr_fs_rename_pending_count != 0; }
+
+static bool scr_fs_renames_dispatch(void) {
+  ScrFsRenameOp *op = scr_fs_rename_done;
+  if (op == NULL) return false;
+  scr_fs_rename_done = op->next;
+  if (scr_fs_rename_done == NULL) scr_fs_rename_done_tail = &scr_fs_rename_done;
+  scr_fs_rename_pending_count--;
+
+  if (op->generic) {
+    op->work(op->payload);
+    op->after(op->payload);
+    op->destroy(op->payload);
+    free(op);
+    return true;
+  }
+
+  int error = scr_fs_rename_raw(op->oldpath, op->newpath);
+  ScrCaught *caught = NULL;
+  ScrError *err = NULL;
+  if (error != 0) {
+    scr_fs_rename_error(error, op->oldpath, op->newpath);
+    caught = scr_exc_take();
+    if (caught && caught->kind == SCR_EXC_OBJ && scr_error_is(caught->payload)) {
+      err = (ScrError *)caught->payload;
+    }
+  }
+  op->fn(op->cb, err);
+  scr_caught_release(caught);
+  scr_str_release(op->oldpath);
+  scr_str_release(op->newpath);
+  scr_closure_release(op->cb);
+  free(op);
+  return true;
+}
+#else
+static bool scr_fs_renames_pending(void) { return false; }
+static bool scr_fs_renames_dispatch(void) { return false; }
+#endif
+
+void scr_fs_rename_async(ScrStr *oldpath, ScrStr *newpath,
+                         ScrClosure *cb, ScrFsRenameFn fn) {
+#if defined(SCR_LIB)
+  (void)oldpath; (void)newpath; (void)fn;
+  scr_closure_release(cb);
+  scr_trap("scriptc: callback-style fs.rename is not supported by this target\n");
+#elif defined(__wasi__)
+  ScrFsRenameOp *op = calloc(1, sizeof *op);
+  if (!op) scr_trap("scriptc: out of memory\n");
+  op->oldpath = scr_str_retain(oldpath);
+  op->newpath = scr_str_retain(newpath);
+  op->cb = cb;
+  op->fn = fn;
+  *scr_fs_rename_done_tail = op;
+  scr_fs_rename_done_tail = &op->next;
+  scr_fs_rename_pending_count++;
+#else
+  ScrFsRenameOp *op = calloc(1, sizeof *op);
+  if (!op) scr_trap("scriptc: out of memory\n");
+  op->oldpath = scr_str_retain(oldpath);
+  op->newpath = scr_str_retain(newpath);
+  op->cb = cb;
+  op->fn = fn;
+  scr_work_submit_op(op);
+#endif
+}
+
+void scr_work_submit(void *payload, ScrWorkFn work, ScrWorkFn after,
+                     ScrWorkFn destroy) {
+#if defined(SCR_LIB)
+  (void)work; (void)after;
+  destroy(payload);
+  scr_trap("scriptc: native worker jobs are not supported by this target\n");
+#elif defined(__wasi__)
+  ScrFsRenameOp *op = calloc(1, sizeof *op);
+  if (!op) scr_trap("scriptc: out of memory\n");
+  op->generic = true;
+  op->payload = payload;
+  op->work = work;
+  op->after = after;
+  op->destroy = destroy;
+  *scr_fs_rename_done_tail = op;
+  scr_fs_rename_done_tail = &op->next;
+  scr_fs_rename_pending_count++;
+#else
+  ScrFsRenameOp *op = calloc(1, sizeof *op);
+  if (!op) scr_trap("scriptc: out of memory\n");
+  op->generic = true;
+  op->payload = payload;
+  op->work = work;
+  op->after = after;
+  op->destroy = destroy;
+  scr_work_submit_op(op);
+#endif
+}
+
+void scr_fs_rename_thunk0(ScrClosure *cb, ScrError *err) {
+  (void)err;
+  ((void (*)(ScrClosure *))cb->fn)(cb);
+}
+
+/* Callback-style crypto completions. The operation validates and computes
+ * before enqueueing; delivery is a later macrotask, never synchronous with
+ * the API call. The queue owns both references and transfers the Buffer to
+ * the emitted callback adapter. */
+typedef struct ScrCryptoBytesOp {
+  ScrBytes *value;
+  ScrClosure *cb;
+  ScrCryptoBytesFn fn;
+  struct ScrCryptoBytesOp *next;
+} ScrCryptoBytesOp;
+
+static ScrCryptoBytesOp *scr_crypto_bytes_head = NULL;
+static ScrCryptoBytesOp **scr_crypto_bytes_tail = &scr_crypto_bytes_head;
+static size_t scr_crypto_bytes_pending_count = 0;
+static bool scr_crypto_bytes_cleanup_registered = false;
+
+static void scr_crypto_bytes_shutdown(void) {
+  while (scr_crypto_bytes_head != NULL) {
+    ScrCryptoBytesOp *op = scr_crypto_bytes_head;
+    scr_crypto_bytes_head = op->next;
+    scr_bytes_release(op->value);
+    scr_closure_release(op->cb);
+    free(op);
+  }
+  scr_crypto_bytes_tail = &scr_crypto_bytes_head;
+  scr_crypto_bytes_pending_count = 0;
+}
+
+static bool scr_crypto_bytes_pending(void) {
+  return scr_crypto_bytes_pending_count != 0;
+}
+
+static bool scr_crypto_bytes_dispatch(void) {
+  ScrCryptoBytesOp *op = scr_crypto_bytes_head;
+  if (op == NULL) return false;
+  scr_crypto_bytes_head = op->next;
+  if (scr_crypto_bytes_head == NULL) scr_crypto_bytes_tail = &scr_crypto_bytes_head;
+  scr_crypto_bytes_pending_count--;
+  op->fn(op->cb, op->value); /* adapter consumes value */
+  scr_closure_release(op->cb);
+  free(op);
+  return true;
+}
+
+void scr_crypto_defer_bytes(ScrBytes *value, ScrClosure *cb, ScrCryptoBytesFn fn) {
+  if (!scr_crypto_bytes_cleanup_registered) {
+    if (atexit(scr_crypto_bytes_shutdown) != 0) {
+      scr_bytes_release(value);
+      scr_closure_release(cb);
+      scr_trap("scriptc: could not register crypto callback cleanup\n");
+    }
+    scr_crypto_bytes_cleanup_registered = true;
+  }
+  ScrCryptoBytesOp *op = malloc(sizeof *op);
+  if (!op) {
+    scr_bytes_release(value);
+    scr_closure_release(cb);
+    scr_trap("scriptc: out of memory\n");
+  }
+  op->value = value;
+  op->cb = cb;
+  op->fn = fn;
+  op->next = NULL;
+  *scr_crypto_bytes_tail = op;
+  scr_crypto_bytes_tail = &op->next;
+  scr_crypto_bytes_pending_count++;
+}
+
+/* ── node:timers/promises ────────────────────────────────────────────
+ * The promisified pair: a PENDING void promise a one-shot heap timer
+ * (setTimeout) or the immediate queue (setImmediate) fulfills — the
+ * resolve closure is scr_make_resolve's void thunk, so settlement wakes
+ * awaiting fibers exactly like every other promise, and the armed
+ * timer/immediate keeps the loop alive until it fires (Node: an awaited
+ * timers/promises sleep holds the process open). Neither throws; the
+ * delay rides scr_set_timeout's own Node-exact coercion (clamp to 1,
+ * truncate). */
+
+ScrPromise *scr_tp_set_timeout(double ms) {
+  ScrPromise *p = scr_promise_new();
+  scr_set_timeout(scr_make_resolve(p, 3 /* void */), ms);
+  return p;
+}
+
+ScrPromise *scr_tp_set_immediate(void) {
+  ScrPromise *p = scr_promise_new();
+  scr_set_immediate(scr_make_resolve(p, 3 /* void */));
+  return p;
+}
+
+/* ── setImmediate as a first-class dyn value ─────────────────────────
+ * The global passed AS A FUNCTION VALUE into untyped code (the Node-suite
+ * traceCallback shape: `channel.traceCallback(setImmediate, 0, ctx, null,
+ * cb)`). The minted dyn callable schedules its own immediate that calls
+ * args[0] with the remaining arguments (Node's setImmediate(cb, ...args)
+ * contract) and answers undefined (the Immediate handle object has no dyn
+ * story — clearImmediate over this value is not modeled). A non-function
+ * first argument throws Node's ERR_INVALID_ARG_TYPE synchronously. */
+
+static void scr_imm_dyn_fire(ScrClosure *c) {
+  ScrDyn *fn = (ScrDyn *)scr_box_get_ref(c->caps[0]);
+  ScrDyn *args = (ScrDyn *)scr_box_get_ref(c->caps[1]);
+  ScrDyn *r = scr_dyn_call(fn, args->v.arr.items, args->v.arr.len, "immediate callback");
+  if (r != NULL) scr_dyn_release(r);
+  scr_dyn_release(fn);
+  scr_dyn_release(args);
+}
+
+static ScrDyn *scr_imm_dyn_thunk(ScrClosure *clo, ScrDyn *const *args, size_t argc) {
+  (void)clo;
+  if (argc == 0 || args[0]->kind != SCR_DYN_FUNC) {
+    static const char msg[] = "The \"callback\" argument must be of type function.";
+    scr_throw_error_msg_code(SCR_ERR_TYPE, msg, sizeof msg - 1, "ERR_INVALID_ARG_TYPE");
+    return NULL;
+  }
+  ScrDyn *rest = scr_dyn_new_arr();
+  for (size_t i = 1; i < argc; i++) scr_dyn_arr_push(rest, scr_dyn_retain(args[i]));
+  ScrClosure *c = scr_closure_new((void *)scr_imm_dyn_fire, 2);
+  c->caps[0] = scr_box_new_obj(scr_dyn_retain_v, scr_dyn_release_v, NULL);
+  scr_box_set_ref(c->caps[0], scr_dyn_retain(args[0]));
+  c->caps[1] = scr_box_new_obj(scr_dyn_retain_v, scr_dyn_release_v, NULL);
+  scr_box_set_ref(c->caps[1], rest); /* moves */
+  scr_set_immediate(c);
+  return scr_dyn_retain(scr_dyn_undefined());
+}
+
+ScrDyn *scr_set_immediate_dyn_value(void) {
+  return scr_dyn_new_func(scr_closure_new((void *)scr_imm_dyn_thunk, 0), scr_imm_dyn_thunk,
+                          1, "(cb,...)", "setImmediate");
+}
+
+/* ── queueMicrotask ──────────────────────────────────────────────────
+ * The READY queue's FIFO is the microtask queue (promise continuations
+ * ride it as parked fibers), so a queueMicrotask callback enters the SAME
+ * queue as a stackless envelope — one order across both producers, like
+ * V8's single microtask queue. scr_resume_fiber runs the closure on the
+ * main stack; a throw is an uncaught exception (Node's queueMicrotask),
+ * never a rejection. */
+void scr_queue_microtask(ScrClosure *cb) {
+  ScrFiber *f = calloc(1, sizeof *f);
+  if (!f) scr_oom();
+  f->micro_cb = cb; /* ownership moves in */
+  scr_ready_push(f);
+}
+
+/* The checked-dynamic argument form (JS files — common.mustCall wrappers
+ * and the suite's invalid-input probes): a non-function throws Node's
+ * ERR_INVALID_ARG_TYPE synchronously; a function value is called with
+ * zero arguments (Node passes none — extra queueMicrotask arguments are
+ * ignored). */
+static void scr_micro_dyn_fire(ScrClosure *c) {
+  ScrDyn *fn = (ScrDyn *)scr_box_get_ref(c->caps[0]);
+  ScrDyn *r = scr_dyn_call(fn, NULL, 0, "queueMicrotask callback");
+  if (r != NULL) scr_dyn_release(r);
+  scr_dyn_release(fn);
+}
+
+void scr_queue_microtask_dyn(const ScrDyn *cb) {
+  if (cb->kind != SCR_DYN_FUNC) {
+    static const char msg[] = "The \"callback\" argument must be of type function.";
+    scr_throw_error_msg_code(SCR_ERR_TYPE, msg, sizeof msg - 1, "ERR_INVALID_ARG_TYPE");
+    return;
+  }
+  ScrClosure *c = scr_closure_new((void *)scr_micro_dyn_fire, 1);
+  c->caps[0] = scr_box_new_obj(scr_dyn_retain_v, scr_dyn_release_v, NULL);
+  scr_box_set_ref(c->caps[0], scr_dyn_retain((ScrDyn *)cb));
+  scr_queue_microtask(c);
+}
+
+/* ── the event hooks (scr_events.c) ──────────────────────────────────
+ * Process signal/exit events and the piped-stdin surface live in
+ * scr_events.c, which links ONLY into binaries whose IR uses those
+ * surfaces (the io-hook precedent, gated like scr_regex/scr_fetch):
+ * scr_events_install() fills these slots before %main. Event-free
+ * builds keep every slot NULL and the loop is byte-identical in
+ * behavior. The scr_lib.c hooks live in scr_lib.c and the abnormal-exit
+ * code hint in scr_exception.c, so each unit stays self-contained (the
+ * runtime C tests link them separately). */
+
+static bool (*scr_events_pending_fn)(void) = NULL;  /* keeps the loop alive */
+static bool (*scr_events_watching_fn)(void) = NULL; /* wants the poll sleep */
+static void (*scr_events_dispatch_fn)(void) = NULL; /* fire due listeners */
+static int (*scr_events_pollfds_fn)(int out[2]) = NULL;
+
+void scr_loop_set_events(bool (*pending)(void), bool (*watching)(void),
+                          void (*dispatch)(void), int (*pollfds)(int out[2])) {
+  scr_events_pending_fn = pending;
+  scr_events_watching_fn = watching;
+  scr_events_dispatch_fn = dispatch;
+  scr_events_pollfds_fn = pollfds;
+}
+
+/* ── the event loop ───────────────────────────────────────────────────── */
+
+static void scr_resume_fiber(ScrFiber *f) {
+  if (f->gen_job != NULL) {
+    ScrGen *g = f->gen_job;
+    free(f);
+    scr_async_gen_job(g);
+    scr_gen_release(g);
+    return;
+  }
+  /* A queueMicrotask envelope: run the closure on the main stack, no
+   * context switch. A throw leaves the exception cell pending — every
+   * drain site returns to main's uncaught report, Node's queueMicrotask
+   * semantics (uncaughtException, never an unhandled rejection). */
+  if (f->micro_cb != NULL) {
+    ScrClosure *cb = f->micro_cb;
+    free(f);
+    ((void (*)(ScrClosure *))cb->fn)(cb);
+    scr_closure_release(cb);
+    return;
+  }
+  if (f->gen != NULL) {
+    scr_async_gen_resume_ready(f->gen);
+    return;
+  }
+#ifdef _WIN32
+  /* The loop runs on the main stack; make sure scr_loop_ctx names its
+   * fiber handle (a no-op after the first conversion). */
+  (void)scr_win_self();
+#endif
+  f->return_to = &scr_loop_ctx;
+#ifdef __wasi__
+  scr_current = f;
+  scr_exc_swap_cell(&f->exc);
+  scr_als_active = &f->als;
+  if (f->coro != NULL) {
+    scr_wasi_coro_resume(f->coro);
+  } else {
+    /* scr_async_spawn_after's runtime C continuation. Its dependency is
+     * settled, so the entry can extract through scr_await_* without a
+     * stack switch and then completes in this turn. */
+    f->entry(f, f->argpack);
+    scr_fiber_finish(f);
+  }
+  scr_current = NULL;
+  scr_exc_swap_cell(NULL);
+  scr_als_active = &scr_als_main_slot;
+#else
+  scr_switch(&scr_loop_ctx, &f->ctx, f);
+#endif
+  if (f->done) {
+    scr_promise_release(f->promise);
+    scr_fiber_destroy(f);
+    scr_fibers_live--;
+  }
+}
+
+/* Reap granularity while children are pending on the POLLING fallback:
+ * the loop polls waitpid(WNOHANG) at least this often instead of sleeping
+ * to the next timer deadline. The primary path has no cap — the sleep
+ * waits on the child watch (scr_children_wait — kqueue NOTE_EXIT on BSD,
+ * pidfd epoll on Linux, the timer deadline riding the wait timeout) and a
+ * child's exit wakes it immediately. The fallback covers platforms with
+ * neither backend, children whose exit watch could not be armed, and
+ * turns where the io poll takes the sleep (curl's fds can't join it). */
+#define SCR_CHILD_POLL_MS 1.0
+
+/* Sleep cap while only external io is pending (no timer models its
+ * wakeup): the io poll sleeps on REAL fds inside this window (a fetch
+ * transfer's sockets wake it early), so a generous cap costs nothing. */
+#define SCR_IO_POLL_MS 1000.0
+
+/* Sleep cap while signal handlers or stdin consumers are pending AND the
+ * io poll owns the sleep (curl's fd wait retries on EINTR and can't watch
+ * the wake pipe): bounded signal/stdin latency during a fetch at 20
+ * wakeups/second. The dedicated poll(2) sleep needs no cap — it waits on
+ * the wake pipe and fd 0 directly. */
+#define SCR_SIGNAL_POLL_MS 50.0
+
+/* The external io hook (scr_runtime.h): the dynamic island registers
+ * engine-job draining + fetch transfer polling here. Static builds never
+ * set it — both slots stay NULL and the loop is byte-identical in
+ * behavior. */
+static bool (*scr_io_pending_fn)(void) = NULL;
+static void (*scr_io_poll_fn)(double max_wait_ms) = NULL;
+
+void scr_loop_set_io(bool (*pending)(void), void (*poll)(double)) {
+  scr_io_pending_fn = pending;
+  scr_io_poll_fn = poll;
+}
+
+/* The island's earliest armed timer deadline (HUGE_VAL when none): caps
+ * the sleep below so an armed AbortSignal.timeout fires on time while
+ * the loop waits on socket readiness — without keeping the loop alive by
+ * itself (the liveness test never consults it, Node's unref'd timer). */
+static double (*scr_island_deadline_fn)(void) = NULL;
+
+void scr_loop_set_island_deadline(double (*fn)(void)) { scr_island_deadline_fn = fn; }
+
+/* The net hook (scr_net.c, when linked — the events-hook shape):
+ * `pending` keeps the loop alive while servers/sockets are live,
+ * `dispatch` fires at every turn top, and `pollfd` is the unit's poller
+ * fd for the idle poll(2) sleep (readable while events are pending —
+ * kqueue and epoll fds both behave this way).
+ * Net-free builds keep every slot NULL and the loop is byte-identical. */
+static bool (*scr_net_pending_fn)(void) = NULL;
+static void (*scr_net_dispatch_fn)(void) = NULL;
+static int (*scr_net_pollfd_fn)(void) = NULL;
+
+void scr_loop_set_net(bool (*pending)(void), void (*dispatch)(void), int (*pollfd)(void)) {
+  scr_net_pending_fn = pending;
+  scr_net_dispatch_fn = dispatch;
+  scr_net_pollfd_fn = pollfd;
+}
+
+/* The dgram hook (scr_dgram.c, when linked) — the net hook's exact shape:
+ * one more set of nullable slots, byte-identical loop behavior when
+ * unset. */
+static bool (*scr_dgram_pending_fn)(void) = NULL;
+static void (*scr_dgram_dispatch_fn)(void) = NULL;
+static int (*scr_dgram_pollfd_fn)(void) = NULL;
+
+void scr_loop_set_dgram(bool (*pending)(void), void (*dispatch)(void), int (*pollfd)(void)) {
+  scr_dgram_pending_fn = pending;
+  scr_dgram_dispatch_fn = dispatch;
+  scr_dgram_pollfd_fn = pollfd;
+}
+
+/* The fs.watch hook (scr_watch.c, when linked) — the net hook's exact
+ * shape: one more set of nullable slots, byte-identical loop behavior
+ * when unset. */
+static bool (*scr_watch_pending_fn)(void) = NULL;
+static void (*scr_watch_dispatch_fn)(void) = NULL;
+static int (*scr_watch_pollfd_fn)(void) = NULL;
+
+void scr_loop_set_watch(bool (*pending)(void), void (*dispatch)(void), int (*pollfd)(void)) {
+  scr_watch_pending_fn = pending;
+  scr_watch_dispatch_fn = dispatch;
+  scr_watch_pollfd_fn = pollfd;
+}
+
+/* The foreign-FFI queue hook (scr_ffi.c when a format-5 descriptor exists).
+ * Its pending count includes live registrations (reffed by default) and
+ * queued deliveries; dispatch runs exactly one callback per loop turn. */
+static bool (*scr_ffi_pending_fn)(void) = NULL;
+static bool (*scr_ffi_dispatch_fn)(void) = NULL;
+static int (*scr_ffi_pollfd_fn)(void) = NULL;
+static void (*scr_ffi_stop_fn)(void) = NULL;
+
+void scr_loop_set_ffi(bool (*pending)(void), bool (*dispatch)(void),
+                      int (*pollfd)(void), void (*stop)(void)) {
+  scr_ffi_pending_fn = pending;
+  scr_ffi_dispatch_fn = dispatch;
+  scr_ffi_pollfd_fn = pollfd;
+  scr_ffi_stop_fn = stop;
+}
+
+/* The stream hook (scr_stream.c, when linked): `pending` keeps the loop
+ * alive while deferred stream ticks exist, `dispatch` drains them at the
+ * TOP of every turn — before the events/net stations, the closest
+ * placement to Node's process.nextTick (whose stream emissions these
+ * are). No poller fd: ticks are pure CPU work, always ready. */
+static bool (*scr_stream_pending_fn)(void) = NULL;
+static void (*scr_stream_dispatch_fn)(void) = NULL;
+
+void scr_loop_set_stream(bool (*pending)(void), void (*dispatch)(void)) {
+  scr_stream_pending_fn = pending;
+  scr_stream_dispatch_fn = dispatch;
+}
+
+/* Dispatch hooks yield between event batches when fibers are already
+ * queued, so promise jobs interleave with event delivery (Node's
+ * microtask checkpoints between macrotasks). */
+bool scr_loop_has_ready(void) { return scr_ready_len > 0; }
+
+static bool scr_loop_run_pass(ScrPromise *top_level, bool first_checkpoint) {
+  /* The FIRST checkpoint after the synchronous main body runs promise
+   * jobs BEFORE the first tick drain: Node's main-module evaluation is
+   * itself awaited (the runMain continuation is a microtask queued after
+   * the body's own), so microtasks scheduled during the body beat ticks
+   * scheduled during the body exactly once, at startup — differentially
+   * pinned. Every later checkpoint drains ticks first. */
+  bool rejection_failed = false;
+  for (;;) {
+    if (top_level != NULL && top_level->state == SCR_PROM_REJECTED) break;
+    /* process.nextTick callbacks BEFORE promise jobs (Node's checkpoint
+     * order): the tick queue to exhaustion, then the microtask queue to
+     * exhaustion, and back while either has work — a microtask's
+     * nextTick runs before the turn proceeds, but never preempts the
+     * microtask queue mid-drain (V8 drains it fully). A tick's uncaught
+     * throw ends the loop like any listener's (main reports it). */
+    if (!first_checkpoint) {
+      while (scr_nt_head != NULL) {
+        ScrNtick *t = scr_nt_head;
+        scr_nt_head = t->next;
+        if (scr_nt_head == NULL) scr_nt_tail = NULL;
+        ScrClosure *cb = t->cb;
+        ScrFsRenameFn error_cb = t->error_cb;
+        void (*raw)(void) = t->raw;
+        free(t);
+        if (cb) {
+          if (error_cb) error_cb(cb, NULL);
+          else ((void (*)(ScrClosure *))cb->fn)(cb);
+          scr_closure_release(cb);
+        } else {
+          raw(); /* one stream tick, FIFO with the user ticks around it */
+        }
+        if (scr_exc_pending()) return false;
+      }
+    }
+    /* Microtasks to exhaustion (Node: promise jobs before timers). */
+    while (scr_ready_len > 0) {
+      ScrFiber *f = scr_ready[scr_ready_head++];
+      scr_ready_len--;
+      scr_resume_fiber(f);
+    }
+    first_checkpoint = false;
+    /* The ESM loader observes a rejected entry evaluation at a promise-job
+     * checkpoint and terminates before later ref'd timers or I/O can run.
+     * Wait until this checkpoint's ready queue is drained — the root's
+     * rejection handler is itself promise-job ordered — then leave through
+     * the normal teardown below. A fulfilled root deliberately does not
+     * stop the loop. */
+    if (top_level != NULL && top_level->state == SCR_PROM_REJECTED) break;
+    if (scr_nt_head != NULL) continue;
+    /* Node decides unhandled rejections at the END of each complete
+     * nextTick/microtask checkpoint, before advancing to timers or I/O.
+     * A rejected executable module root wins over OTHER rejections from
+     * this same checkpoint (the root check above); rejections from an
+     * earlier checkpoint have already delivered here. Handled listeners
+     * may enqueue more jobs or ref'd work, so return to the checkpoint
+     * head instead of declaring the loop exhausted underneath them. */
+    if (scr_report_unhandled_rejections()) {
+      rejection_failed = true;
+      break;
+    }
+    if (scr_ready_len > 0 || scr_nt_head != NULL || scr_nunhandled > 0) continue;
+    /* Quiescent between turns (microtasks drained, nothing running):
+     * collect any cycles the turn left behind. One SCHEDULED pass — almost
+     * always a nursery one, sized to the turn's own garbage. A full sweep
+     * here would walk the whole live heap every turn, which is precisely
+     * what the generations exist to avoid. No-op on an empty buffer. */
+    scr_cyc_collect_scheduled();
+    /* Completed callback-style worker jobs are delivered on the main
+     * runtime thread. A worker may have finished while synchronous user code
+     * occupied that thread. Deliver exactly one completion per checkpoint:
+     * Node drains nextTicks and microtasks after each native callback before
+     * invoking the next ready callback. */
+    if (scr_fs_renames_pending()) {
+      bool dispatched = scr_fs_renames_dispatch();
+      if (scr_exc_pending()) return false;
+      if (dispatched) continue;
+    }
+    if (scr_crypto_bytes_pending()) {
+      bool dispatched = scr_crypto_bytes_dispatch();
+      if (scr_exc_pending()) return false;
+      if (dispatched) continue;
+    }
+    /* Foreign native callbacks are macrotasks. Deliver one, then restart at
+     * the microtask checkpoint before considering the next queued post. */
+    if (scr_ffi_dispatch_fn != NULL && scr_ffi_dispatch_fn()) {
+      if (scr_exc_pending()) return false;
+      if (scr_ready_len > 0 || scr_nt_head != NULL || scr_nunhandled > 0) continue;
+    }
+    /* Stream tick dispatch (scr_stream.c, when linked): the deferred
+     * next-tick emissions ('data' flow kicks, 'readable'/'end'/'finish'/
+     * 'drain'/'error'/'close') fire now, FIRST — the nextTick station.
+     * Listeners may enqueue microtasks or more ticks — restart the turn
+     * so those drain first. */
+    if (scr_stream_dispatch_fn != NULL) {
+      scr_stream_dispatch_fn();
+      if (scr_exc_pending()) return false; /* uncaught throw in a listener */
+      if (scr_ready_len > 0) continue;
+      if (scr_stream_pending_fn != NULL && scr_stream_pending_fn()) continue;
+    }
+    /* Event dispatch (scr_events.c, when linked): watched signals
+     * delivered since the last turn fire their listeners now (macrotasks,
+     * like timers), then stdin — while a consumer exists, probe fd 0 and
+     * deliver one arrived chunk / the EOF events. Both may enqueue
+     * microtasks — restart the turn so those drain first. */
+    if (scr_events_dispatch_fn != NULL) {
+      scr_events_dispatch_fn();
+      if (scr_exc_pending()) return false; /* uncaught throw in a listener */
+      if (scr_ready_len > 0) continue;
+    }
+    /* Net dispatch (scr_net.c, when linked): accepts, arrived data,
+     * connect completions, and the deferred listening/error/close emits
+     * fire now (macrotasks, like the events hook). Callbacks may enqueue
+     * microtasks — restart the turn so those drain first. */
+    if (scr_net_dispatch_fn != NULL) {
+      scr_net_dispatch_fn();
+      if (scr_exc_pending()) return false; /* uncaught throw in a listener */
+      if (scr_ready_len > 0) continue;
+    }
+    /* Dgram dispatch (scr_dgram.c, when linked): arrived datagrams and
+     * the deferred listening/connect/error/close emits (plus pending
+     * dns.lookup callbacks) fire now — the net hook's exact station. */
+    if (scr_dgram_dispatch_fn != NULL) {
+      scr_dgram_dispatch_fn();
+      if (scr_exc_pending()) return false; /* uncaught throw in a listener */
+      if (scr_ready_len > 0) continue;
+    }
+    /* Watch dispatch (scr_watch.c, when linked): file events queued on
+     * the unit's event backend fire their FSWatcher listeners now — the
+     * net hook's exact station. */
+    if (scr_watch_dispatch_fn != NULL) {
+      scr_watch_dispatch_fn();
+      if (scr_exc_pending()) return false; /* uncaught throw in a listener */
+      if (scr_ready_len > 0) continue;
+    }
+    /* Reap spawned children and fire their listeners (before timers,
+     * like Node's nextTick-ish spawn-failure "error"). A callback may
+     * enqueue microtasks or spawn again — restart the turn if so.
+     * EXCEPT when ONLY unref'd children remain: Node's liveness check
+     * precedes any further poll/reap, so once nothing reffed holds the
+     * loop their pending exits are DROPPED at loop exit — never
+     * delivered. Without this gate the kill-then-exit shape (a timer's
+     * last act kills an unref'd child) would race machine timing: a
+     * death landing before this sweep fired its 'exit' where Node
+     * deterministically never does. The teardown below releases the
+     * undelivered listeners, so the RC audit stays clean. */
+    if (scr_children_pending()) {
+      bool held =
+          scr_reffed_timers > 0 || scr_reffed_immediates > 0 || scr_children_reffed_pending() ||
+          /* spawn failures settle regardless: their 'error' is a
+           * next-tick event, delivered even on an unref'd child */
+          scr_children_failed_pending() ||
+          (scr_io_pending_fn != NULL && scr_io_pending_fn()) ||
+          (scr_events_pending_fn != NULL && scr_events_pending_fn()) ||
+          (scr_net_pending_fn != NULL && scr_net_pending_fn()) ||
+          (scr_dgram_pending_fn != NULL && scr_dgram_pending_fn()) ||
+          (scr_watch_pending_fn != NULL && scr_watch_pending_fn()) ||
+          (scr_ffi_pending_fn != NULL && scr_ffi_pending_fn()) ||
+          scr_fs_renames_pending() || scr_crypto_bytes_pending();
+      if (held) {
+        scr_children_poll();
+        if (scr_exc_pending()) return false; /* uncaught throw in a listener */
+        if (scr_ready_len > 0) continue;
+      }
+    }
+    /* Deferred stream ticks are always-ready work: never sleep or exit
+     * while any exist (the dispatch above drains them, so reaching here
+     * with ticks pending means a dispatch station enqueued more). User
+     * nextTicks enqueued by a station listener are the same story. */
+    if (scr_stream_pending_fn != NULL && scr_stream_pending_fn()) continue;
+    if (scr_nt_head != NULL) continue;
+    bool kids = scr_children_pending();
+    bool io = scr_io_pending_fn != NULL && scr_io_pending_fn();
+    bool events = scr_events_pending_fn != NULL && scr_events_pending_fn();
+    bool net = scr_net_pending_fn != NULL && scr_net_pending_fn();
+    bool dgram = scr_dgram_pending_fn != NULL && scr_dgram_pending_fn();
+    bool watch = scr_watch_pending_fn != NULL && scr_watch_pending_fn();
+    bool ffi = scr_ffi_pending_fn != NULL && scr_ffi_pending_fn();
+    bool renames = scr_fs_renames_pending();
+    bool crypto = scr_crypto_bytes_pending();
+    /* Timer liveness counts only REF'd timers: an unref'd timer stays in
+     * the heap (and fires if the loop runs on for other reasons) but does
+     * not by itself keep the process alive — Node's unref semantics.
+     * Children follow the same rule: an unref'd child is still REAPED
+     * while the loop runs (kids drives the sweeps and sleeps above) but
+     * only reffed ones keep the process alive. */
+    if (scr_reffed_timers == 0 && scr_reffed_immediates == 0 && !scr_children_reffed_pending() && !io && !events && !net && !dgram && !watch && !ffi && !renames && !crypto) break;
+    /* Sleep to the earliest deadline, then run every due timer (each may
+     * enqueue microtasks, which the next iteration drains first). Who
+     * sleeps depends on what is pending:
+     * - external io: the io POLL takes the sleep — it waits on real fds
+     *   up to the deadline (socket readiness wakes it early) and makes
+     *   progress (engine jobs, arrived response data) before timers run.
+     *   A child's exit can't wake curl's fds, so children re-impose the
+     *   reap-granularity cap on this path.
+     * - signals watched or stdin consumed: poll(2) takes the sleep — the
+     *   wake pipe, fd 0, and the child watch's fd together, so any of
+     *   them ends it immediately (details at the branch).
+     * - children only: the sleep waits on the child watch — the exit
+     *   wakeup and the timer deadline together via the wait timeout — so a
+     *   child's exit ends the sleep immediately and the next turn's reap
+     *   pass fires its listeners. scr_children_wait declines when it
+     *   can't wake for every pending child; the ~1ms polling cap is the
+     *   fallback.
+     * - timers only: plain nanosleep to the deadline. */
+    double now = scr_now_ms();
+    double due = scr_ntimers > 0 ? scr_timers[0].deadline_ms : now + SCR_IO_POLL_MS;
+    /* The worker pool has no platform poll handle yet. Bound the idle
+     * wait exactly like the portable child fallback so completion is noticed
+     * promptly even when another poller owns the sleep. */
+    if (renames && due > now + SCR_CHILD_POLL_MS) due = now + SCR_CHILD_POLL_MS;
+    /* An armed island timer (AbortSignal.timeout) caps the sleep: it must
+     * fire on time even while the poller waits on socket readiness. */
+    if (scr_island_deadline_fn != NULL) {
+      double isl_due = scr_island_deadline_fn();
+      if (isl_due < due) due = isl_due;
+    }
+    /* Pending immediates are always-ready work: no sleep — run due timers
+     * (Node's timers phase precedes check), then the check phase below. */
+    if (scr_pending_immediates > 0) due = now;
+    /* Child polling can drain a pipe into the IPC queue after its dispatch
+     * station has run. Completed sends and disconnects can also become ready
+     * during that turn. No fd will wake us for this userspace work: return to
+     * dispatch without sleeping, still allowing due timers and immediates. */
+    if (scr_children_ready()) due = now;
+#if !defined(_WIN32) && !defined(__wasi__)
+    if (due > now) scr_fiber_stacks_clear();
+#endif
+    bool evw = scr_events_watching_fn != NULL && scr_events_watching_fn();
+    if (io) {
+      if (kids && due > now + SCR_CHILD_POLL_MS) due = now + SCR_CHILD_POLL_MS;
+      /* Signals/stdin/net can't wake curl's fd wait (and its poll retries
+       * on EINTR), so they re-impose a coarser cap — bounded Ctrl-C and
+       * socket latency during a fetch, without the reap-granularity
+       * cost. */
+      else if ((evw || net || dgram || watch || ffi) && due > now + SCR_SIGNAL_POLL_MS) due = now + SCR_SIGNAL_POLL_MS;
+      scr_io_poll_fn(due > now ? due - now : 0);
+      now = scr_now_ms();
+      if (scr_ready_len > 0) continue; /* io callbacks woke fibers */
+    } else if (evw || net || dgram || watch || ffi) {
+#if defined(_WIN32) || defined(__wasi__)
+      /* The win32 arm, and WASI hosts whose poll_oneoff adapters do not
+       * reliably wake for a closed inherited stdin pipe: the sleep is a capped nanosleep and
+       * the next turn's dispatch does the work — signal flags (set by the
+       * CRT handler on msvcrt's console-ctrl thread) and the stdin probe
+       * at SCR_SIGNAL_POLL_MS granularity (scr_events.c), socket/dgram
+       * readiness at the child-reap cap (scr_loop_wsapoll.c's zero-timeout
+       * WSAPoll drains at dispatch; ~1ms keeps loopback exchanges inside
+       * the granularity Node's IOCP-woken loop delivers, where 50ms could
+       * reorder a socket emit past a short timer). When the caps ever
+       * show up in a profile, the upgrade is a real waitable arm —
+       * WaitForMultipleObjects over WSAEVENTs, or IOCP. */
+      if (evw && due > now + SCR_SIGNAL_POLL_MS) due = now + SCR_SIGNAL_POLL_MS;
+      if ((net || dgram || watch) && due > now + SCR_CHILD_POLL_MS) due = now + SCR_CHILD_POLL_MS;
+      if (ffi && due > now + SCR_CHILD_POLL_MS) due = now + SCR_CHILD_POLL_MS;
+      if (kids && due > now + SCR_CHILD_POLL_MS) due = now + SCR_CHILD_POLL_MS;
+      if (due > now) {
+        double wait = due - now;
+        struct timespec ts = {(time_t)(wait / 1000.0), (long)((wait - (double)((time_t)(wait / 1000.0)) * 1000.0) * 1e6)};
+        nanosleep(&ts, NULL);
+      }
+      now = scr_now_ms();
+#else
+      /* poll(2) takes the sleep: the events unit's fds — the signal wake
+       * pipe (a signal delivered BEFORE the poll entered still has its
+       * byte in the pipe — no lost-wakeup race) and fd 0 while stdin has
+       * a consumer — plus the child watch's fd when it can represent
+       * every pending child (kqueue and epoll fds poll readable when
+       * events are pending); unrepresentable children keep the ~1ms reap cap
+       * instead. Dispatch happens at the next turn's top — the poll only
+       * decides how long to sleep. */
+      struct pollfd fds[7];
+      int nfds = 0;
+      int evfds[2];
+      int nev = evw && scr_events_pollfds_fn != NULL ? scr_events_pollfds_fn(evfds) : 0;
+      for (int i = 0; i < nev; i++) {
+        fds[nfds].fd = evfds[i];
+        fds[nfds].events = POLLIN;
+        fds[nfds++].revents = 0;
+      }
+      if (net) {
+        /* The net unit's poller fd: readable while socket events are
+         * pending, so arrived data/accepts end the sleep immediately.
+         * Dispatch (which also DRAINS the poller) happens at the next
+         * turn's top — the poll only decides how long to sleep. */
+        int nfd = scr_net_pollfd_fn != NULL ? scr_net_pollfd_fn() : -1;
+        if (nfd >= 0) {
+          fds[nfds].fd = nfd;
+          fds[nfds].events = POLLIN;
+          fds[nfds++].revents = 0;
+        } else if (due > now + SCR_SIGNAL_POLL_MS) {
+          due = now + SCR_SIGNAL_POLL_MS;
+        }
+      }
+      if (dgram) {
+        /* The dgram unit's poller fd — the net slot's exact story. */
+        int dfd = scr_dgram_pollfd_fn != NULL ? scr_dgram_pollfd_fn() : -1;
+        if (dfd >= 0) {
+          fds[nfds].fd = dfd;
+          fds[nfds].events = POLLIN;
+          fds[nfds++].revents = 0;
+        } else if (due > now + SCR_SIGNAL_POLL_MS) {
+          due = now + SCR_SIGNAL_POLL_MS;
+        }
+      }
+      if (watch) {
+        /* The watch unit's event fd — the net slot's exact story. */
+        int wfd = scr_watch_pollfd_fn != NULL ? scr_watch_pollfd_fn() : -1;
+        if (wfd >= 0) {
+          fds[nfds].fd = wfd;
+          fds[nfds].events = POLLIN;
+          fds[nfds++].revents = 0;
+        } else if (due > now + SCR_SIGNAL_POLL_MS) {
+          due = now + SCR_SIGNAL_POLL_MS;
+        }
+      }
+      if (ffi) {
+        int ffd = scr_ffi_pollfd_fn != NULL ? scr_ffi_pollfd_fn() : -1;
+        if (ffd >= 0) {
+          fds[nfds].fd = ffd;
+          fds[nfds].events = POLLIN;
+          fds[nfds++].revents = 0;
+        } else if (due > now + SCR_CHILD_POLL_MS) {
+          due = now + SCR_CHILD_POLL_MS;
+        }
+      }
+      if (kids) {
+        int cfd = scr_children_wake_fd();
+        if (cfd >= 0) {
+          fds[nfds].fd = cfd;
+          fds[nfds].events = POLLIN;
+          fds[nfds++].revents = 0;
+        } else if (due > now + SCR_CHILD_POLL_MS) {
+          due = now + SCR_CHILD_POLL_MS;
+        }
+      }
+      double wait = due > now ? due - now : 0;
+      if (wait > 1e9) wait = 1e9; /* poll takes int ms */
+      (void)poll(fds, (nfds_t)nfds, (int)(wait + 0.999));
+      /* Consume queued child-exit events NOW (zero-timeout drain) — an
+       * unconsumed event keeps the kqueue fd readable and would turn the
+       * next poll into a busy spin while other children still run. The
+       * WNOHANG sweep at the next turn's top stays the reaper. */
+      if (kids) (void)scr_children_wait(0);
+      now = scr_now_ms();
+#endif /* !_WIN32 && !__wasi__ */
+    } else if (kids && scr_children_wait(due > now ? due - now : 0)) {
+      now = scr_now_ms(); /* woke early on an exit, or hit the deadline */
+    } else {
+      if (kids && due > now + SCR_CHILD_POLL_MS) due = now + SCR_CHILD_POLL_MS;
+      if (due > now) {
+        double wait = due - now;
+        struct timespec ts = {(time_t)(wait / 1000.0), (long)((wait - (double)((time_t)(wait / 1000.0)) * 1000.0) * 1e6)};
+        nanosleep(&ts, NULL);
+        now = due;
+      }
+    }
+    while (scr_ntimers > 0 && scr_timers[0].deadline_ms <= now) {
+      ScrTimer t = scr_timer_pop();
+      /* Timer callbacks are plain sync closures, run on the main stack.
+       * An interval entry is OUT of the heap while its callback runs;
+       * clearInterval(its own id) from inside sets the firing flag and
+       * the re-arm below drops it instead. A throw kills the interval
+       * with the program (Node: the uncaught exception ends the process
+       * before any rescheduled tick could run). */
+      if (t.id != 0) {
+        scr_firing_id = t.id;
+        scr_firing_cleared = false;
+        scr_firing_reffed = t.reffed; /* unref/ref during the callback lands here */
+        scr_firing_refresh = false;
+      }
+      ((void (*)(ScrClosure *))t.cb->fn)(t.cb);
+      if (t.id != 0 && t.repeat_ms > 0 && !scr_firing_cleared) {
+        /* Re-arm relative to the post-callback clock (libuv's uv_timer
+         * repeat behavior: no catch-up bursts after a slow callback). The
+         * ref state carries across ticks (a self-unref'd interval stays
+         * unref'd). */
+        ScrTimer again = {scr_now_ms() + t.repeat_ms, scr_timer_seq++, t.cb, t.repeat_ms, t.id, scr_firing_reffed, t.delay_ms};
+        scr_timer_push(again);
+      } else if (t.id != 0 && scr_firing_refresh && !scr_firing_cleared) {
+        /* refresh() from inside the one-shot's own callback: re-arm to
+         * now + the original delay, ref state carried (Node's
+         * Timeout.refresh — the timer fires again). */
+        ScrTimer again = {scr_now_ms() + t.delay_ms, scr_timer_seq++, t.cb, 0, t.id, scr_firing_reffed, t.delay_ms};
+        scr_timer_push(again);
+      } else {
+        scr_closure_release(t.cb);
+      }
+      scr_firing_id = 0;
+      if (scr_exc_pending()) return false; /* uncaught throw in a callback: main handles it */
+      if (scr_ready_len > 0) break; /* drain microtasks before more timers */
+    }
+    /* The check phase: immediates queued BEFORE this phase started run
+     * now, FIFO; ones a callback queues wait for the next turn (the end
+     * snapshot — Node's once-per-turn rule, so a setImmediate chain can't
+     * starve the loop). Cleared slots (cb NULL) skip. Microtasks drain
+     * between callbacks, like Node's inter-macrotask checkpoints; a woken
+     * fiber restarts the turn first (the timer loop's break above), so the
+     * phase only runs on a drained queue. */
+    if (scr_ready_len == 0 && scr_pending_immediates > 0) {
+      size_t end = scr_nimmediates;
+      while (scr_immediates_head < end) {
+        size_t i = scr_immediates_head++;
+        ScrClosure *cb = scr_immediates[i].cb;
+        if (cb == NULL) continue; /* cleared */
+        scr_immediates[i].cb = NULL; /* fired: clear/ref/unref on this id no-op now */
+        if (scr_pending_immediates > 0) scr_pending_immediates--;
+        if (scr_immediates[i].reffed) {
+          scr_immediates[i].reffed = false;
+          if (scr_reffed_immediates > 0) scr_reffed_immediates--;
+        }
+        ((void (*)(ScrClosure *))cb->fn)(cb);
+        scr_closure_release(cb);
+        if (scr_exc_pending()) return false; /* uncaught throw: main handles it */
+        while (scr_ready_len > 0) {
+          ScrFiber *f = scr_ready[scr_ready_head++];
+          scr_ready_len--;
+          scr_resume_fiber(f);
+        }
+        if (top_level != NULL && top_level->state == SCR_PROM_REJECTED) break;
+      }
+      /* Fully drained: reset the cursor so the array reuses its slots. */
+      if (scr_immediates_head == scr_nimmediates) {
+        scr_immediates_head = 0;
+        scr_nimmediates = 0;
+      }
+    }
+  }
+  return rejection_failed;
+}
+
+bool scr_loop_run(ScrPromise *top_level) {
+  bool first_checkpoint = true;
+  bool rejection_failed;
+  for (;;) {
+    rejection_failed = scr_loop_run_pass(top_level && !top_level->top_level_handled ? top_level : NULL, first_checkpoint);
+    first_checkpoint = false;
+    if (scr_exc_pending()) {
+      if (scr_exc_handle_uncaught(false)) continue;
+      break; /* main reports the original or handler-thrown exception */
+    }
+    if (rejection_failed) break;
+    if (top_level && top_level->state == SCR_PROM_REJECTED && !top_level->top_level_handled) {
+      scr_promise_rethrow(top_level);
+      if (scr_exc_handle_uncaught(true)) {
+        top_level->top_level_handled = true;
+        scr_prom_observe(top_level);
+        continue;
+      }
+      scr_discard_unhandled_rejections();
+      break;
+    }
+    break;
+  }
+  /* Exit can now leave UNREF'd timers armed in the heap (ordinary
+   * exhaustion, a fatal module root, or an unhandled rejection). They
+   * never fire, so release their closures here or the RC audit counts
+   * them as leaks. Uncaught callback throws still return above for main
+   * to report through the existing exceptional teardown path. */
+  scr_timers_teardown();
+  /* No reffed foreign registration or queued invocation remains at ordinary
+   * exhaustion. Disarm posting before exit listeners/global teardown so a
+   * straggling native thread becomes a safe silent drop. */
+  if (scr_ffi_stop_fn != NULL) scr_ffi_stop_fn();
+  /* Retained native FFI registrations are deliberately NOT torn down
+   * here: process 'exit' listeners run after the loop returns (inline in
+   * main, before any atexit handler) and may legitimately release a
+   * registration or pump a native library one last time — exactly like
+   * the process.exit() path, where the ledger is also still intact. On
+   * returns that reach atexit, the sweep scr_ffi_retain registered (LIFO
+   * before the RC audit) then drops the ledger's references. It does NOT
+   * run on process.exit(): scr_process_exit ends in _Exit, skipping every
+   * atexit handler — the sweep and the RC audit alike — and leaves the
+   * ledger to the OS. */
+  /* Unref'd children the loop never reaped: release the
+   * registry's references (their listeners never fire — the process is
+   * exiting, Node's behavior; the OS reparents the children). */
+  scr_children_teardown();
+  scr_fibers_abandoned = scr_fibers_live;
+  scr_note_abandoned_fibers(scr_fibers_abandoned);
+  return rejection_failed;
+}
+
+/* The island's half of the unhandled-rejection report (scr_island.c
+ * registers it at engine boot): called with print=true when the static
+ * ledger below reported nothing — one report, one voice, like Node's
+ * first-unhandled-rejection death. Returns whether the island had any.
+ * Static builds never set it. */
+static bool (*scr_island_rejections_fn)(bool print) = NULL;
+static int (*scr_island_jobs_drain_fn)(void) = NULL;
+
+void scr_loop_set_island_rejections(bool (*fn)(bool print),
+                                    int (*drain_jobs)(void)) {
+  scr_island_rejections_fn = fn;
+  scr_island_jobs_drain_fn = drain_jobs;
+}
+
+/* ── process.on('unhandledRejection') ─────────────────────────────────
+ * Dyn listeners are called per never-observed rejection at the completed
+ * nextTick/microtask checkpoint. A registered listener suppresses the
+ * default report and exit 1, exactly Node's handled-event contract. */
+
+/* The unhandled-rejection LISTENER hook (scr_async_dyn.c installs it at
+ * registration — the loop-hook pattern, so listener-free binaries keep
+ * their size class): called per never-observed rejection; false = a
+ * listener threw (the uncaught crash path). */
+bool (*scr_urj_deliver_fn)(ScrPromise *p) = NULL;
+
+/* Unhandled rejections at a completed nextTick/microtask checkpoint:
+ * Node prints an error and exits 1 when no listener handles the event.
+ * Snapshot the ledger: a listener can reject another promise, but that
+ * new rejection belongs to the NEXT checkpoint rather than this report. */
+bool scr_report_unhandled_rejections(void) {
+  /* Static fibers drain before engine jobs in this runtime's documented
+   * island ordering. Complete BOTH halves of that microtask checkpoint
+   * before deciding either rejection ledger; engine reactions can attach
+   * a handler to a rejection that would otherwise look unhandled here.
+   * A host callback can wake a static fiber/tick, in which case the loop
+   * must drain that work before reporting too. */
+  if (scr_island_jobs_drain_fn != NULL) {
+    scr_island_jobs_drain_fn();
+    if (scr_ready_len > 0 || scr_nt_head != NULL) return false;
+  }
+  bool any = false;
+  bool crashed = false;
+  size_t report_count = scr_nunhandled;
+  for (size_t i = 0; i < report_count; i++) {
+    ScrPromise *p = scr_maybe_unhandled[i];
+    if (p->state == SCR_PROM_REJECTED && !p->rejection_observed && !crashed) {
+      if (scr_urj_deliver_fn != NULL) {
+        /* Listener dispatch (scr_async_dyn.c installed the hook at
+         * registration): the event handles it, like Node — per entry,
+         * no report, exit 0. A listener throw is the uncaught crash.
+         * reported_unhandled arms the 'rejectionHandled' window: a
+         * handler the listener itself attaches fires the sibling event
+         * (scr_prom_observe). */
+        p->rejection_observed = true;
+        p->reported_unhandled = true;
+        if (!scr_urj_deliver_fn(p) && !scr_exc_handle_uncaught(false)) crashed = true;
+      } else if (scr_uncaught_exception_hook != NULL) {
+        scr_promise_rethrow(p);
+        if (scr_exc_handle_uncaught(true)) {
+          p->rejection_observed = true;
+          p->reported_unhandled = true;
+        } else {
+          crashed = true;
+        }
+      } else if (!any) {
+        any = true;
+        fflush(stdout);
+        fputs("Unhandled promise rejection: ", stderr);
+        switch (p->payload_kind) {
+        case SCR_EXC_F64: {
+          char buf[32];
+          scr_f64_to_str(p->f64, buf);
+          fputs(buf, stderr);
+          break;
+        }
+        case SCR_EXC_BOOL: fputs(p->b ? "true" : "false", stderr); break;
+        case SCR_EXC_STR: fputs(((ScrStr *)p->payload)->data, stderr); break;
+        case SCR_EXC_OBJ:
+          /* Error rejections render "name: message", like the uncaught path. */
+          if (scr_error_is(p->payload)) {
+            ScrStr *s = scr_error_to_string((ScrError *)p->payload);
+            fwrite(s->data, 1, s->len, stderr);
+            scr_str_release(s);
+            break;
+          }
+          /* fall through */
+        default: fputs("[object]", stderr); break;
+        }
+        fputc('\n', stderr);
+      }
+    }
+    scr_promise_release(p);
+  }
+  size_t remaining = scr_nunhandled - report_count;
+  if (remaining > 0) {
+    memmove(scr_maybe_unhandled, scr_maybe_unhandled + report_count,
+            remaining * sizeof *scr_maybe_unhandled);
+  }
+  scr_nunhandled = remaining;
+  if (crashed) {
+    scr_exc_print_uncaught();
+    return true;
+  }
+  if (scr_island_rejections_fn != NULL) {
+    bool island = scr_island_rejections_fn(!any);
+    any = any || island;
+  }
+  /* main returns 1 on a reported rejection — the 'exit' listeners (atexit)
+   * must see that code, like Node's. */
+  if (any) scr_exit_code_note(1);
+  return any;
+}
+
+/* A fatal executable-module rejection suppresses unrelated rejections
+ * created in the SAME checkpoint. Drop their retained ledger references
+ * without delivering process events or a competing default report. */
+void scr_discard_unhandled_rejections(void) {
+  for (size_t i = 0; i < scr_nunhandled; i++) {
+    scr_promise_release(scr_maybe_unhandled[i]);
+  }
+  scr_nunhandled = 0;
+}
+
+/* ── new Promise(executor) ────────────────────────────────────────────── */
+
+/* resolve closures: caps[0] is a box whose slot holds the promise (+1). */
+static ScrPromise *scr_resolve_target(ScrClosure *self) {
+  return (ScrPromise *)scr_box_get_ref(self->caps[0]);
+}
+static void scr_resolve_thunk_f64(ScrClosure *self, double v) {
+  ScrPromise *p = scr_resolve_target(self);
+  scr_promise_fulfill_f64(p, v);
+  scr_promise_release(p);
+}
+static void scr_resolve_thunk_bool(ScrClosure *self, bool v) {
+  ScrPromise *p = scr_resolve_target(self);
+  scr_promise_fulfill_bool(p, v);
+  scr_promise_release(p);
+}
+static void scr_resolve_thunk_str(ScrClosure *self, ScrStr *v) {
+  ScrPromise *p = scr_resolve_target(self);
+  scr_promise_fulfill_str(p, v); /* callee owns its params: moves in */
+  scr_promise_release(p);
+}
+static void scr_resolve_thunk_void(ScrClosure *self) {
+  ScrPromise *p = scr_resolve_target(self);
+  scr_promise_fulfill_void(p);
+  scr_promise_release(p);
+}
+/* Ref payloads: emitted code passes the concrete retain/release via a
+ * second box slot? Keep simple: ref resolve thunks are monomorphized by the
+ * EMITTER (it knows the arm type) — it emits a tiny static thunk calling
+ * scr_promise_fulfill_ref with the concrete helpers, wrapped over caps[0].
+ * The runtime provides only the scalar/str/void thunks. */
+
+ScrClosure *scr_make_resolve(ScrPromise *p, int kind /*0 f64,1 bool,2 str,3 void*/) {
+  void *fns[4] = {(void *)&scr_resolve_thunk_f64, (void *)&scr_resolve_thunk_bool,
+                  (void *)&scr_resolve_thunk_str, (void *)&scr_resolve_thunk_void};
+  ScrClosure *c = scr_closure_new(fns[kind], 1);
+  ScrBox *b = scr_box_new_obj(scr_promise_retain_v, scr_promise_release_v,
+                               scr_promise_trace_v);
+  scr_box_set_ref(b, scr_promise_retain(p));
+  c->caps[0] = b;
+  return c;
+}
+
+/* Emitted ref-kind resolve thunks call this. */
+void scr_resolve_ref_impl(ScrClosure *self, void *v, void *(*retain)(void *), void (*release)(void *), ScrTraceFn trace) {
+  ScrPromise *p = scr_resolve_target(self);
+  scr_promise_fulfill_ref(p, v, retain, release, trace);
+  scr_promise_release(p);
+}
+
+/* ── emitter support ──────────────────────────────────────────────────── */
+
+ScrPromise *scr_fiber_promise(ScrFiber *f) { return f->promise; }
+
+/* Generalized resolve construction: `fn` is an emitted thunk with closure
+ * calling convention (self, value) that forwards to scr_resolve_ref_impl
+ * with the concrete RC helpers. */
+ScrClosure *scr_make_resolve_fn(ScrPromise *p, void *fn) {
+  ScrClosure *c = scr_closure_new(fn, 1);
+  ScrBox *b = scr_box_new_obj(scr_promise_retain_v, scr_promise_release_v,
+                               scr_promise_trace_v);
+  scr_box_set_ref(b, scr_promise_retain(p));
+  c->caps[0] = b;
+  return c;
+}
+
+/* ── reject closures (new Promise((resolve, reject) => ...)) ──────────
+ * Mirror of the resolve closures: caps[0] is a box whose slot holds the
+ * promise (+1). The reason is always an Error-hierarchy instance (the
+ * frontend pins the reject parameter to `(reason: Error) => void`), and it
+ * stores as SCR_EXC_OBJ — exactly the thrown-Error representation — so an
+ * await rethrows it, catch-side instanceof answers, and the uncaught
+ * printer prints "name: message". First settle wins, exactly JS: reject
+ * after any settle (resolve-then-reject, double reject) releases the
+ * reason and does nothing. The closure calling convention makes the thunk
+ * own its param, like every compiled callee. */
+static void scr_reject_thunk(ScrClosure *self, ScrError *reason) {
+  ScrPromise *p = scr_resolve_target(self);
+  if (p->state != SCR_PROM_PENDING) {
+    scr_error_release(reason);
+  } else {
+    p->state = SCR_PROM_REJECTED;
+    p->payload_kind = SCR_EXC_OBJ;
+    p->payload = reason; /* ownership moves in */
+    p->retain_fn = scr_error_retain_v;
+    p->release_fn = scr_error_release_v;
+    p->trace_fn = scr_error_trace_arg();
+    scr_promise_settle_wake(p);
+  }
+  scr_promise_release(p);
+}
+
+ScrClosure *scr_make_reject(ScrPromise *p) {
+  ScrClosure *c = scr_closure_new((void *)&scr_reject_thunk, 1);
+  ScrBox *b = scr_box_new_obj(scr_promise_retain_v, scr_promise_release_v,
+                               scr_promise_trace_v);
+  scr_box_set_ref(b, scr_promise_retain(p));
+  c->caps[0] = b;
+  return c;
+}
+
+/* Shared executor epilogue: an escaping throw rejects a still-pending
+ * promise (payload moved out of the active exception cell); a throw after
+ * the promise settled is swallowed, like resolve-then-throw in JS. */
+static void scr_executor_check_throw(ScrPromise *p) {
+  if (!scr_exc_pending()) return;
+  if (p->state == SCR_PROM_PENDING) {
+    scr_promise_reject_from_cell(p, scr_exc_current_cell());
+    scr_promise_settle_wake(p);
+  } else {
+    scr_exc_clear();
+  }
+}
+
+/* Runs a `new Promise(executor)` executor synchronously; an escaping throw
+ * rejects the promise (JS-exact). Borrows `p` and `exec`; `resolve`
+ * ownership moves INTO the executor call (compiled callees own and finally
+ * release their params). */
+void scr_promise_run_executor(ScrPromise *p, ScrClosure *exec, ScrClosure *resolve) {
+  ((void (*)(ScrClosure *, ScrClosure *))exec->fn)(exec, resolve);
+  scr_executor_check_throw(p);
+}
+
+/* Zero-param executor (`new Promise(() => {})`): nothing can ever resolve —
+ * a deliberately-forever-pending promise. Borrows both. */
+void scr_promise_run_executor0(ScrPromise *p, ScrClosure *exec) {
+  ((void (*)(ScrClosure *))exec->fn)(exec);
+  scr_executor_check_throw(p);
+}
+
+/* Two-param executor (`new Promise((resolve, reject) => ...)`): same as
+ * scr_promise_run_executor with the reject closure's +1 also moving into
+ * the call. */
+void scr_promise_run_executor2(ScrPromise *p, ScrClosure *exec, ScrClosure *resolve,
+                                ScrClosure *reject) {
+  ((void (*)(ScrClosure *, ScrClosure *, ScrClosure *))exec->fn)(exec, resolve, reject);
+  scr_executor_check_throw(p);
+}
+
+/* ── sync generators (function*) ──────────────────────────────────────
+ * Contract in scr_runtime.h. The fiber machinery above is the engine:
+ * scr_switch is a synchronous coroutine hop, so generators need no event
+ * loop — .next() from the main stack (or any fiber) blocks until the
+ * yield. */
+
+enum {
+  SCR_GEN_UNSTARTED = 0,
+  SCR_GEN_SUSPENDED,
+  SCR_GEN_AWAITING,
+  SCR_GEN_RUNNING,
+  SCR_GEN_DONE
+};
+
+/* One type-erased value slot: the exception cell's payload technique,
+ * with only the release entry point (takes MOVE — nothing here retains). */
+typedef struct {
+  ScrExcKind kind; /* NONE / F64 / BOOL / REF (strings ride REF) */
+  double f64;
+  bool b;
+  void *payload;
+  void (*release_fn)(void *);
+} ScrGenSlot;
+
+enum { SCR_AG_NEXT = 0, SCR_AG_RETURN, SCR_AG_THROW };
+
+typedef struct ScrAsyncGenRequest {
+  int mode;
+  ScrGenSlot arg;
+  ScrExcCell thrown;
+  ScrPromise *promise; /* owned until this request settles */
+  struct ScrAsyncGenRequest *next;
+} ScrAsyncGenRequest;
+
+static void scr_gen_slot_reset(ScrGenSlot *s) {
+  if (s->kind == SCR_EXC_REF && s->payload != NULL) s->release_fn(s->payload);
+  s->kind = SCR_EXC_NONE;
+  s->payload = NULL;
+  s->release_fn = NULL;
+}
+
+static void scr_gen_exc_reset(ScrExcCell *cell) {
+  if (cell->kind == SCR_EXC_STR) {
+    scr_str_release((ScrStr *)cell->payload);
+  } else if (cell->kind == SCR_EXC_REF || cell->kind == SCR_EXC_OBJ ||
+             cell->kind == SCR_EXC_PRIMITIVE_REF) {
+    cell->release_fn(cell->payload);
+  }
+  ScrStackFrame *stack = cell->stack;
+  memset(cell, 0, sizeof *cell);
+  cell->stack = stack;
+}
+
+static void scr_gen_exc_move(ScrExcCell *dst, ScrExcCell *src) {
+  scr_gen_exc_reset(dst);
+  ScrStackFrame *dst_stack = dst->stack;
+  ScrStackFrame *src_stack = src->stack;
+  *dst = *src;
+  dst->stack = dst_stack;
+  memset(src, 0, sizeof *src);
+  src->stack = src_stack;
+}
+
+struct ScrGen {
+  size_t rc;
+  ScrFiber *fiber; /* NULL once torn down (done, or unstarted release) */
+  int state;
+  ScrGenSlot out; /* yielded value / completion value */
+  ScrGenSlot in;  /* the .next(v) argument */
+  ScrGenSlot ret; /* a parked .return(v) value */
+  /* The never-started teardown: drops the packed (already-retained)
+   * arguments the spawn wrapper built. Emitted per generator function. */
+  void (*drop_args)(void *);
+  bool is_async;
+  void (*settle_async)(ScrGen *, ScrPromise *);
+  ScrAsyncGenRequest *requests_head;
+  ScrAsyncGenRequest *requests_tail;
+  ScrPromise *active_request;
+  ScrExcCell async_error;
+  bool settle_queued;
+  bool success_hop_done;
+};
+
+static ScrGen *scr_gen_new_common(void (*entry)(ScrFiber *, void *), void *argpack,
+                                  void (*drop_args)(void *),
+                                  void (*settle_async)(ScrGen *, ScrPromise *)) {
+  ScrGen *g = calloc(1, sizeof *g);
+  if (!g) scr_oom();
+  g->rc = 1;
+  g->state = SCR_GEN_UNSTARTED;
+  g->drop_args = drop_args;
+  g->is_async = settle_async != NULL;
+  g->settle_async = settle_async;
+  scr_obj_alloc_note();
+
+  ScrFiber *f = calloc(1, sizeof *f);
+  if (!f) scr_oom();
+  f->gen = g;
+  f->entry = entry;
+  f->argpack = argpack; /* owned by the fiber start; drop_args if never started */
+  /* AsyncLocalStorage: generator bodies run in their CREATOR's context
+   * (the spawn-inheritance stance; resumes swap it active like any
+   * fiber switch). */
+  f->als = scr_als_ctx_retain(*scr_als_active);
+  scr_fibers_live++;
+#ifdef _WIN32
+  f->ctx = CreateFiberEx(4096, 0, 0, scr_trampoline, NULL);
+  if (f->ctx == NULL) scr_oom();
+#elif defined(__wasi__)
+  f->ctx = 0;
+#else
+  scr_fiber_context_init(f);
+#endif
+  g->fiber = f;
+  return g;
+}
+
+ScrGen *scr_gen_new(void (*entry)(ScrFiber *, void *), void *argpack,
+                    void (*drop_args)(void *)) {
+  return scr_gen_new_common(entry, argpack, drop_args, NULL);
+}
+
+ScrGen *scr_async_gen_new(void (*entry)(ScrFiber *, void *), void *argpack,
+                          void (*drop_args)(void *),
+                          void (*settle)(ScrGen *, ScrPromise *)) {
+  return scr_gen_new_common(entry, argpack, drop_args, settle);
+}
+
+ScrGen *scr_gen_retain(ScrGen *g) {
+  if (g) g->rc++;
+  return g;
+}
+
+void scr_gen_release(ScrGen *g) {
+  if (!g || --g->rc != 0) return;
+  scr_gen_slot_reset(&g->out);
+  scr_gen_slot_reset(&g->in);
+  scr_gen_slot_reset(&g->ret);
+  scr_gen_exc_reset(&g->async_error);
+  if (g->fiber != NULL) {
+    if (g->state == SCR_GEN_UNSTARTED) {
+      /* Never ran: nothing on the stack owns anything — clean teardown.
+       * The packed arguments (+1 each) drop through the emitted helper. */
+      if (g->drop_args != NULL) g->drop_args(g->fiber->argpack);
+      else free(g->fiber->argpack);
+      scr_fiber_destroy(g->fiber);
+      scr_fibers_live--;
+    } else {
+      /* Suspended: the stack holds live locals — unwinding would run user
+       * finally blocks Node's GC never runs, so the fiber is deliberately
+       * ABANDONED (it stays in the live count; a program that drops a
+       * suspended generator gets the abandoned-fiber audit note, the
+       * loop-exhaustion story). The fiber's back-pointer clears so a
+       * dangling resume can never reach the freed ScrGen. */
+      g->fiber->gen = NULL;
+    }
+  }
+  scr_obj_free_note();
+  free(g);
+}
+
+void *scr_gen_retain_v(void *g) { return scr_gen_retain((ScrGen *)g); }
+void scr_gen_release_v(void *g) { scr_gen_release((ScrGen *)g); }
+
+bool scr_gen_done(ScrGen *g) { return g->state == SCR_GEN_DONE; }
+ScrGen *scr_gen_of_fiber(ScrFiber *f) { return f->gen; }
+
+/* Slot setters (release the previous occupant; payloads MOVE in). */
+static void scr_gen_slot_f64(ScrGenSlot *s, double v) {
+  scr_gen_slot_reset(s);
+  s->kind = SCR_EXC_F64;
+  s->f64 = v;
+}
+static void scr_gen_slot_bool(ScrGenSlot *s, bool v) {
+  scr_gen_slot_reset(s);
+  s->kind = SCR_EXC_BOOL;
+  s->b = v;
+}
+static void scr_gen_slot_ref(ScrGenSlot *s, void *v, void (*release)(void *)) {
+  scr_gen_slot_reset(s);
+  s->kind = SCR_EXC_REF;
+  s->payload = v;
+  s->release_fn = release;
+}
+
+void scr_gen_in_f64(ScrGen *g, double v) { scr_gen_slot_f64(&g->in, v); }
+void scr_gen_in_bool(ScrGen *g, bool v) { scr_gen_slot_bool(&g->in, v); }
+void scr_gen_in_ref(ScrGen *g, void *v, void (*release)(void *)) {
+  scr_gen_slot_ref(&g->in, v, release);
+}
+void scr_gen_in_none(ScrGen *g) { scr_gen_slot_reset(&g->in); }
+
+void scr_gen_ret_f64(ScrGen *g, double v) { scr_gen_slot_f64(&g->ret, v); }
+void scr_gen_ret_bool(ScrGen *g, bool v) { scr_gen_slot_bool(&g->ret, v); }
+void scr_gen_ret_ref(ScrGen *g, void *v, void (*release)(void *)) {
+  scr_gen_slot_ref(&g->ret, v, release);
+}
+void scr_gen_ret_none(ScrGen *g) { scr_gen_slot_reset(&g->ret); }
+
+void scr_gen_out_f64(ScrGen *g, double v) { scr_gen_slot_f64(&g->out, v); }
+void scr_gen_out_bool(ScrGen *g, bool v) { scr_gen_slot_bool(&g->out, v); }
+void scr_gen_out_ref(ScrGen *g, void *v, void (*release)(void *)) {
+  scr_gen_slot_ref(&g->out, v, release);
+}
+
+/* Takes MOVE the slot's value out (the slot resets — a done generator's
+ * later resumes answer NONE, JS's undefined). */
+static double scr_gen_slot_take_f64(ScrGenSlot *s) {
+  double v = s->kind == SCR_EXC_F64 ? s->f64 : 0;
+  s->kind = SCR_EXC_NONE;
+  return v;
+}
+static bool scr_gen_slot_take_bool(ScrGenSlot *s) {
+  bool v = s->kind == SCR_EXC_BOOL ? s->b : false;
+  s->kind = SCR_EXC_NONE;
+  return v;
+}
+static void *scr_gen_slot_take_ref(ScrGenSlot *s) {
+  void *v = s->kind == SCR_EXC_REF ? s->payload : NULL;
+  s->kind = SCR_EXC_NONE;
+  s->payload = NULL;
+  s->release_fn = NULL;
+  return v;
+}
+
+bool scr_gen_out_has(ScrGen *g) { return g->out.kind != SCR_EXC_NONE; }
+double scr_gen_take_out_f64(ScrGen *g) { return scr_gen_slot_take_f64(&g->out); }
+bool scr_gen_take_out_bool(ScrGen *g) { return scr_gen_slot_take_bool(&g->out); }
+void *scr_gen_take_out_ref(ScrGen *g) { return scr_gen_slot_take_ref(&g->out); }
+
+/* The running fiber's generator — the body-side helpers' anchor. */
+static ScrGen *scr_gen_self(void) {
+  ScrGen *g = scr_current != NULL ? scr_current->gen : NULL;
+  if (g == NULL) {
+    fputs("scriptc: internal error: yield outside a generator\n", stderr);
+    abort();
+  }
+  return g;
+}
+
+double scr_gen_take_in_f64(void) { return scr_gen_slot_take_f64(&scr_gen_self()->in); }
+bool scr_gen_take_in_bool(void) { return scr_gen_slot_take_bool(&scr_gen_self()->in); }
+void *scr_gen_take_in_ref(void) { return scr_gen_slot_take_ref(&scr_gen_self()->in); }
+
+ScrDyn *scr_gen_delegate_resume(ScrDyn *(*next_ref)(void *),
+                                ScrDyn *(*return_ref)(void *),
+                                ScrDyn *(*caught_ref)(const ScrCaught *)) {
+  ScrGen *g = scr_gen_self();
+  ScrExcCell *cell = scr_exc_current_cell();
+  int mode = cell->kind == SCR_EXC_GENRET ? 1 : cell->kind != SCR_EXC_NONE ? 2 : 0;
+  ScrDyn *value;
+  if (mode == 2) {
+    ScrCaught *caught = scr_exc_take();
+    value = caught_ref(caught);
+    scr_caught_release(caught);
+  } else {
+    if (mode == 1) scr_exc_clear();
+    ScrGenSlot *slot = mode == 1 ? &g->ret : &g->in;
+    ScrDyn *(*convert)(void *) = mode == 1 ? return_ref : next_ref;
+    switch (slot->kind) {
+    case SCR_EXC_F64: value = scr_dyn_new_num(slot->f64); break;
+    case SCR_EXC_BOOL: value = scr_dyn_new_bool(slot->b); break;
+    case SCR_EXC_REF: value = convert(slot->payload); break;
+    default: value = scr_dyn_retain(scr_dyn_undefined()); break;
+    }
+    scr_gen_slot_reset(slot);
+  }
+  if (scr_exc_pending()) {
+    scr_dyn_release(value);
+    return NULL;
+  }
+  ScrDyn *result = scr_dyn_new_obj();
+  ScrDyn *kind = scr_dyn_new_num(mode);
+  scr_dyn_obj_set(result, "kind", 4, kind);
+  scr_dyn_obj_set(result, "value", 5, value);
+  /* Object insertion takes ownership of both completion fields. */
+  return result;
+}
+
+/* yield: park the value in OUT and hop back to the resumer. Control
+ * returns here at the next resume — possibly with an injected .throw
+ * payload or the GENRET sentinel pending (the emitted check handles it). */
+static void scr_gen_yield_switch(void) {
+  ScrFiber *self = scr_current;
+  scr_switch(&self->ctx, self->return_to, NULL);
+}
+void scr_gen_yield_f64(double v) {
+  scr_gen_slot_f64(&scr_gen_self()->out, v);
+  scr_gen_yield_switch();
+}
+void scr_gen_yield_bool(bool v) {
+  scr_gen_slot_bool(&scr_gen_self()->out, v);
+  scr_gen_yield_switch();
+}
+void scr_gen_yield_ref(void *v, void (*release)(void *)) {
+  scr_gen_slot_ref(&scr_gen_self()->out, v, release);
+  scr_gen_yield_switch();
+}
+
+void scr_async_gen_hop_done(void) {
+  ScrGen *g = scr_gen_self();
+  if (!g->is_async) {
+    fputs("scriptc: internal error: async-generator hop on sync generator\n", stderr);
+    abort();
+  }
+  g->success_hop_done = true;
+}
+
+bool scr_exc_genret_pending(void) {
+  return scr_exc_current_cell()->kind == SCR_EXC_GENRET;
+}
+
+void scr_gen_ret_to_out(ScrGen *g) {
+  scr_gen_slot_reset(&g->out);
+  g->out = g->ret;
+  g->ret.kind = SCR_EXC_NONE;
+  g->ret.payload = NULL;
+  g->ret.release_fn = NULL;
+}
+
+/* The consumer→fiber hop shared by every resume mode: switches in, and on
+ * the way back moves a pending body exception into the RESUMER's cell
+ * (synchronous propagation — the await-rethrow analog) and tears down a
+ * finished fiber. The gen is retained across the hop so a body that
+ * releases the consumer's last reference cannot free it mid-run. */
+static void scr_gen_switch_in(ScrGen *g) {
+  ScrFiber *f = g->fiber;
+  scr_gen_retain(g);
+  g->state = SCR_GEN_RUNNING;
+#ifdef _WIN32
+  ScrCtx here = scr_win_self();
+#elif defined(__wasi__)
+  ScrCtx here = 0;
+#else
+  ucontext_t here;
+#endif
+  f->return_to = &here;
+  ScrFiber *me = scr_current;
+#ifdef __wasi__
+  scr_current = f;
+  scr_exc_swap_cell(&f->exc);
+  scr_als_active = &f->als;
+  if (f->coro == NULL) f->entry(f, f->argpack);
+  else scr_wasi_coro_resume(f->coro);
+  scr_current = me;
+  scr_exc_swap_cell(me != NULL ? &me->exc : NULL);
+  scr_als_active = me != NULL ? &me->als : &scr_als_main_slot;
+#else
+  scr_switch(&here, &f->ctx, f);
+  /* Back on the consumer: restore identity + the consumer's cell (the
+   * yield/finish switch targeted NULL — main's cell). */
+  scr_current = me;
+  scr_exc_swap_cell(me != NULL ? &me->exc : NULL);
+#endif
+  if (f->done) {
+    g->state = SCR_GEN_DONE;
+    if (f->exc.kind != SCR_EXC_NONE) {
+      if (g->is_async) {
+        /* Async-generator failures reject the active request; they never
+         * throw synchronously from next/return/throw. */
+        scr_gen_exc_move(&g->async_error, &f->exc);
+      } else {
+        /* The sync body's exception escapes at the resume site. */
+        ScrExcCell *mine = scr_exc_current_cell();
+        scr_gen_exc_move(mine, &f->exc);
+      }
+    }
+    scr_fiber_destroy(f);
+    g->fiber = NULL;
+    scr_fibers_live--;
+  } else {
+    /* An async generator can return to its resumer because it yielded OR
+     * because await parked it. OUT distinguishes the two. */
+    g->state = g->is_async && !scr_gen_out_has(g)
+        ? SCR_GEN_AWAITING
+        : SCR_GEN_SUSPENDED;
+  }
+  scr_gen_release(g);
+}
+
+void scr_gen_resume(ScrGen *g) {
+  switch (g->state) {
+  case SCR_GEN_RUNNING: {
+    static const char sc_m[] = "Generator is already running";
+    scr_throw_error_msg(SCR_ERR_TYPE, sc_m, sizeof sc_m - 1);
+    return;
+  }
+  case SCR_GEN_DONE:
+    /* next() on a done generator: { value: undefined, done: true } —
+     * OUT is NONE (the completing resume took it). */
+    return;
+  default:
+    scr_gen_switch_in(g);
+  }
+}
+
+void scr_gen_resume_return(ScrGen *g) {
+  switch (g->state) {
+  case SCR_GEN_RUNNING: {
+    static const char sc_m[] = "Generator is already running";
+    scr_throw_error_msg(SCR_ERR_TYPE, sc_m, sizeof sc_m - 1);
+    return;
+  }
+  case SCR_GEN_DONE:
+    scr_gen_ret_to_out(g);
+    return;
+  case SCR_GEN_UNSTARTED:
+    /* The body never runs: tear the fiber down cleanly (drop the packed
+     * arguments) and complete with the parked value. */
+    if (g->drop_args != NULL) g->drop_args(g->fiber->argpack);
+    else free(g->fiber->argpack);
+    scr_fiber_destroy(g->fiber);
+    g->fiber = NULL;
+    scr_fibers_live--;
+    g->state = SCR_GEN_DONE;
+    scr_gen_ret_to_out(g);
+    return;
+  default:
+    /* Suspended at a yield: inject the sentinel (an earlier .return whose
+     * unwind a finally-yield parked leaves it already set) and resume. */
+    if (g->fiber->exc.kind == SCR_EXC_NONE) g->fiber->exc.kind = SCR_EXC_GENRET;
+    scr_gen_switch_in(g);
+  }
+}
+
+void scr_gen_resume_throw(ScrGen *g) {
+  switch (g->state) {
+  case SCR_GEN_RUNNING: {
+    /* Replace the parked payload with Node's reentrancy TypeError. */
+    static const char sc_m[] = "Generator is already running";
+    scr_throw_error_msg(SCR_ERR_TYPE, sc_m, sizeof sc_m - 1);
+    return;
+  }
+  case SCR_GEN_DONE:
+    /* .throw on a done generator: it stays done; the payload stays
+     * pending in the caller — the call site's check rethrows it. */
+    return;
+  case SCR_GEN_UNSTARTED:
+    /* The body never runs; the generator becomes done and the payload
+     * stays pending in the caller (probed Node behavior). */
+    if (g->drop_args != NULL) g->drop_args(g->fiber->argpack);
+    else free(g->fiber->argpack);
+    scr_fiber_destroy(g->fiber);
+    g->fiber = NULL;
+    scr_fibers_live--;
+    g->state = SCR_GEN_DONE;
+    return;
+  default: {
+    /* Move the caller's pending payload into the fiber's cell (an
+     * earlier sentinel parked by a finally-yield is replaced — the
+     * injected throw wins, like a throw inside that finally). */
+    ScrExcCell *mine = scr_exc_current_cell();
+    ScrExcCell *dst = &g->fiber->exc;
+    dst->kind = mine->kind;
+    dst->f64 = mine->f64;
+    dst->b = mine->b;
+    dst->payload = mine->payload;
+    dst->retain_fn = mine->retain_fn;
+    dst->release_fn = mine->release_fn;
+    dst->trace_fn = mine->trace_fn;
+    mine->kind = SCR_EXC_NONE;
+    mine->payload = NULL;
+    mine->trace_fn = NULL;
+    scr_gen_switch_in(g);
+  }
+  }
+}
+
+/* ── async generators ────────────────────────────────────────────────
+ * The body is the same suspended generator fiber, now allowed to await.
+ * Each resume request owns a promise and a retained generator. The first
+ * idle request starts synchronously; successful yield/return settlement
+ * takes AsyncGeneratorYield's Await hop (the emitter marks a yielded
+ * promise whose await already supplied it), while thrown completions reject
+ * immediately. Requests queued behind an active one begin in that
+ * settlement continuation, preserving Node's ordering. */
+
+static void scr_async_gen_start_queued(ScrGen *g);
+
+static void scr_async_gen_schedule_settle(ScrGen *g) {
+  if (g->settle_queued) return;
+  ScrFiber *job = calloc(1, sizeof *job);
+  if (!job) scr_oom();
+  job->gen_job = scr_gen_retain(g);
+  g->settle_queued = true;
+  scr_ready_push(job);
+}
+
+static void scr_async_gen_finish_active(ScrGen *g, bool rejected) {
+  ScrPromise *p = g->active_request;
+  if (p == NULL) {
+    fputs("scriptc: internal error: async generator settled without a request\n", stderr);
+    abort();
+  }
+  g->active_request = NULL;
+  if (rejected) {
+    scr_promise_reject_from_cell(p, &g->async_error);
+    scr_promise_settle_wake(p);
+  } else {
+    g->settle_async(g, p);
+  }
+  scr_promise_release(p);
+  /* Every queued request retains its generator until settlement. */
+  scr_gen_release(g);
+}
+
+static void scr_async_gen_capture_caller_error(ScrGen *g) {
+  ScrExcCell *cell = scr_exc_current_cell();
+  if (cell->kind != SCR_EXC_NONE) scr_gen_exc_move(&g->async_error, cell);
+}
+
+static void scr_async_gen_run_active(ScrGen *g, ScrAsyncGenRequest *req) {
+  const bool was_done = g->state == SCR_GEN_DONE;
+  if (req->mode == SCR_AG_NEXT) {
+    scr_gen_slot_reset(&g->in);
+    g->in = req->arg;
+    memset(&req->arg, 0, sizeof req->arg);
+    scr_gen_resume(g);
+  } else if (req->mode == SCR_AG_RETURN) {
+    scr_gen_slot_reset(&g->ret);
+    g->ret = req->arg;
+    memset(&req->arg, 0, sizeof req->arg);
+    scr_gen_resume_return(g);
+  } else {
+    scr_gen_exc_move(scr_exc_current_cell(), &req->thrown);
+    scr_gen_resume_throw(g);
+  }
+  scr_async_gen_capture_caller_error(g);
+
+  if (g->async_error.kind != SCR_EXC_NONE) {
+    scr_async_gen_finish_active(g, true);
+    return;
+  }
+  if (g->state == SCR_GEN_AWAITING) return;
+
+  /* next() on an already-completed generator produces its undefined done
+   * result immediately. Yield, body return, and return(value) all pass
+   * through AsyncGeneratorYield's successful Await step. */
+  if (was_done && req->mode == SCR_AG_NEXT) {
+    scr_async_gen_finish_active(g, false);
+  } else if (g->success_hop_done) {
+    g->success_hop_done = false;
+    scr_async_gen_finish_active(g, false);
+  } else {
+    scr_async_gen_schedule_settle(g);
+  }
+}
+
+static void scr_async_gen_start_queued(ScrGen *g) {
+  /* Keep the object alive while an immediately-settled request releases
+   * its own retain and the loop inspects the next queue node. */
+  scr_gen_retain(g);
+  while (g->active_request == NULL && g->requests_head != NULL) {
+    ScrAsyncGenRequest *req = g->requests_head;
+    g->requests_head = req->next;
+    if (g->requests_head == NULL) g->requests_tail = NULL;
+    g->active_request = req->promise; /* request ownership transfers */
+    scr_async_gen_run_active(g, req);
+    scr_gen_slot_reset(&req->arg);
+    scr_gen_exc_reset(&req->thrown);
+    free(req);
+    /* Awaiting or deferred successful settlement keeps this request
+     * active. Immediate done/error settlement cleared it and can consume
+     * the next queued request in this same turn. */
+    if (g->active_request != NULL) break;
+  }
+  scr_gen_release(g);
+}
+
+static ScrPromise *scr_async_gen_enqueue(ScrGen *g, int mode,
+                                         ScrGenSlot *arg,
+                                         ScrExcCell *thrown) {
+  if (!g->is_async || g->settle_async == NULL) {
+    fputs("scriptc: internal error: async resume of a sync generator\n", stderr);
+    abort();
+  }
+  ScrAsyncGenRequest *req = calloc(1, sizeof *req);
+  if (!req) scr_oom();
+  req->mode = mode;
+  if (arg != NULL) {
+    req->arg = *arg;
+    memset(arg, 0, sizeof *arg);
+  }
+  if (thrown != NULL) scr_gen_exc_move(&req->thrown, thrown);
+  req->promise = scr_promise_new();
+  /* The request owns the original promise reference and one generator
+   * reference; the caller receives a second promise reference. */
+  ScrPromise *result = scr_promise_retain(req->promise);
+  scr_gen_retain(g);
+  if (g->requests_tail != NULL) g->requests_tail->next = req;
+  else g->requests_head = req;
+  g->requests_tail = req;
+  if (g->active_request == NULL && !g->settle_queued) {
+    scr_async_gen_start_queued(g);
+  }
+  return result;
+}
+
+ScrPromise *scr_async_gen_next_none(ScrGen *g) {
+  ScrGenSlot arg = {0};
+  return scr_async_gen_enqueue(g, SCR_AG_NEXT, &arg, NULL);
+}
+
+ScrPromise *scr_async_gen_next_f64(ScrGen *g, double value) {
+  ScrGenSlot arg = {0};
+  scr_gen_slot_f64(&arg, value);
+  return scr_async_gen_enqueue(g, SCR_AG_NEXT, &arg, NULL);
+}
+
+ScrPromise *scr_async_gen_next_bool(ScrGen *g, bool value) {
+  ScrGenSlot arg = {0};
+  scr_gen_slot_bool(&arg, value);
+  return scr_async_gen_enqueue(g, SCR_AG_NEXT, &arg, NULL);
+}
+
+ScrPromise *scr_async_gen_next_ref(ScrGen *g, void *value,
+                                   void (*release)(void *)) {
+  ScrGenSlot arg = {0};
+  scr_gen_slot_ref(&arg, value, release);
+  return scr_async_gen_enqueue(g, SCR_AG_NEXT, &arg, NULL);
+}
+
+ScrPromise *scr_async_gen_return_none(ScrGen *g) {
+  ScrGenSlot arg = {0};
+  return scr_async_gen_enqueue(g, SCR_AG_RETURN, &arg, NULL);
+}
+
+ScrPromise *scr_async_gen_return_f64(ScrGen *g, double value) {
+  ScrGenSlot arg = {0};
+  scr_gen_slot_f64(&arg, value);
+  return scr_async_gen_enqueue(g, SCR_AG_RETURN, &arg, NULL);
+}
+
+ScrPromise *scr_async_gen_return_bool(ScrGen *g, bool value) {
+  ScrGenSlot arg = {0};
+  scr_gen_slot_bool(&arg, value);
+  return scr_async_gen_enqueue(g, SCR_AG_RETURN, &arg, NULL);
+}
+
+ScrPromise *scr_async_gen_return_ref(ScrGen *g, void *value,
+                                     void (*release)(void *)) {
+  ScrGenSlot arg = {0};
+  scr_gen_slot_ref(&arg, value, release);
+  return scr_async_gen_enqueue(g, SCR_AG_RETURN, &arg, NULL);
+}
+
+ScrPromise *scr_async_gen_throw(ScrGen *g) {
+  ScrExcCell thrown = {0};
+  scr_gen_exc_move(&thrown, scr_exc_current_cell());
+  return scr_async_gen_enqueue(g, SCR_AG_THROW, NULL, &thrown);
+}
+
+static void scr_async_gen_job(ScrGen *g) {
+  g->settle_queued = false;
+  if (g->active_request == NULL || g->async_error.kind != SCR_EXC_NONE) {
+    fputs("scriptc: internal error: invalid async generator settlement job\n", stderr);
+    abort();
+  }
+  scr_async_gen_finish_active(g, false);
+  scr_async_gen_start_queued(g);
+}
+
+static void scr_async_gen_resume_ready(ScrGen *g) {
+  if (!g->is_async || g->active_request == NULL) {
+    fputs("scriptc: internal error: async generator resumed without a request\n", stderr);
+    abort();
+  }
+  /* The active request normally owns the generator. Keep an independent
+   * reference across settlement, which releases that request ownership. */
+  scr_gen_retain(g);
+  scr_gen_switch_in(g);
+  if (g->async_error.kind != SCR_EXC_NONE) {
+    scr_async_gen_finish_active(g, true);
+    scr_async_gen_start_queued(g);
+  } else if (g->state != SCR_GEN_AWAITING) {
+    if (g->success_hop_done) {
+      g->success_hop_done = false;
+      scr_async_gen_finish_active(g, false);
+      scr_async_gen_start_queued(g);
+    } else {
+      scr_async_gen_schedule_settle(g);
+    }
+  }
+  scr_gen_release(g);
+}