@solidjs/signals 2.0.0-rc.0 → 2.0.0-rc.10

package/dist/observe/core/scheduler.js

@@ -0,0 +1,1547 @@
+import { EFFECT_RENDER, EFFECT_USER, STATUS_PENDING, CONFIG_ADOPTED_UNFLUSHED, NOT_PENDING, REACTIVE_MANUAL_WRITE, STATUS_UNINITIALIZED, REACTIVE_OPTIMISTIC_DIRTY, REACTIVE_IN_HEAP_HEIGHT, REACTIVE_DISPOSED, CONFIG_HELD_TRUTH, CONFIG_HAS_COMPANIONS, EFFECT_TRACKED, CONFIG_INPUTS_PUBLISHED, REACTIVE_ZOMBIE, CONFIG_LANE_FRAME, CONFIG_HAS_LANE, CONFIG_HAS_SNAPSHOT, REACTIVE_RECOMPUTING_DEPS, REACTIVE_MISSED_WAKE, CONFIG_IN_SNAPSHOT_SCOPE, REACTIVE_SNAPSHOT_STALE, CONFIG_SLOT_NODE, CONFIG_HELD_CHILDREN, REACTIVE_REASK, CONFIG_AUTHORITATIVE_READ } from "./constants.js";
+
+import { attrHooks } from "./attribution-hooks.js";
+
+import { resyncUnflushedCompanions, currentOptimisticLane, slotUnobservedHook, markUnflushedStaged, hasActiveOverride } from "./core.js";
+
+import { DEV, GRAPH_SIZE_WARN_AT, noteFanOut } from "./dev.js";
+
+import { NotReadyError } from "./error.js";
+
+import { sweepDormant, trimStaleDeps } from "./graph.js";
+
+import { runHeap, deleteFromHeap, enqueueSub } from "./heap.js";
+
+import { activeLanes, assignOrMergeLane, findLane } from "./lanes.js";
+
+export { getOrCreateLane, mergeLanes, resolveLane } from "./lanes.js";
+
+import { devCheckFlushStart, devCheckActiveOverrides, devCensusCompanions, devCheckQuiescent } from "./invariants.js";
+
+const transitions = new Set;
+
+const dirtyQueue = {
+    eE: new Array(2e3).fill(undefined),
+    tE: false,
+    st: 0,
+    EE: 0
+};
+
+const zombieQueue = {
+    eE: new Array(2e3).fill(undefined),
+    tE: false,
+    st: 0,
+    EE: 0
+};
+
+/** runHeap callback that discards a queued zombie recompute instead of running
+ * it: unlink pure recompute entries; strip just the recompute bit from dirtied
+ * height-adjust entries so their height work still happens. A zombie dirtied
+ * through the lane channel (OPTIMISTIC_DIRTY — an override or a `latest()`
+ * companion) runs instead (#3444): a zombie renders mainline until the commit
+ * disposes it, and the lane's values ARE the mainline frame — the still-visible
+ * branch a held `Show` is removing showed the old `latest(count)` beside the
+ * new one outside. Its pass runs under the lane and its run lands on the
+ * lane's effect queue, so a held lane defers it exactly as it defers every
+ * other reader's. */ function cancelZombieRecompute(e) {
+    if (e.oe & REACTIVE_OPTIMISTIC_DIRTY && !laneZombie(e)) return GlobalQueue.nt(e);
+    if (e.oe & REACTIVE_IN_HEAP_HEIGHT) e.oe &= -140; else {
+        deleteFromHeap(e, zombieQueue);
+        e.oe &= -132;
+    }
+}
+
+/** A member of a parked LANE frame (CONFIG_LANE_FRAME on the owner whose pass
+ * parked it, #3662). The #3444 exception is for transaction zombies; a lane
+ * frame's member is retired by the run that carries the lane's values, so its
+ * lane-channel recompute is cancelled with the rest — run, it republished the
+ * retired frame under those values. Mainline writes still reach it (#3463). */ function laneZombie(e) {
+    let t = e;
+    while (t !== null && t.oe & REACTIVE_ZOMBIE) t = t._parent;
+    return t !== null && (t.C & CONFIG_LANE_FRAME) !== 0;
+}
+
+let clock = 0;
+
+let activeTransition = null;
+
+let scheduled = false;
+
+let halted = false;
+
+let haltNotified = false;
+
+let syncDepth = 0;
+
+let projectionWriteActive = false;
+
+/** > 0 while an action's generator body is on the stack (the synchronous
+ * slice between yields). Maintained by action.ts around `it.next()`. */ let actionStepDepth = 0;
+
+function enterActionStep() {
+    actionStepDepth++;
+}
+
+function exitActionStep() {
+    actionStepDepth--;
+}
+
+// Store property nodes whose last subscriber left while they carried state
+// the backing cannot reconstruct — an optimistic override (overrides live on
+// nodes, over a clone the setter discards) or a staged write. Releasing the
+// slot then would drop the override: an optimistic store key read `0` the
+// moment its only reader gated away while the action was live (S7). Swept
+// after each flush — a node still without subs whose override and staging
+// have resolved is released through the slot hook; one that regained a
+// subscriber leaves the set.
+const transientStoreNodes = new Set;
+
+/** Slot hook's deferral: release this node when its carried state resolves. */ function deferSlotRelease(e) {
+    transientStoreNodes.add(e);
+}
+
+function canUseSimpleSyncFlush(e) {
+    const t = e.m;
+    return transitions.size === 0 && activeLanes.size === 0 && e.Xn.length === 0 && t.bn.length === 0 && t.A.length === 0 && t.En.size === 0 && transientStoreNodes.size === 0 && pendingRearms.size === 0;
+}
+
+function sweepTransientStoreNodes() {
+    if (transientStoreNodes.size === 0) return;
+    for (const e of transientStoreNodes) {
+        if (e.u !== null) {
+            transientStoreNodes.delete(e);
+            continue;
+        }
+        if (e._e !== NOT_PENDING) continue;
+        if (e.o?.we !== undefined && e.o?.we !== NOT_PENDING) continue;
+        // A live affects() mark keeps the node addressable: sweeping it would
+        // detach the refcount from the slot (a fresh probe would upsert a new,
+        // unmarked node for the same property).
+                if (e.o?.t) continue;
+        transientStoreNodes.delete(e);
+        if (e.C & CONFIG_SLOT_NODE) slotUnobservedHook(e); else e.o?.bt?.();
+    }
+}
+
+/**
+ * Toggles the dev-mode "must be inside a `<Loading>` boundary" enforcement
+ * window. Only `render()` calls this — wrapping the initial mount so that a
+ * top-level uncaught async read surfaces the diagnostic. Not part of the
+ * user-facing API.
+ *
+ * @internal
+ */ function enforceLoadingBoundary(e) {}
+
+function setProjectionWriteActive(e) {
+    projectionWriteActive = e;
+}
+
+/**
+ * Ambient work IS a transaction: the global queue always carries one
+ * current-transaction-shaped batch (`globalQueue._batch`). With no transition
+ * active, registrations (pending commits, optimistic nodes, affects marks,
+ * optimistic stores) land in a plain ambient batch that the plain flush
+ * finalizes; when a transition initializes it adopts the ambient batch's
+ * contents and `_batch` becomes the transition itself, so later registrations
+ * land there directly — no per-field aliasing.
+ */ function createBatch() {
+    return {
+        Ge: clock,
+        In: [],
+        ae: new Map,
+        bn: [],
+        A: [],
+        En: new Set,
+        tt: [],
+        ei: {
+            ti: [ [], [] ],
+            Xn: []
+        },
+        Dt: false,
+        Tt: new Set,
+        ht: null
+    };
+}
+
+function mergeTransitionState(e, t) {
+    if (attrHooks !== null) attrHooks.transitionMerged(e, t);
+    t.Dt = e;
+    e.tt.push(...t.tt);
+    e.et ||= t.et;
+    for (const n of activeLanes) if (n.me === t) n.me = e;
+    if (t.bn.length) {
+        // Move (don't copy): the global queue's batch may still be the outgoing
+        // transition, and the adoption pass in initTransition would re-push its
+        // contents into the target — duplicating every entry.
+        e.bn.push(...t.bn);
+        t.bn.length = 0;
+    }
+    if (t.A.length) {
+        // Move (don't copy): the global queue's batch may still be the outgoing
+        // transition, and the adoption pass in initTransition would re-push its
+        // contents into the target — double-releasing every mark.
+        e.A.push(...t.A);
+        t.A.length = 0;
+    }
+    for (const n of t.En) e.En.add(n);
+    for (const [n, i] of t.ae) {
+        let t = e.ae.get(n);
+        if (!t) e.ae.set(n, t = new Set);
+        for (const e of i) t.add(e);
+    }
+    for (const n of t.Tt) e.Tt.add(n);
+    if (t.ht) (e.ht ??= []).push(...t.ht);
+}
+
+/**
+ * Flip-entanglement (#3164 follow-up): `until()` is a declaration of
+ * relatedness — the predicate names the condition that confirms the awaiting
+ * transaction. When the predicate settles truthy, every live foreign
+ * transition whose staged write it read IS the confirming event by the
+ * user's own definition, so it merges into the awaiting transaction and
+ * reveals at the joint settle — the cross-primitive twin of the family fold
+ * (a landing on an optimism-carrying family joins the retaining
+ * transaction). Non-flipping updates never pass through here: falsy
+ * evaluations don't entangle, so unrelated traffic on the watched sources
+ * reveals freely on its own schedule.
+ *
+ * Runs inside the predicate's compute (pure phase) — the confirming
+ * transition's stamps are still live and its commit decision hasn't run, so
+ * the merge lands before any reveal. Only the tree-shaken graphs that call
+ * `until()` retain this.
+ */ function entangleConfirmingTransitions(e, t) {
+    t = currentTransition(t);
+    if (t.Dt === true) return;
+    // The confirming evidence is a dep whose value is STAGED (pending,
+    // uncommitted) at flip evaluation — committed deps are public already and
+    // carry nothing to entangle. A staged dep lives in one of two carriers: a
+    // stamped transition, or the queue's current batch (ambient registrations
+    // don't stamp; "ambient work IS a transaction" — the batch is the
+    // carrier). The entangle STEALS the carrier's staged cargo — its pending
+    // nodes move (re-stamped) into the awaiting transaction and reveal at its
+    // settle — but never the carrier itself: its async reporters, actions,
+    // and stashes are its own future (a live stream's flight must not chain
+    // the awaiting transaction to landings that haven't happened; a merged
+    // reporter deadlocked exactly that way).
+        let n = false;
+    for (let i = e.Ee; i !== null; i = i.Ie) {
+        const r = i.De;
+        if (r._e !== NOT_PENDING) {
+            const e = r.me;
+            const i = e != null ? currentTransition(e) : null;
+            // Skip the awaiting transaction's own cargo (t === target: a
+            // fold-staged landing or a write the action itself issued — hold and
+            // reveal already correct) and dead carriers. Ambient-batch staging
+            // (t === null) must leave the batch NOW — it commits at this flush's
+            // end, which would reveal the confirmation under the live optimism
+            // it just confirmed.
+                        const s = i === null ? currentBatch.In : i !== t && i.Dt !== true ? i.In : null;
+            if (s !== null) n = stealEntangledCargo(s, t) || n;
+        }
+        if (i === e.Je) break;
+    }
+    // The steal never activates the awaiting transaction: the predicate can
+    // flip inside another transaction's finalize heap, and adopting the queue
+    // batch there hands the stolen cargo to that finalize's commit sweep — a
+    // premature reveal at a foreign settle. Subscribers that computed against
+    // the pre-steal world were re-dirtied by the steal itself, so this
+    // flush's applies paint the mid-hold view (committed for lane and stale
+    // readers; a deriving reader is held with the cargo); the cargo commits
+    // at the awaiting transaction's own settle.
+}
+
+/** Move a confirming carrier's staged nodes into the awaiting transaction:
+ * re-stamp and arm the held-truth mask (override-covered nodes skip it —
+ * the override already hides their staged value per A17, and is usually
+ * the very optimism this confirmation settles); the mask's commit
+ * registers the settle-side post-revert wake. The carrier's array is
+ * emptied so its own commit point commits none of the stolen cargo.
+ *
+ * EFFECT subs of stolen nodes re-run: any that recomputed against the
+ * staging BEFORE the steal (the carrier's landing notified them as a plain
+ * write) hold a private result derived from the carrier's world — one
+ * that the next paint gate (stash-point lane run, a foreign flush's
+ * completion drain) would apply as the carrier's. Re-running them now that
+ * the cargo is the awaiting transaction's re-routes each by its posture in
+ * this same heap pass: a lane pass keeps committed (the mask), a stale
+ * render effect keeps committed and registers for the reveal's replay
+ * (stale-of-foreign), and a deriving user effect is served the truth and
+ * held with the transaction (A29). PURE computeds are deliberately NOT
+ * re-run: a staged value of theirs is itself stolen cargo — held with the
+ * transaction, so ordinary readers already serve their committed value —
+ * while re-running them would re-derive the OLD world and re-stage it over
+ * the held truth. The reveal re-notifies (commitPendingNodes), which is
+ * when they re-derive for real. */ function stealEntangledCargo(e, t) {
+    if (e === t.In || e.length === 0) return false;
+    for (let n = 0; n < e.length; n++) {
+        const i = e[n];
+        i.me = t;
+        t.In.push(i);
+        // Override-covered nodes stay silent AND unmasked: the override is the
+        // display (A17 — its staging never notified, its revert will), so their
+        // subs saw nothing and re-running one would break the silence with a
+        // duplicate fire of an unchanged view.
+                if (!hasActiveOverride(i)) {
+            i.C |= CONFIG_HELD_TRUTH;
+            for (let e = i.u; e !== null; e = e.Pe) {
+                const t = e.Ae;
+                if (t.He && !(t.C & CONFIG_AUTHORITATIVE_READ)) enqueueSub(t);
+            }
+        }
+    }
+    e.length = 0;
+    transitions.add(t);
+    return true;
+}
+
+/** `schedule()` armed `scheduled` but withheld the microtask because a
+ * projection draft was writing (see below). Consumed by `scheduleWithheld`. */ let withheld = false;
+
+function schedule() {
+    if (halted) {
+        notifyHalted();
+        return;
+    }
+    if (scheduled) return;
+    scheduled = true;
+    if (!syncDepth && !globalQueue.$t) {
+        // A projection draft's writes withhold the microtask: a flight's landing
+        // drains them (asyncWrite's flush), and a drain BEFORE the landing tears
+        // the pre- and post-await halves of the run apart (createProjection.async
+        // "notifies only changed paths"). A draft write that no landing will
+        // follow re-arms through scheduleWithheld (proj R37).
+        withheld = projectionWriteActive;
+        if (!withheld) queueMicrotask(flush);
+    }
+}
+
+/** Arm the microtask `schedule()` withheld under projectionWriteActive. The
+ * projection draft calls this after a write made outside its run with no
+ * flight up (proj R37): nothing else will drain, and leaving `scheduled`
+ * armed with no microtask strands the whole scheduler — every later
+ * `schedule()` early-returns — until something calls `flush()` by hand.
+ * (`withheld` was set in the same synchronous slice as this call, so the
+ * syncDepth/_running gates it passed still hold; a stale mark left by a
+ * landing's own flush arms at worst one no-op drain.) */ function scheduleWithheld() {
+    if (withheld) {
+        withheld = false;
+        queueMicrotask(flush);
+    }
+}
+
+/**
+ * Parked transactions whose reporter set changed without a write. A
+ * transaction completes when nothing live reports a flight it waits on, but
+ * the flush only judges the ACTIVE transaction: a parked one is re-entered by
+ * a stamped node's landing or an action's resume. A reporter that stops
+ * counting for another reason — its loading boundary flipped to the fallback
+ * (#3375), or it was disposed by ambient work (#3372) — is neither: the
+ * pruning in `reporterBlocksSource` would drop it at the next check, but no
+ * check comes, and the writes held with it stay staged. Such sites record the
+ * transaction here (deduped: one idle pass per transaction, however many
+ * reporters changed); the flush re-enters it on an otherwise idle pass, so
+ * the re-evaluation adopts no unrelated ambient work.
+ */ const wokenTransitions = [];
+
+/** Wake every parked transaction — for a site that knows a reporter stopped
+ * counting but not whose (a boundary reset). */ function wakeParked() {
+    for (const e of transitions) wokenTransitions.includes(e) || wokenTransitions.push(e);
+    schedule();
+}
+
+/**
+ * Boundaries whose `on` dependencies notified this flush (#3540; a Set:
+ * many notifications, one re-arm). The notification arrives inside a pass,
+ * at the height of the `on` reads — before the readers the write put in
+ * flight are registered — so the re-arm waits for the heap and runs before
+ * the verdict (GlobalQueue.run → drainRearms), still under the write's
+ * transaction: the release is seen by the verdict that follows, and the
+ * staged fallback swap lands with the write's frame.
+ */ const pendingRearms = new Set;
+
+function queueRearm(e) {
+    pendingRearms.add(e);
+    schedule();
+}
+
+/** The drain (see pendingRearms): each boundary decides for itself, from
+ * what is pending under it now, whether re-arming means its fallback or
+ * nothing. Snapshot first: a re-arm can queue another — a write it makes
+ * notifying an `on` that reads it — which is the next pass's. */ function drainRearms() {
+    const e = Array.from(pendingRearms);
+    pendingRearms.clear();
+    for (let t = 0; t < e.length; t++) e[t].se();
+}
+
+/** Transactions a mainline tick has PROPOSED against (A34, #3494): a write to a
+ * node one of them holds — the same value or another — is a second proposal
+ * on a contested node, and the tick reveals with the hold ("both are
+ * suggesting a value; if one finished before the other that would be odd").
+ * Entered at the next flush's start, where the ambient batch is adopted;
+ * never from the write itself, which left `activeTransition` set across the
+ * caller's block and made creation after the write the transaction's (A29). */ const batchJoins = [];
+
+/**
+ * Permanently halts the reactive system. Called when a user error escapes
+ * every boundary — app state is undefined at that point, so scheduling stops
+ * entirely rather than limping along with a half-applied update.
+ */
+/**
+ * The key a root owner carries its client error hook under (`render`'s
+ * `onError`) — registered, so a runtime writes it with no import of the hook
+ * module (core/error-hooks.ts) and no property mangling in the way. Defined
+ * HERE, not there: a runtime that only writes the key must not retain the
+ * hook machinery (pay-for-use).
+ */ const ROOT_ERROR_HOOK = Symbol.for("solid-js/root-error-hook");
+
+function haltReactivity(e) {
+    if (halted) return;
+    halted = true;
+    let t = "[REACTIVITY_HALTED]";
+    // Surface the cause here too: callers rethrow it, but a creation-time throw
+    // unwinds through ancestor recomputes that convert it to status instead of
+    // surfacing it (#2884), so the rethrow alone cannot guarantee visibility.
+    // Where the platform has one, hand the cause to its uncaught-error channel
+    // (`reportError` → `error` event → window.onerror / error monitoring): a
+    // halt that only reaches console.error leaves a page that LOOKS alive with
+    // nothing an app or its telemetry can act on (#3338 — an uncaught throw
+    // during the hydration render). The rethrow may reach the top as well in
+    // the non-swallowed cases; a duplicate report beats a silent one.
+        const n = e !== undefined && globalThis.reportError;
+    n || e === undefined ? console.error(t) : console.error(t, e);
+    n && n(e);
+}
+
+// Logs on the first write after a halt so a frozen interaction is traceable.
+function notifyHalted() {
+    if (haltNotified) return;
+    haltNotified = true;
+    console.error("[REACTIVITY_HALTED]");
+}
+
+/** @internal Test/dev-reload hook. Revives scheduling after a halt. */ function resetErrorHalt() {
+    halted = false;
+    haltNotified = false;
+}
+
+// Identifies one child-traversal pass in `Queue.run` so a rescan after the
+// child list shifts can tell "already run this pass" from "still pending".
+let queueRunToken = 0;
+
+class Queue {
+    _parent=null;
+    ti=[ [], [] ];
+    Xn=[];
+    ni=0;
+    created=clock;
+    addChild(e) {
+        this.Xn.push(e);
+        e._parent = this;
+    }
+    removeChild(e) {
+        const t = this.Xn.indexOf(e);
+        if (t >= 0) {
+            this.Xn.splice(t, 1);
+            e._parent = null;
+        }
+    }
+    notify(e, t, n, i) {
+        if (this._parent) return this._parent.notify(e, t, n, i);
+        return false;
+    }
+    run(e) {
+        if (this.ti[e - 1].length) {
+            const t = this.ti[e - 1];
+            this.ti[e - 1] = [];
+            runQueue(t, e);
+        }
+        // Effects run here can dispose owners, and disposal removes queues from
+        // this list — the running child itself, an earlier sibling, or several at
+        // once. A plain index walk then skips whatever shifted into the cursor.
+        // Stamping each child before it runs makes the pass idempotent, so a shift
+        // can be recovered by rescanning from the front and every child still runs
+        // exactly once. Children appended mid-pass carry a stale stamp and run,
+        // matching the previous live-array behaviour.
+                const t = this.Xn;
+        const n = ++queueRunToken;
+        for (let i = 0; i < t.length; ) {
+            const r = t[i];
+            if (r.ni !== n) {
+                r.ni = n;
+                r.run?.(e);
+                if (t[i] !== r) {
+                    i = 0;
+                    continue;
+                }
+            }
+            i++;
+        }
+    }
+    enqueue(e, t) {
+        if (e) {
+            // Route to lane's effect queue if we're in an optimistic recomputation
+            if (currentOptimisticLane) {
+                const n = findLane(currentOptimisticLane);
+                n.It[e - 1].push(t);
+            } else {
+                this.ti[e - 1].push(t);
+            }
+        }
+        schedule();
+    }
+    stashQueues(e) {
+        // Attribution hook: the parking transition's lane effects have run; its
+        // queues are being stashed. Root call only (children recurse below).
+        if (attrHooks !== null && this === globalQueue) attrHooks.holdEnd();
+        e.ti[0].push(...this.ti[0]);
+        e.ti[1].push(...this.ti[1]);
+        this.ti = [ [], [] ];
+        for (let t = 0; t < this.Xn.length; t++) {
+            let n = this.Xn[t];
+            let i = e.Xn[t];
+            if (!i) {
+                i = {
+                    ti: [ [], [] ],
+                    Xn: []
+                };
+                e.Xn[t] = i;
+            }
+            n.stashQueues(i);
+        }
+    }
+    restoreQueues(e) {
+        this.ti[0].push(...e.ti[0]);
+        this.ti[1].push(...e.ti[1]);
+        for (let t = 0; t < e.Xn.length; t++) {
+            const n = e.Xn[t];
+            let i = this.Xn[t];
+            if (i) i.restoreQueues(n);
+        }
+    }
+}
+
+class GlobalQueue extends Queue {
+    $t=false;
+    // The current transaction-shaped batch: a plain ambient batch while no
+    // transition is active, the active transition itself after initTransition.
+    m=createBatch();
+    static nt;
+    static ut;
+    static Rt;
+    static ii=null;
+    // Store-side hook: drops a keyless affects() mark's identity scope when the
+    // carrier node's last registration releases (wired by store.ts, mirroring
+    // _clearOptimisticStore).
+    static p=null;
+    // affects()-side hooks (wired by affects.ts, mirroring _update): the mark
+    // engine — count/register/release — lives with the feature. Every call site
+    // is gated by state only that module creates, so `!` invocations are safe
+    // once the gate holds.
+    static G=null;
+    static M=null;
+    static j=null;
+    // External-source bridge (wired by enableExternalSource(); null while no
+    // config is active — including after _resetExternalSourceConfig()).
+    static Bt=null;
+    static Jt=null;
+    // Verdict-layer hooks (wired by verdict.ts when isPending()/latest() are
+    // imported; null in apps that never use them). Call sites either guard for
+    // null or sit behind state only the verdict layer can create (`!` is safe
+    // there: `_pendingSignal`/`_latestValueComputed` are only ever assigned by
+    // verdict.ts, and `pendingCheckActive`/`latestReadActive` only flip inside
+    // isPending()/latest()).
+    static Le=null;
+    static Ue=null;
+    static qe=null;
+    static Qn=null;
+    static An=null;
+    static Cn=null;
+    static Hn=null;
+    static At=null;
+    static k=null;
+    static ri=null;
+    // Re-asks probes whose verdict was provisionally suppressed by a fresh read
+    // of a held value, once the transaction gains an async blocker (#3028).
+    static si=null;
+    // Optimistic-engine hooks (wired by core/optimistic.ts via
+    // installOptimisticEngine(), called from verdict.ts / createOptimistic /
+    // createOptimisticStore — every module that can create optimistic state).
+    // Call sites are gated by state only the engine can create: an
+    // `_overrideValue` slot, a lane in `activeLanes`, an `_optimisticNodes`
+    // entry, or a non-null `currentOptimisticLane`, so `!` invocations are safe
+    // once the gate holds.
+    static Ln=null;
+    static Bn=null;
+    static Yn=null;
+    static Wn=null;
+    static Kn=null;
+    /** Patch-channel optimistic drain (next/patch.ts): optimistic emissions
+     * apply at lane-effect timing — visible in flight, unlike the regular
+     * effect queues an action stashes. Injected; null when unused. */
+    static Zn=null;
+    static Pn=null;
+    static Gn=null;
+    /** Is the node routed through a LIVE lane (`resolveLane`)? read()'s reveal
+     * carve-out asks before showing a foreign-held pending node's committed
+     * value: a lane-derived flight's inputs are already revealed through the
+     * lane (#3334). Gated on CONFIG_HAS_LANE, which only the engine sets. */
+    static Rn=null;
+    static On=null;
+    static _t=null;
+    static Ot=null;
+    /** Authoritative-view reader wakeup: installed by until() and refresh() before
+     * their first read. Call sites are gated by CONFIG_AUTHORITATIVE_OBSERVED, which
+     * only such a reader's carve-out read can set, so `!` invocations are safe once
+     * the gate holds (#3303). */
+    static Xt=null;
+    static St=null;
+    /** A18 supersession (#3331): own-source truth `value` landed under an active
+     * override. The engine decides whether the graph re-derives — the value
+     * differs from the override and is not a stale (older-action) answer (mark
+     * the node, demote its lane cascade, notify), or returns to it after an
+     * earlier differing arrival (clear the mark, notify) — and owns the
+     * authoritative-observer wake for a silent confirm. Installed with the
+     * optimistic engine; only reachable on a node that has an override. */
+    static Qe=null;
+    /** The flush's pre-verdict step (#3427): once the transaction's action
+     * bodies have all ended and nothing authoritative is left in flight, the
+     * engine supersedes every override still in force with the truth it
+     * reverts to, so the graph re-derives from it now, as the transaction's
+     * held work, instead of after the flights the overrides fed have landed.
+     * True when it superseded something: the caller re-runs the heap ahead of
+     * the verdict. The engine owns every gate (acted, actions drained, has
+     * overrides, no store edits, no authoritative flight); null without it. */
+    static zn=null;
+    /** read()'s value for a TRACKED reader of a superseded node (#3331): the
+     * staged truth, unless the reader is a stale (render) reader of another
+     * transaction — then the displayed override, as it keeps a foreign
+     * transaction's committed value over its staged write. */
+    /** A tracked read of an active override: the lane outside-view rule
+     * (#3460) and the A18 supersession selection (#3331) — see optimistic.ts. */
+    static Dn=null;
+    /** A lane pass's publish for a memo (#3479, lanes stage): the speculative
+     * result becomes a DERIVED override, `_value` stays committed — see
+     * optimistic.ts laneOverride. Set with the engine, which a lane implies. */
+    static je=null;
+    /** Verdict-layer recompute in progress (companion creation, latest()/
+     * isPending() pulls): never born held — see core.ts enterStagedRead. */
+    static Tn=false;
+    /** setSignal's authoritative (projection-write) landing on an override-
+     * covered node (#3331 store twin): stage the truth for its transaction's
+     * commit whatever its relation to the committed value — a landing equal to
+     * committed still differs from the override — then _supersedeOverride
+     * decides. Installed with the optimistic engine; only reachable on a node
+     * that has an override. */
+    static Vn=null;
+    static Jn=null;
+    flush() {
+        if (this.$t) return;
+        // Fast drain: nothing in flight but plain pending commits — no dirty
+        // computeds, no queued effects, no child queues, no transitions/lanes/
+        // optimistic state. Commit and go; anything a commit hook schedules
+        // (companion snaps, store folds notifying subs) re-arms `scheduled`
+        // below and the outer drain loop takes the full spine next round.
+                if (activeTransition === null && dirtyQueue.EE < dirtyQueue.st && this.ti[0].length === 0 && this.ti[1].length === 0 && this.Xn.length === 0 && !wokenTransitions.length && !batchJoins.length && // a join must drain in its own tick (#3519 review)
+        canUseSimpleSyncFlush(this)) {
+            this.$t = true;
+            try {
+                // A28: companions of nodes written since the last flush mirror the
+                // flushed world — re-synced inside the running window (the write is
+                // "flushed" from here on).
+                resyncUnflushedCompanions();
+                // Sweep first: unobserved() pulls swept nodes out of the dirty heap,
+                // so a dormant memo dirtied in the same tick is reclaimed instead of
+                // recomputed (matching the old inline dispose-on-read counts).
+                                sweepDormant();
+                commitPendingNodes();
+            } finally {
+                this.$t = false;
+            }
+            clock++;
+            scheduled = dirtyQueue.EE >= dirtyQueue.st || this.ti[0].length !== 0 || this.ti[1].length !== 0 || this.m.In.length !== 0;
+            return;
+        }
+        this.$t = true;
+        resyncUnflushedCompanions();
+ // A28, see above
+                try {
+            // The tick proposed against a hold (#3494): adopt its batch into it.
+            // Inside the try: the adoption runs user comparators (the no-proposal
+            // drop), and a throw there must not leave `_running` set.
+            while (batchJoins.length) this.initTransition(batchJoins.pop());
+            if (false) ;
+            // Before runHeap for the same reason as the fast drain above; late
+            // subscribers (an effect reading a swept memo this flush) revive it,
+            // which is the pay-for-use contract.
+                        sweepDormant();
+            runHeap(dirtyQueue, GlobalQueue.nt);
+            // The action bodies are over: the overrides they leave in force
+            // revert at this settle, and the correction is this transaction's
+            // held work — re-derived here, under it, ahead of the verdict — not a
+            // waterfall after the flights the overrides fed (#3427). After the
+            // heap, not before: a synchronous body's own writes (a refresh that
+            // puts an override node's source in flight) are judged applied.
+                        if (activeTransition && GlobalQueue.zn?.(activeTransition)) runHeap(dirtyQueue, GlobalQueue.nt);
+            // Re-arm the boundaries whose `on` notified this pass (pendingRearms):
+            // after the heap — what the notification put in flight is registered —
+            // and before the verdict, under the transaction that carried it. The
+            // boundary releases its hold and stages its fallback swap with the
+            // frame; the heap re-runs so its output pass is staged too, ahead of
+            // the verdict that commits or parks it (#3540).
+                        if (pendingRearms.size) {
+                drainRearms();
+                runHeap(dirtyQueue, GlobalQueue.nt);
+            }
+            if (activeTransition) {
+                // A boundary whose fallback read something not ready is judged before
+                // the verdict, under the transaction (boundaries.ts `_judgeHeld`,
+                // #3540): its output is pending on that read and holds the frame,
+                // and only a sweep re-runs it — the commit sweep, after the verdict
+                // its own read keeps parking. Ready, it stages `_disabled` false with
+                // the frame; the heap re-runs so the output drops the read ahead of
+                // the verdict.
+                if (this.Xn.length) {
+                    checkBoundaryChildren(this, true);
+                    if (dirtyQueue.EE >= dirtyQueue.st) runHeap(dirtyQueue, GlobalQueue.nt);
+                }
+                const e = transitionComplete(activeTransition);
+                if (!e) {
+                    const e = activeTransition;
+                    // Parked: the unchanged passes' inputs are held; their tails stay (A30).
+                                        heldTrims.length = 0;
+                    // When the parking batch IS the transition, all of its writes commit
+                    // only with it — every zombie recompute they queued would run against
+                    // a world the zombie never displays (zombies render mainline until
+                    // commit), so cancel them instead of running them. Only an ambient
+                    // batch's mainline writes (the #2916 shape below) legitimately reach
+                    // zombies here. Height-adjust entries still process normally: a
+                    // dirtied one keeps its height flag and falls through to runHeap's
+                    // adjustHeight path on the next pass of the bucket.
+                                        runHeap(zombieQueue, this.m === e ? cancelZombieRecompute : GlobalQueue.nt);
+                    // Detach: the stashed transition keeps its batch; ambient work that
+                    // follows lands in a fresh one. If the batch is already a separate
+                    // ambient one — action done() restored activeTransition without
+                    // adopting the batch, and an ordinary write landed there before
+                    // the scheduled flush (#2916) — keep it: replacing it would strand
+                    // its queued pending nodes with held _pendingValues forever.
+                                        if (this.m === e) currentBatch = this.m = createBatch();
+                    // Run lane effects immediately (before stashing) - lanes with no pending async
+                                        if (activeLanes.size) {
+                        GlobalQueue.Kn(EFFECT_RENDER);
+                        GlobalQueue.Kn(EFFECT_USER);
+                    }
+                    this.stashQueues(e.ei);
+                    clock++;
+                    // A kept ambient batch may hold pending nodes (#2916): stay
+                    // scheduled so the outer drain loop commits them via the plain
+                    // flush path instead of leaving them until the next natural flush.
+                                        scheduled = dirtyQueue.EE >= dirtyQueue.st || this.m.In.length > 0;
+                    reassignPendingTransition(e.In);
+                    activeTransition = null;
+                    finalizePureQueue(null, true);
+                    return;
+                }
+                const t = activeTransition;
+                const n = this.m;
+                n !== t && n.In.push(...t.In);
+                this.restoreQueues(t.ei);
+                transitions.delete(t);
+                activeTransition = null;
+                reassignPendingTransition(n.In);
+                finalizePureQueue(t);
+                if (n === t) {
+                    // Drop the dead Transition wrapper but keep its (drained) containers
+                    // as the ambient batch — late registrations during finalization live
+                    // there and must survive to the next flush.
+                    const e = createBatch();
+                    e.In = n.In;
+                    e.bn = n.bn;
+                    e.A = n.A;
+                    e.En = n.En;
+                    currentBatch = this.m = e;
+                }
+            } else {
+                if (canUseSimpleSyncFlush(this)) {
+                    commitPendingNodes();
+                    if (dirtyQueue.EE >= dirtyQueue.st) {
+                        runHeap(dirtyQueue, GlobalQueue.nt);
+                        commitPendingNodes();
+                    }
+                } else {
+                    // Parked transactions elsewhere: their owners' zombies render
+                    // mainline until the commit that disposes them, so a mainline write
+                    // reaches them here (#2916, #3463). Commit THIS flush's pending nodes
+                    // first (#3546): a zombie whose owner commits now is disposed by that
+                    // commit and never reruns — the same fate it has when no transaction
+                    // is parked, where this queue is not run at all. Run before the
+                    // commit, it reran and notified its owner through the previous
+                    // pass's dependency tail (kept by A30 until the commit trims it),
+                    // and the owner recomputed a second time with identical inputs,
+                    // creating and disposing one more child per write. Only the zombies
+                    // of actually parked owners survive the commit and rerun; the
+                    // finalize's own commit picks up whatever those reruns stage.
+                    if (transitions.size) {
+                        commitPendingNodes();
+                        runHeap(zombieQueue, GlobalQueue.nt);
+                    }
+                    finalizePureQueue();
+                }
+            }
+            clock++;
+            // Check if finalization added items to the heap (from optimistic reversion).
+            // Finalization may also have ENTERED a transaction (a commit hook, boundary
+            // sweep or recompute wrote a node it owns): effects computed under it
+            // since are its to apply, not this flush's — runEffect leaves them queued
+            // and the next pass parks them with it (#3319). Everything computed
+            // mainline applies now.
+                        scheduled = dirtyQueue.EE >= dirtyQueue.st || activeTransition !== null;
+            // Run lane effects first (for ready lanes), then regular effects
+                        activeLanes.size && GlobalQueue.Kn(EFFECT_RENDER);
+            this.run(EFFECT_RENDER);
+            activeLanes.size && GlobalQueue.Kn(EFFECT_USER);
+            this.run(EFFECT_USER);
+            if (false) ;
+            if (false && !scheduled && !activeTransition && transitions.size === 0 && activeLanes.size === 0) ;
+            if (false) ;
+        } finally {
+            // Re-enter a woken transaction (see wokenTransitions) only from an
+            // idle pass: entering adopts the ambient batch, and staged or dirty
+            // ambient work would be held behind flights it never read. `scheduled`
+            // is that test here — after the park exit as well as the normal one:
+            // it was recomputed from the heap this pass, every write since re-armed
+            // it, and optimistic ambient nodes reverted with the finalize — so a
+            // wake in a pass with work simply falls to the next. Entering re-arms
+            // it itself; a dead (completed) wake is a bare return in
+            // initTransition, and the loop moves on to the next.
+            while (!scheduled && !activeTransition && wokenTransitions.length) this.initTransition(wokenTransitions.pop());
+            this.$t = false;
+        }
+    }
+    notify(e, t, n, i) {
+        // Only track async if the boundary is propagating STATUS_PENDING (not caught by boundary)
+        if (t & STATUS_PENDING) {
+            if (n & STATUS_PENDING) {
+                // Callers pass either nothing or this node's own `_x._error`, so `??`
+                // is exact (a null error falls back to the same null).
+                const t = i ?? e.o?._;
+                // A visibility-only mark notification (the affects() boundary
+                // channel) updates display state on its way up but must be invisible
+                // to completion accounting BY CONSTRUCTION: it never registers a
+                // reporter and never counts toward the loading-boundary diagnostic.
+                                if (t?.l) return true;
+                if (t) {
+                    // A reveal can discover a flight started in an earlier flush. Hold
+                    // the staged writes with that reader (A15), even if the reader is
+                    // new. Fresh/reset loading boundaries consume pending before it
+                    // reaches here. A reader already parked in a transition must not
+                    // open a second one.
+                    if (!activeTransition && !e.me && currentBatch.In.length) this.initTransition();
+                    if (activeTransition) {
+                        const n = t.source;
+                        let i = activeTransition.ae.get(n);
+                        if (!i) activeTransition.ae.set(n, i = new Set);
+                        const r = i.size;
+                        i.add(e);
+                        if (i.size !== r) {
+                            schedule();
+                            GlobalQueue.si?.(activeTransition);
+                        }
+                    }
+                }
+            }
+            return true;
+        }
+        return false;
+    }
+    initTransition(e) {
+        if (e) {
+            e = currentTransition(e);
+            // A finished transaction cannot be re-entered: its state is committed
+            // or reverted, so "rejoining" it (A26) is meaningless and re-activating
+            // it spins the drain loop (#3140). The refusal must be a bare return —
+            // redirecting the caller to a fresh batch would re-arm the loop with a
+            // new transaction identity each pass. Stamps are cleared at commit, so
+            // this is a belt for paths that hand over a chased-dead reference
+            // (merged chains, async settles racing completion).
+                        if (e.Dt === true || e === activeTransition) return;
+        }
+        if (!e && activeTransition && activeTransition.Ge === clock) return;
+        if (!activeTransition) {
+            activeTransition = e ?? createBatch();
+        } else if (e) {
+            const t = activeTransition;
+            mergeTransitionState(e, t);
+            // Effects the outgoing transaction parked belong to the surviving one
+            // now: back onto the live queue, where this flush parks them under
+            // `transition` or runs them at its completion. The outgoing stash is
+            // never read again — the transaction is dead (#3310).
+                        this.restoreQueues(t.ei);
+            transitions.delete(t);
+            activeTransition = e;
+        }
+        transitions.add(activeTransition);
+        activeTransition.Ge = clock;
+        const t = this.m;
+        if (t !== activeTransition) {
+            // Adopt the ambient batch into the transaction, then make the
+            // transaction the batch so later registrations land there directly.
+            // Pending and optimistic nodes are re-stamped as the transaction's;
+            // marks don't hijack the node's _transition — a mark on a plain signal
+            // must not entangle unrelated writes to it; the same rule holds one hop
+            // downstream: propagation never queues pended subscribers as pending
+            // nodes, see propagateAffectsMark, #2893.
+            // Adopted outside a flush: the staging is still unflushed — the stamp
+            // must not make it read as held-and-carried (CONFIG_ADOPTED_UNFLUSHED;
+            // the carrying flush clears it in reassignPendingTransition).
+            const e = this.$t ? 0 : CONFIG_ADOPTED_UNFLUSHED;
+            for (let n = 0; n < t.In.length; n++) {
+                const i = t.In[n];
+                // A tick that nets to the committed value proposed nothing (A34, #3494):
+                // `setShow(false); setShow(true)` beside a write that opens a hold
+                // left `show` staged at its own value, stamped, pending to the
+                // verdict, and its next mainline write held by a flight it never
+                // derived from. Unstage it here — its subscribers were walked at the
+                // write and re-derive the same value. Writes only: a signal's staging
+                // is always one, a computed's only under REACTIVE_MANUAL_WRITE
+                // (`createSignal(fn)`'s setter, #3519 review) — otherwise it is its
+                // pass's result, which may equal an uninitialized `undefined` (a
+                // born-held first pass). `_equals: false` opts out. The unstaging is
+                // the commit's own path (commitPendingNode with nothing staged): the
+                // manual-write flag, companions and the rest are cleaned up as a
+                // commit would, and the node is stamped nowhere. Unstamped
+                // only: a node already a transaction's — arriving here as a parked
+                // batch folds into a merge — carries a FLUSHED proposal a later
+                // rewrite brought back to the committed value; it is held, not
+                // proposal-free. Dropped, it kept the dead stamp, and the next write
+                // to it queued under the merged transaction a value the commit then
+                // skipped as another's (fuzzer latest-2 #1470, S3).
+                                if (i.me === null && i._e !== NOT_PENDING && (!i.Se || i.oe & REACTIVE_MANUAL_WRITE && !(i.S & STATUS_UNINITIALIZED)) && i.Fe && i.Fe(i.ce, i._e)) {
+                    i._e = NOT_PENDING;
+                    commitPendingNode(i);
+                    continue;
+                }
+                i.me = activeTransition;
+                i.C |= e;
+                activeTransition.In.push(i);
+            }
+            for (let e = 0; e < t.bn.length; e++) {
+                const n = t.bn[e];
+                n.me = activeTransition;
+                activeTransition.bn.push(n);
+            }
+            if (t.A.length) activeTransition.A.push(...t.A);
+            for (const e of t.En) activeTransition.En.add(e);
+            // Gated readers recorded against the ambient batch move with it: their
+            // replay-at-commit now happens at the transaction's completion.
+                        if (t.Tt.size) {
+                for (const e of t.Tt) activeTransition.Tt.add(e);
+                t.Tt.clear();
+            }
+            currentBatch = this.m = activeTransition;
+        }
+        for (const e of activeLanes) {
+            if (!e.me) e.me = activeTransition;
+        }
+        // A transaction's ambient window is one flush. Entering must therefore
+        // guarantee a flush: a transaction opened with no writes (an action whose
+        // first statements only await) otherwise leaves activeTransition and the
+        // adopted batch armed across the async gap, and the next unrelated work
+        // to arrive — an optimistic store's authoritative landing, a plain async
+        // settle — is adopted into a transaction it has nothing to do with
+        // (#3141). The scheduled flush parks the incomplete transaction through
+        // the normal machinery and detaches the ambient slots first.
+                schedule();
+    }
+}
+
+function queuePendingNode(e) {
+    currentBatch.In.push(e);
+    if (!globalQueue.$t) markUnflushedStaged();
+ // A28
+}
+
+// Sticky: flips true on the first refresh() ever (the only setter of
+// REACTIVE_REASK) so the hot notification loop skips the per-subscriber flag
+// clear entirely in apps that never refresh.
+let reaskArmed = false;
+
+/** §12d: bumped by every recompute and every new subscriber edge. A node's
+ * staged-rewrite skip is sound only while NOTHING recomputed or linked since
+ * its last notify — a mid-batch pull can clean a marked subscriber, and a
+ * skipped re-write would leave it stale. */ let notifyEpoch = 0;
+
+function bumpNotifyEpoch() {
+    notifyEpoch++;
+}
+
+function armReaskClear() {
+    reaskArmed = true;
+}
+
+/** Provenance of the work currently running (A18 supersession, #3331): the
+ * invocation sequence of the action whose ambient window this is — set by
+ * action() for each slice; the flush that ends the window clears it — or,
+ * inside an async landing, the sequence captured when that flight was
+ * registered (asyncWrite sets it for the landing's synchronous propagation,
+ * so a sync recompute downstream of the landing — an optimistic wrapper over
+ * the async source — derives under the flight's provenance, and flights it
+ * registers inherit it). 0 is mainline: no action, always the current
+ * question. An override stamps this at its write (`_overrideStamp`); an
+ * answer whose flight an OLDER action issued is a stale question the user
+ * has since changed — it holds silently to commit instead of superseding. A
+ * slow source must not leak back in over a newer intent. Transactions merge,
+ * so the transition object cannot say WHICH action asked; this can. */ let origin = 0;
+
+function setOrigin(e) {
+    const t = origin;
+    origin = e;
+    return t;
+}
+
+function insertSubs(e, t = false) {
+    // §12d: stamp before walking — setSignal's staged-rewrite fast path skips
+    // the next walk for this node while the epoch holds (marking is idempotent).
+    e.Lt = notifyEpoch;
+    // Get source lane: prefer node's own lane over current context
+    // This is important for isPending signals which need their own lane to flush immediately
+    // Presence bits gate the optional-slot probes (see constants.ts): one
+    // masked read of the always-present _config instead of missing-property
+    // lookups in the hottest notify loop. Bits are sticky — the field read
+    // stays authoritative when a bit is set.
+        const n = e.C;
+    const i = (n & CONFIG_HAS_LANE ? e.o?.Ce : undefined) || currentOptimisticLane;
+    const r = (n & CONFIG_HAS_SNAPSHOT) !== 0 && e.o?.ct !== undefined;
+    const s = reaskArmed;
+    // Observe-tier fan-out: this walk visits every subscriber edge anyway, so
+    // the graph-size count is one local increment here and no field anywhere.
+        let o = 0;
+    for (let n = e.u; n !== null; n = n.Pe) {
+        const e = n.Ae;
+        o++;
+        // A value-change notification is a new question for the subscriber: any
+        // pending re-ask mark (refresh) it carried is superseded.
+                if (s) e.oe &= ~REACTIVE_REASK;
+        // Missed-wake latch (#3037): this write is landing while the subscriber
+        // is mid-recompute (a nested pull committing beneath its reads), and the
+        // heap refuses RECOMPUTING nodes. A gen-current link means the pass
+        // already validated this dep — the value it read is now stale — so latch
+        // for recompute's tail to reschedule. Untouched links need no latch (the
+        // pass either re-reads them fresh or trims them), and neither does the
+        // tail link: it is the read IN FLIGHT — read() links before it pulls, so
+        // this very commit is what that read returns.
+                if (e.oe & REACTIVE_RECOMPUTING_DEPS && n.Ze === e.We && n !== e.Je) e.oe |= REACTIVE_MISSED_WAKE;
+        if (r && e.C & CONFIG_IN_SNAPSHOT_SCOPE) {
+            e.oe |= REACTIVE_SNAPSHOT_STALE;
+            continue;
+        }
+        if (t && i) {
+            e.oe |= REACTIVE_OPTIMISTIC_DIRTY;
+            assignOrMergeLane(e, i);
+        } else if (t) {
+            e.oe |= REACTIVE_OPTIMISTIC_DIRTY;
+            // No source lane means reversion - clear subscriber's lane so effects go to regular queue
+                        if (e.o) e.o.Ce = undefined;
+        }
+        enqueueSub(e);
+    }
+    if (o >= GRAPH_SIZE_WARN_AT) noteFanOut(e, o);
+}
+
+function commitPendingNode(e) {
+    const t = e;
+    if (!t.Se) {
+        if (e._e !== NOT_PENDING) {
+            e.ce = e._e;
+            e._e = NOT_PENDING;
+        }
+        if (e.C & CONFIG_HAS_COMPANIONS) GlobalQueue.Qn(e);
+        return;
+    }
+    if (e._e !== NOT_PENDING) {
+        e.ce = e._e;
+        e._e = NOT_PENDING;
+        // A node born held (recompute) initializes at this commit.
+                t.S &= ~STATUS_UNINITIALIZED;
+        // Set _modified for effects, but not for tracked effects (they handle their own scheduling)
+                if (e.He && e.He !== EFFECT_TRACKED) e.it = true;
+        // A quiet re-ask classification preserved through a held landing dies
+        // with the value commit — the commit IS the reveal (#3178). Gated on the
+        // staged value: status propagation queues pending nodes whose windows
+        // are still OPEN (no staged value), and their live classification must
+        // survive this sweep.
+                if (e.o) e.o.ve = false;
+    }
+    // The committed hold is the first observable answer for a loading-window
+    // node — the window closes here, not at compute time (#2990). Unconditional
+    // store to an always-present computed slot.
+        t.ye = false;
+    t.oe &= ~REACTIVE_MANUAL_WRITE;
+    // The children this commit publishes are the frame's now (#3404) — and so
+    // are the dependencies of the pass that produced the value: the previous
+    // frame's tail goes (A30, #3410; `recompute` left it for a staged pass). Only
+    // after a clean pass: `_error` is cleared by a clean pass or by the node's
+    // own landing (whose pass was clean), so a set `_error` means the last pass
+    // threw, kept its full list, and `_depsTail` marks where it stopped.
+        if (t.o?._ == null) trimStaleDeps(t);
+    // A LANE frame still parked here (#3662) rode a hold that stashed the run
+    // that would have retired it: this commit applies that run, so it goes too.
+        t.C &= -68157441;
+    if (!(t.S & STATUS_PENDING)) t.S &= ~STATUS_UNINITIALIZED;
+    // A flight this commit leaves in the air (unobserved, or observed only by
+    // a boundary) now has PUBLISHED inputs: its committed value is stale
+    // against the frame. read()'s reveal carve-out keys on the mark (#3305).
+     else e.C |= CONFIG_INPUTS_PUBLISHED;
+    if (t.o != null && (t.o.dt !== null || t.o.Et !== null)) GlobalQueue.ut(t, false, true);
+    if (e.C & CONFIG_HAS_COMPANIONS) GlobalQueue.Qn(e);
+}
+
+// Store commit hook (INTERNALS-STORE-STATE.md §3): installed by the store
+// module at init (same treeshakeable pattern as _resolveOptimistic /
+// _clearOptimisticStores). Folds committed store-node values into their
+// backing objects at the same moment pending values commit — the single
+// mutation point of the owned-raw model.
+let storeCommitHook = null;
+
+function setStoreCommitHook(e) {
+    storeCommitHook = e;
+}
+
+/** Held truth committed this finalize, awaiting its post-revert wake (see
+ * finalizePureQueue): the commit IS the reveal, but subscribers must not
+ * re-derive until the settling transaction's optimistic overrides have
+ * reverted — a commit-time wake recomputes them in the window where
+ * confirming truth is committed and the override still displays, a torn
+ * frame no timeline contains. */ const heldRevealed = [];
+
+/** Unchanged passes with a stale dependency tail, waiting on this flush's
+ * verdict (A30, #3469). A pass that changed nothing replaced nothing either —
+ * and cannot know at its own tail whether the flush that ran it will park:
+ * parked, its inputs are held and the committed frame still derives from the
+ * tail (`b() ? b() : a()` computed `1` from the held `b`, equal to the `1` it
+ * had from `a` — with `a` trimmed, the mainline `a = 2` never reached it).
+ * Trimmed when the flush commits; dropped with a park, the tail stays linked
+ * until a committing pass trims it (one spurious recompute at most). */ const heldTrims = [];
+
+function commitPendingNodes() {
+    while (heldTrims.length) trimStaleDeps(heldTrims.pop());
+    const e = currentBatch.In;
+    for (let t = 0; t < e.length; t++) {
+        const n = e[t];
+        commitPendingNode(n);
+        // The stamp dies with the commit (#3143) — symmetric with
+        // resolveOptimisticNodes clearing optimistic stamps. A stamp outliving
+        // its transaction let any later write (even a value-equal no-op, which
+        // re-opens before the equality bail) resurrect the finished transaction;
+        // a boundary flag rewritten every finalize pass then spun the drain loop
+        // forever (#3140). The held-truth mark dies the same death — the commit
+        // IS the reveal — but its wake defers to the post-revert pass: ordinary
+        // subscribers were masked to committed all hold (some re-derived against
+        // that old view and cached it), and commits are otherwise silent
+        // (staging already notified), so without a wake they'd hold the old
+        // world forever.
+                n.me = null;
+        if (n.C & CONFIG_HELD_TRUTH) {
+            n.C &= ~CONFIG_HELD_TRUTH;
+            heldRevealed.push(n);
+        }
+    }
+    e.length = 0;
+    storeCommitHook?.();
+}
+
+function finalizePureQueue(e = null, t = false) {
+    // For incomplete transitions, skip pending resolution and optimistic reversion
+    // For completing transitions or no-transition, resolve pending and revert optimistic
+    const n = currentBatch;
+    const i = !t;
+    if (i) commitPendingNodes();
+    // A parked finalize sweeps nothing: the boundaries' staged swaps are the
+    // transaction's, and a boundary whose own output parks the verdict was
+    // judged under it in run(), ahead of the verdict (#3540).
+        if (!t && globalQueue.Xn.length) checkBoundaryChildren(globalQueue);
+    // Contested effects (#3322) re-derive from the world this commit just
+    // produced. Ahead of the heap run — not the post-heap gated replay — so
+    // the recompute and the effect phase land in this same pass and the value
+    // the other transaction wrote into the slot is never published.
+    // (No clear: a completed transition is never finalized again.)
+    // A transaction whose settle reverts optimism re-derives them post-revert
+    // instead (below, with the gated replay): between commitPendingNodes and
+    // _resolveOptimistic the truth is committed but the overrides still
+    // display, and a re-derive here would compose the two — the #3164 tear,
+    // one window later. The slot meanwhile holds the frame that is on screen.
+        const r = e?.ht;
+    const s = i && (e ?? n).bn.length !== 0;
+    if (r && !s) for (const e of r) if (!(e.oe & REACTIVE_DISPOSED)) enqueueSub(e);
+    const o = dirtyQueue.EE >= dirtyQueue.st;
+    if (o) runHeap(dirtyQueue, GlobalQueue.nt);
+    if (i) {
+        // Boundary checks, commit hooks and recomputes can enter a transaction,
+        // which adopts the batch this finalize was settling: nothing batch-derived
+        // may be committed or reverted here — the entered transaction owns it now
+        // (#3319). A completing transaction's OWN containers are a different
+        // matter: when the ambient batch was separate from it (the #2916 shape),
+        // adoption never touched them and it must still settle them; when the
+        // batch WAS the completing transaction, adoption re-stamped its contents
+        // into the entered one and there is nothing left to settle.
+        if (currentBatch !== n) {
+            if (e === null || e === n) return;
+        } else if (o) commitPendingNodes();
+        // The settling batch: the completing transaction's, or the ambient one.
+                const t = e ?? n;
+        // Optimistic reversion: a non-empty batch means _optimisticWrite ran,
+        // which installed the engine's hooks.
+                if (t.bn.length) GlobalQueue.Bn(t.bn);
+        if (r && s) {
+            for (const e of r) if (!(e.oe & REACTIVE_DISPOSED)) enqueueSub(e);
+            schedule();
+        }
+        // Replay entanglement: subs recorded by the read-time gate get rescheduled
+        // so they re-run with the now-committed values visible. The ambient batch
+        // replays too — laneReadsCommitted records readers whose committed-view
+        // read hid a same-tick plain write that just committed above (#2963).
+                if (t.Tt.size) {
+            for (const e of t.Tt) {
+                if (e.oe & REACTIVE_DISPOSED) continue;
+                enqueueSub(e);
+            }
+            t.Tt.clear();
+            // A completing transition keeps the outer flush loop alive by itself;
+            // the ambient batch needs the re-arm or the replay sits in the heap
+            // until the next unrelated write.
+                        schedule();
+        }
+        // Declared motion ends with the transaction: settle (or plain flush end
+        // for ambient marks) releases each registration's refcount. A non-empty
+        // batch means registerAffectsMark ran, which installed the hook. Marks
+        // held boundary display state through the visual channel, and their
+        // release is the display-state update point — re-run the boundary sweep
+        // (the earlier sweep above ran while the marks were still live).
+                if (t.A.length) {
+            GlobalQueue.G(t.A);
+            if (globalQueue.Xn.length) checkBoundaryChildren(globalQueue);
+        }
+        // A non-empty set means trackOptimisticStore ran, which installed the
+        // hook; the hook iterates, clears, and schedules (keeping the loop out of
+        // core lets esbuild shake it — rollup already folds the null guard). The
+        // completing transition scopes the clear to its own layer keys (#2899).
+                if (t.En.size) GlobalQueue.ii(t.En, e);
+        // Held-truth reveal wake (#3164), post-revert by construction: this
+        // finalize committed confirming truth whose subscribers were masked all
+        // hold — some re-derived against the committed view (the staging, or a
+        // confirming carrier's landing, notified them as a plain write) and
+        // cached it, and stash-restored applies may carry those torn values.
+        // Waking and recomputing HERE — after _resolveOptimistic and the store
+        // clears above — means every apply paints the settled view; a wake at
+        // commit time would recompute them in the window where truth is
+        // committed but the settling transaction's overrides still display.
+                if (heldRevealed.length !== 0) {
+            while (heldRevealed.length) insertSubs(heldRevealed.pop());
+            if (dirtyQueue.EE >= dirtyQueue.st) {
+                runHeap(dirtyQueue, GlobalQueue.nt);
+                commitPendingNodes();
+            }
+        }
+        sweepTransientStoreNodes();
+        // Lanes only enter activeLanes through the engine's getOrCreateLane.
+                if (activeLanes.size) GlobalQueue.Wn(e);
+    }
+}
+
+/** The boundary sweep: the commit's (`_checkSources`), or — `held`, before
+ * the verdict (#3540) — the one for a collecting boundary whose output is
+ * pending on its fallback's read (`_judgeHeld`). */ function checkBoundaryChildren(e, t) {
+    for (const n of e.Xn) {
+        t ? n.Re?.() : n.Te?.();
+        checkBoundaryChildren(n, t);
+    }
+}
+
+/**
+ * Count of live `affects()` registrations across the system (including
+ * store-scope inherited marks). Gates the read-path mark check in `read()` so
+ * graphs that never use the feature pay one integer compare.
+ */ let activeAffectsMarks = 0;
+
+/**
+ * Counter mutation seam for the mark engine in affects.ts: an imported `let`
+ * binding is read-only, and the read-path gate above must stay a plain module
+ * variable so `read()` pays one integer compare, not a function call.
+ *
+ * @internal
+ */ function shiftAffectsMarks(e) {
+    activeAffectsMarks += e;
+}
+
+function reassignPendingTransition(e) {
+    for (let t = 0; t < e.length; t++) {
+        e[t].me = activeTransition;
+        e[t].C &= ~CONFIG_ADOPTED_UNFLUSHED;
+ // this flush carried it
+        }
+}
+
+const globalQueue = new GlobalQueue;
+
+// Hot-path mirror of `globalQueue._batch`: `queuePendingNode` runs once per
+// staged write and `commitPendingNodes` once per flush, and the extra
+// property hop through `_batch` was a measured instruction-count regression
+// (CodSpeed update1to1, PR #2905). The field stays authoritative for
+// cross-module readers; every `_batch` assignment updates both.
+let currentBatch = globalQueue.m;
+
+function flush(e) {
+    // Inside an action body the drain is incoherent (#3333): the action's
+    // writes are held by its transaction until it settles, so a drain can't
+    // reveal them — and the loop below only exits once `activeTransition` is
+    // null, so it would PARK the transaction mid-slice and every write after it
+    // in the body would land as a plain, committed write. The reporter's
+    // "leading flush()" workaround was exactly that leak. Prod: run `fn` if
+    // given (its writes stay in the transaction) and skip the drain.
+    if (actionStepDepth > 0) {
+        return e ? e() : undefined;
+    }
+    if (e) {
+        syncDepth++;
+        try {
+            return e();
+        } finally {
+            // Decrement even if the drain throws (a throwing effect): a leaked
+            // syncDepth would stop `schedule()` from ever queuing a microtask again.
+            try {
+                flush();
+            } finally {
+                syncDepth--;
+            }
+        }
+    }
+    if (globalQueue.$t) {
+        return;
+    }
+    if (halted) return;
+    // Attribution: whether this call drained anything, so `flushEnd` fires once
+    // per real drain and never for a no-op call. The declaration is dead in
+    // prod (its only write is behind true) and rollup drops it.
+        let t = false;
+    // The drain's opening instant, under the loop's own condition so it fires
+    // exactly when `flushEnd` below will. Outside every try (see the rule in
+    // attribution-hooks.ts).
+        if (attrHooks !== null && (scheduled || activeTransition)) attrHooks.flushStart();
+    // `flush()` is an explicit drain point, so it must also process an active
+    // transition even if no microtask was scheduled for it yet.
+        while (scheduled || activeTransition) {
+        globalQueue.flush();
+        t = true;
+    }
+    // Provenance ends with the drain: every ambient window (an action's first
+    // slice, a landing's propagation) runs to this flush.
+        origin = 0;
+    // Outside every try in this function (see the rule in attribution-hooks.ts):
+    // the drain loop above is the one place all scheduled work funnels through,
+    // so this is the "committed and effects ran, or parked" instant for
+    // everything the loop processed.
+        if (t && attrHooks !== null) attrHooks.flushEnd();
+}
+
+function runQueue(e, t) {
+    for (let n = 0; n < e.length; n++) e[n](t);
+}
+
+/** Does `reporter` still hold the transaction waiting on `source` — live,
+ * routed to no collecting loading boundary, and deriving from the source in
+ * its current pass? The verdict's per-reporter test (sourceObserved), also
+ * a boundary re-arm's (boundaries.ts `_rearm` runs before the verdict prunes
+ * registrations that stopped counting). */ function reporterBlocksSource(e, t, n) {
+    const i = e.oe;
+    if (i & REACTIVE_DISPOSED) return false;
+    // A zombie renders until the commit that disposes it (#3463): while its
+    // removal is staged in a live transaction it is still on screen, and what
+    // it displays must stay consistent with the frame — a held `Show`'s
+    // `Details: 0` beside the lane's `Value: 1` otherwise. Its say is moot for
+    // the verdict of the transaction that stages the removal (`verdict`): done,
+    // and the commit disposes it; not done, and it stays parked regardless. A
+    // zombie whose removal commits this flush (owner's pass not held) is dead.
+    // The owner is stamped when the flush parks; held in this flush, its
+    // staging transaction is the active one.
+        if (i & REACTIVE_ZOMBIE) {
+        let t = e;
+        while (t && t.oe & REACTIVE_ZOMBIE) t = t._parent;
+        let i = t && (t.me || (t.C & CONFIG_HELD_CHILDREN ? activeTransition : null));
+        // A LANE frame's member (#3662) is displayed until its owner's run
+        // applies, and the lane's transaction is what applies it (its completion
+        // runs the lane's queue): moot for that verdict, live for every other.
+                if (!i && t && t.C & CONFIG_LANE_FRAME && t.o?.Ce) i = findLane(t.o.Ce).me;
+        if (!i || (i = currentTransition(i)).Dt === true || i === n) return false;
+    }
+    // Fallback-caught async holds nothing. A collecting loading boundary
+    // consumes the notification, so a reader under a fallback never registers —
+    // but a reader registered while its boundary showed content stays
+    // registered when the boundary's `on` later changes and it flips to the
+    // fallback. The reader is behind the fallback now; if nothing outside the
+    // boundary consumes the flight, the hold is over (A33, ruled 2026-09-12, #3375).
+        for (let t = e.T; t; t = t._parent) if (t.te & STATUS_PENDING && !t.H) return false;
+    // "Still derives from the source" is a question about THIS pass's reads:
+    // the deps up to `_depsTail`. Past it lie the committed frame's — kept
+    // linked by A30 until the commit trims them (a staged pass, an errored
+    // one). Reading them here made a reporter whose pass had stopped reading
+    // the source (a gate closed in the same flush as the write) look live, and
+    // the hold it kept was the commit that would have trimmed the dep that
+    // kept it (spec O3, same-flush form; fuzzer #3446 P1 cases 21/79). A
+    // trimmed list ends at `_depsTail`, so the bound is free there; a pass
+    // that read nothing has a null tail and derives from nothing.
+    // The registration is trusted: a pending mark rides only this pass's links
+    // (notifyStatus skips the kept tail), so a registered reporter read the
+    // source, or threw on it.
+        if (e.o?.ue?.has(t)) return true;
+    const r = e.Je;
+    for (let n = r === null ? null : e.Ee; n; n = n === r ? null : n.Ie) {
+        let e = n.De;
+        while (e) {
+            // Or through a memo pending on the flight (#3494): a stale reader
+            // served a held memo's committed value never turned pending itself, so
+            // its only trace of the flight is the memo between them — `copy` of
+            // `details`. Judged dead, its transaction released `count=1` beside
+            // the `Copy: 0` it displays. `_pendingSources` is transitive, so one
+            // hop covers any depth.
+            if (e === t || e.Ne === t || e.o?.ue?.has(t)) return true;
+            e = e.o?.$n;
+        }
+    }
+    return !!(e.S & STATUS_PENDING && e.o?._ instanceof NotReadyError && e.o?._.source === t);
+}
+
+/**
+ * Does a live reporter of `transition` still observe `source` pending? Dead
+ * reporters (disposed, behind a fallback, no longer reading the source) are
+ * pruned as they are found, and the source's entry with them. Shared by the
+ * settle verdict and the lane's hold check (`waitingTransition`): a live
+ * action parks its transaction without a verdict, so this prune is the only
+ * one an optimistic lane whose last async reader unmounted mid-action ever
+ * gets — without it the lane held on the dead reporter's registration until
+ * the flight it no longer observed landed (#3426).
+ */ function sourceObserved(e, t, n) {
+    const i = e.ae.get(t);
+    let r = false;
+    for (const e of i ?? []) {
+        if (reporterBlocksSource(e, t, n)) return true;
+        // A zombie the verdict passes over is kept, not pruned (#3463): moot for
+        // this verdict, it still holds a lane's reveal while the transaction
+        // stays parked on something else.
+                if (n && e.oe & REACTIVE_ZOMBIE) r = true; else i.delete(e);
+    }
+    if (!r) e.ae.delete(t);
+    return false;
+}
+
+function transitionComplete(e) {
+    if (e.Dt) return true;
+    if (e.tt.length) {
+        // A live action parks the transaction regardless of async state.
+        if (attrHooks !== null) attrHooks.holdStart(e);
+        return false;
+    }
+    let t = true;
+    for (const n of e.ae.keys()) {
+        // The source blocks while its OWN flight is up — the self entry in its
+        // pending sources (added by notifyStatus's source path, with status;
+        // retired by the landing and the supersede sweep, so it implies
+        // STATUS_PENDING). `_error.source` is not that test: propagation from an
+        // input that went pending later overwrites it with the input (#3375 — a
+        // boundary-consumed load re-asked under a held derivation), and the
+        // still-flying source read as settled, committing the writes it was
+        // asked with ahead of its answer.
+        // Not the self entry alone: a source whose own flight an upstream re-ask
+        // superseded is still pending — on that re-ask (#3462). Its reader cannot
+        // render until the chain lands, and the landing folds this transaction in
+        // (enterWaiting). Judged complete instead, a re-entry between the two (a
+        // repeated write to a held signal) committed the held writes beside the
+        // reader's stale frame.
+        if (sourceObserved(e, n, e) && n.o?.ue?.size) {
+            t = false;
+            break;
+        }
+    }
+    // Override blockage lives with the engine (absent hook = "no optimistic
+    // blockage"); the hook's loops over _optimisticNodes/_optimisticStores are
+    // no-ops when the transition holds neither, so no pre-check is needed.
+        if (t && GlobalQueue.Yn?.(e)) t = false;
+    // Attribution hook: this verdict is the fork between settling (held writes
+    // commit next — `_pendingNodes` still lists them) and parking (the flush
+    // runs the lane effects, then stashes; `holdEnd` fires from stashQueues).
+    // Fired here rather than at flush()'s call site because that site is inside
+    // a `try` (see the rule in attribution-hooks.ts).
+        if (attrHooks !== null) t ? attrHooks.transitionSettled(e) : attrHooks.holdStart(e);
+    t && (e.Dt = true);
+    return t;
+}
+
+/** A fresh, unentered transaction (#3146): the optimistic store's truth
+ * flight DECLARES an owned transaction instead of relying on whatever the
+ * ambient adoption machinery stamped on its firewall. Activate it with
+ * initTransition; it is a plain batch until then. */ function createTransition() {
+    return createBatch();
+}
+
+function currentTransition(e) {
+    while (e.Dt && typeof e.Dt === "object") e = e.Dt;
+    return e;
+}
+
+/**
+ * The live transition blocked on `source` — the one whose render reader
+ * observed it pending (INV-3 records the observation in whichever transaction
+ * was active when the reader was notified). The observation is a fact about
+ * the node, so a hold check must not assume it was recorded in the transaction
+ * it happens to hold — lanes merge across transactions (#2912), and a merged
+ * root's transaction knows nothing of the async its members' transactions
+ * observed (#3335). Null when nobody is waiting — a registration whose every
+ * reporter has since died is nobody (#3426).
+ */ function waitingTransition(e) {
+    for (const t of transitions) if (sourceObserved(t, e)) return t;
+    return null;
+}
+
+/** A landing enters EVERY parked transaction still waiting on `source`, folding
+ * them into the active one (A15: each reveal that discovered the flight
+ * completes at its landing). The fold used to happen as the waiters' stamped
+ * readers recomputed under the landing — recompute re-entering an effect's
+ * stamp — which also folded in writes those readers merely shared a hole
+ * with (#3407); effects no longer re-enter, so the landing folds explicitly.
+ * Live iteration is safe: a merge deletes the outgoing (active) entry and
+ * re-adds the visited one. */ function enterWaiting(e) {
+    for (const t of transitions) if (sourceObserved(t, e)) globalQueue.initTransition(t);
+}
+
+function runInTransition(e, t) {
+    const n = activeTransition;
+    try {
+        activeTransition = currentTransition(e);
+        return t();
+    } finally {
+        activeTransition = n;
+    }
+}
+
+/** Run `fn` with `transition` as BOTH the ambient transaction and the
+ * registration batch, restoring both after. runInTransition alone is not
+ * enough for code that WRITES on behalf of a transaction from inside someone
+ * else's window (optimistic replay re-arming a still-open action's edits
+ * during a landing commit, #3123): registrations route through the queue's
+ * batch pointer, and a bare activeTransition swap leaves them in the ambient
+ * batch — a plain batch "completes" at the next flush and reverts optimistic
+ * registrations that were supposed to live with the transaction.
+ * initTransition is the wrong tool here: it MERGES the currently ambient
+ * transaction into the target, entangling whatever the interrupted window
+ * belonged to. */ function runAsTransitionBatch(e, t) {
+    const n = activeTransition;
+    const i = globalQueue.m;
+    try {
+        activeTransition = currentTransition(e);
+        currentBatch = globalQueue.m = activeTransition;
+        return t();
+    } finally {
+        activeTransition = n;
+        currentBatch = globalQueue.m = i;
+    }
+}
+
+export { GlobalQueue, Queue, ROOT_ERROR_HOOK, actionStepDepth, activeAffectsMarks, activeLanes, activeTransition, armReaskClear, assignOrMergeLane, batchJoins, bumpNotifyEpoch, clock, createTransition, currentTransition, deferSlotRelease, dirtyQueue, enforceLoadingBoundary, entangleConfirmingTransitions, enterActionStep, enterWaiting, exitActionStep, finalizePureQueue, findLane, flush, globalQueue, haltReactivity, hasActiveOverride, heldTrims, insertSubs, notifyEpoch, origin, pendingRearms, projectionWriteActive, queuePendingNode, queueRearm, reaskArmed, reporterBlocksSource, resetErrorHalt, runAsTransitionBatch, runInTransition, schedule, scheduleWithheld, setOrigin, setProjectionWriteActive, setStoreCommitHook, shiftAffectsMarks, sourceObserved, storeCommitHook, transitions, waitingTransition, wakeParked, wokenTransitions, zombieQueue };