@solidjs/signals 2.0.0-rc.1 → 2.0.0-rc.10

package/dist/observe/core/core.js

@@ -0,0 +1,1999 @@
+import { releaseFlightTeardown, handleAsync, clearStatus, parkLoadingWindow, notifyStatus, settlePendingSource, settleErroredDependents } from "./async.js";
+
+import { NOT_PENDING, EFFECT_TRACKED, EFFECT_USER, REACTIVE_OPTIMISTIC_DIRTY, CONFIG_DERIVED_OVERRIDE, CONFIG_HELD_CHILDREN, CONFIG_LANE_FRAME, CONFIG_OPTIMISTIC, STATUS_UNINITIALIZED, STATUS_ERROR, STATUS_PENDING, REACTIVE_REASK, REACTIVE_RECOMPUTING_DEPS, REACTIVE_ZOMBIE, CONFIG_SYNC, CONFIG_HAS_LANE, REACTIVE_MISSED_WAKE, REACTIVE_DISPOSED, REACTIVE_SNAPSHOT_STALE, unwrapOverride, CONFIG_DIRECT_COMMIT, CONFIG_HAS_COMPANIONS, CONFIG_PROMOTED, CONFIG_ADOPTED_UNFLUSHED, CONFIG_CHILDREN_FORBIDDEN, CONFIG_FRESH_READ, CONFIG_INPUTS_PUBLISHED, CONFIG_AUTO_DISPOSE, REACTIVE_LAZY, REACTIVE_CHECK, REACTIVE_DIRTY, REACTIVE_IN_HEAP, REACTIVE_IN_HEAP_HEIGHT, REACTIVE_MANUAL_WRITE, CONFIG_AUTHORITATIVE_READ, CONFIG_OVERRIDE_SUPERSEDED, CONFIG_AUTHORITATIVE_OBSERVED, defaultContext, REACTIVE_NONE, CONFIG_HELD_TRUTH, CONFIG_TRANSPARENT, CONFIG_OWNED_WRITE, CONFIG_NO_SNAPSHOT, CONFIG_PLUMBING, CONFIG_IN_SNAPSHOT_SCOPE, NO_SNAPSHOT, CONFIG_HAS_SNAPSHOT, CONFIG_FW_CHILDREN, CONFIG_SLOT_NODE, STORE_SNAPSHOT_PROPS } from "./constants.js";
+
+import { NotReadyError } from "./error.js";
+
+import { clearDeps, trimStaleDeps, link, dormantNodes } from "./graph.js";
+
+import { deleteFromHeap, queueFor, insertIntoHeapHeight, enqueueSub, markNode, markHeap, insertIntoHeap } from "./heap.js";
+
+import { findLane } from "./lanes.js";
+
+import { GRAPH_SIZE_WARN_AT, noteFanIn } from "./dev.js";
+
+import { attrHooks } from "./attribution-hooks.js";
+
+import "./invariants.js";
+
+import { disposeChildren, markDisposal, linkChild, inheritId } from "./owner.js";
+
+import { GlobalQueue, bumpNotifyEpoch, activeTransition, globalQueue, clock, currentTransition, insertSubs, queuePendingNode, wakeParked, heldTrims, runInTransition, schedule, batchJoins, notifyEpoch, dirtyQueue, projectionWriteActive, reaskArmed, armReaskClear } from "./scheduler.js";
+
+// The heap's per-node step. A tracked effect's heap visit is its compute
+// phase — empty, like a user effect whose compute reads nothing — and hands
+// the callback to the user queue. Routing the wake through the heap, rather
+// than straight into the queue at notify time, is what orders the run after
+// the commit regardless of which phase the write came from: a write in a
+// render-effect callback stages its value for the next pass, but a wake pushed
+// directly into the user queue ran in the SAME pass, read the old value, and
+// nothing re-notified it when the value landed (#3291).
+GlobalQueue.nt = e => {
+    if (e.He === EFFECT_TRACKED) {
+        deleteFromHeap(e, queueFor(e));
+        e.it = true;
+        e.T.enqueue(EFFECT_USER, e.lt);
+    } else recompute(e);
+};
+
+GlobalQueue.ut = disposeChildren;
+
+let tracking = false;
+
+/** @internal verdict-module glue */ function setPendingCheckActive(e) {
+    pendingCheckActive = e;
+}
+
+/** @internal verdict-module glue */ function setLatestReadActive(e) {
+    latestReadActive = e;
+}
+
+/** @internal verdict-module glue */ function setContextInternal(e) {
+    context = e;
+}
+
+let stale = false;
+
+let pendingCheckActive = false;
+
+let latestReadActive = false;
+
+let context = null;
+
+let currentOptimisticLane = null;
+
+/** Notify `node`'s subscribers on `lane`'s channel: they recompute as the
+ * lane's passes — a memo publishes a derived override (lanes stage, #3479),
+ * an effect runs from the lane's queue, at the park, ahead of the
+ * transaction — the display-ahead view. The write itself is already staged
+ * (setSignal) and commits with the frame; this walk shows it now. Used by a
+ * boundary re-armed from a lane pass (boundaries.ts `_swap`, #3540). */ function notifyOnLane(e, t) {
+    const n = currentOptimisticLane;
+    currentOptimisticLane = t;
+    try {
+        insertSubs(e, true);
+    } finally {
+        currentOptimisticLane = n;
+    }
+}
+
+let snapshotCaptureActive = false;
+
+let snapshotSources = null;
+
+function ownerInSnapshotScope(e) {
+    while (e) {
+        if (e.ot) return true;
+        e = e._parent;
+    }
+    return false;
+}
+
+function setSnapshotCapture(e) {
+    snapshotCaptureActive = e;
+    if (e && !snapshotSources) snapshotSources = new Set;
+}
+
+function markSnapshotScope(e) {
+    e.ot = true;
+}
+
+function releaseSnapshotScope(e) {
+    e.ot = false;
+    releaseSubtree(e);
+    schedule();
+}
+
+function releaseSubtree(e) {
+    let t = e._n;
+    while (t) {
+        if (t.ot) {
+            t = t.Fn;
+            continue;
+        }
+        if (t.Se) {
+            const e = t;
+            e.C &= ~CONFIG_IN_SNAPSHOT_SCOPE;
+            if (e.oe & REACTIVE_SNAPSHOT_STALE) {
+                e.oe &= ~REACTIVE_SNAPSHOT_STALE;
+                e.oe |= REACTIVE_DIRTY;
+                if (dirtyQueue.st > e.ft) dirtyQueue.st = e.ft;
+                insertIntoHeap(e, dirtyQueue);
+            }
+        }
+        releaseSubtree(t);
+        t = t.Fn;
+    }
+}
+
+function clearSnapshots() {
+    if (snapshotSources) {
+        for (const e of snapshotSources) {
+            // The extension is a fixed-shape object with `_snapshotValue`
+            // pre-initialized to undefined (see ext()), and every reader tests
+            // `!== undefined` — assign, don't `delete`: deleting a field pushes the
+            // object to dictionary mode for every later read of every field.
+            const t = e.o;
+            if (t != null) t.ct = undefined;
+            // StoreNode targets share one pre-initialized hidden class (see
+            // createStoreProxy) — same rule, and only when present so signal-node
+            // sources don't grow the field.
+                        if (e[STORE_SNAPSHOT_PROPS] !== undefined) e[STORE_SNAPSHOT_PROPS] = undefined;
+        }
+        snapshotSources = null;
+    }
+    snapshotCaptureActive = false;
+}
+
+function recompute(e, t = false) {
+    // §12d: any recompute can clean a marked subscriber — invalidate skips.
+    bumpNotifyEpoch();
+    const n = e.He;
+    // Attribution hook: fired before this run touches the dep list — `_deps`
+    // still holds the previous run's links (the subscriptions that could have
+    // triggered this run, and the baseline for the engine's subscription diff).
+        let i = false;
+    if (attrHooks !== null) attrHooks.recomputeStart(e, t);
+    // Lane posture is resolved BEFORE the previous frame is parked below: a
+    // lane pass on an effect direct-commits (#3662, see the parking site), so
+    // the decision must be known there. `lane` is applied to
+    // `currentOptimisticLane` further down, once the previous posture is saved.
+        let l = !!(e.oe & REACTIVE_OPTIMISTIC_DIRTY);
+    let u = null;
+    if (l) {
+        u = GlobalQueue._t(e, true);
+        // `false` = wake-only lane demotion: recompute plain so a mid-tick
+        // latest()/isPending() pull stages instead of direct-committing (#3009).
+        // The predicate lives with the engine (recomputeLane).
+                if (u === false) l = false;
+    } else if (e.C & CONFIG_DERIVED_OVERRIDE) {
+        // Lanes stage (#3479): a pass over a live lane member carrying a derived
+        // override is the lane's pass whatever channel dirtied it (a boundary
+        // reset, an unrelated sync write) — its inputs serve the lane's view, so
+        // its result is the lane's and belongs in the override slot. Run plain,
+        // A18's sync twin below read that re-derived lane view as a differing
+        // truth (a fresh array), superseded the override and demoted the lane;
+        // the lane's next pass then dropped the staged "truth" and left the node
+        // flagged superseded with nothing to serve (fuzzer latest-1 #2481). A
+        // demoted node resolves no lane and stays plain: its pass IS the truth.
+        u = GlobalQueue._t(e, true);
+        if (u) l = true;
+    } else if (activeTransition && !t && activeTransition.bn.length) {
+        // Lane adoption: parent-deeper-than-owned-child can run before its OPT-dirty
+        // child propagates. Walk deps once and inherit the OPT lane so this node
+        // recomputes under the right posture and propagates correctly.
+        u = GlobalQueue._t(e, false);
+        if (u) l = true;
+    }
+    if (!t) {
+        // A stamped memo re-enters its hold: its value is that transaction's work.
+        // An effect's pass belongs to whatever dirtied it (A15 corollary: effects
+        // don't entangle parallel transactions); it joins its stamp only when the
+        // pass observes the held flight — queue notification (#3407).
+        if (e.me && !n && activeTransition !== e.me) globalQueue.initTransition(e.me);
+        deleteFromHeap(e, queueFor(e));
+        if (e.o !== null) {
+            e.o.pe = null;
+            // Supersede is where an iterator flight dies (#3122): close it now.
+            // Its cleanup(close) registration may sit in a zombie-deferred
+            // disposal list that a held transition only drains when the
+            // SUPERSEDING flight settles — cancellation must not wait for the
+            // work that replaced it. Idempotent with the cleanup-channel close.
+                        releaseFlightTeardown(e);
+        }
+        // Tracked effects run after finalizePureQueue, so dispose immediately
+        // instead of deferring. Children built by an uncommitted recompute
+        // (CONFIG_HELD_CHILDREN) die immediately too: no frame ever showed them.
+        // Everything else is the committed frame's and is deferred as zombies
+        // until this node's commit — a transaction-owned node included (#3404):
+        // a parked node's children predate the hold, and tearing them down when
+        // the source lands ran cleanups before the transaction's atomic reveal.
+        // A lane pass on an effect parks a LANE frame instead (CONFIG_LANE_FRAME,
+        // #3662; A15 lanes corollary): the frame it replaces leaves the screen
+        // when the lane's queue applies this run (A30) — not at the action's
+        // commit — and a held lane defers that with the frame still displayed.
+        // The drain is the lane's first render entry for this pass, pushed before
+        // the pass builds the new frame: cleanups before side effects — the
+        // retired frame's `onCleanup`s run ahead of every effect callback of the
+        // new frame. While the frame waits, the live children were never shown:
+        // a superseding pass disposes them here like held children, and the
+        // parked frame stays.
+                if (n === EFFECT_TRACKED || e.C & (CONFIG_HELD_CHILDREN | CONFIG_LANE_FRAME)) disposeChildren(e); else if (e._n !== null || e.Ve !== null) {
+            markDisposal(e);
+            const t = ext(e);
+            t.Et = e.Ve;
+            t.dt = e._n;
+            e.Ve = null;
+            e._n = null;
+            e.Nt = 0;
+            if (n && u) {
+                e.C |= CONFIG_LANE_FRAME;
+                findLane(u).It[0].push(() => {
+                    e.C &= ~CONFIG_LANE_FRAME;
+                    disposeChildren(e, false, true);
+                });
+            }
+        } else ;
+    }
+    // A derived override (lanes stage, #3479) is override-covered like a written
+    // one: a plain pass over it — its source superseded (A18) — stages the truth
+    // and supersedes the override through the sync twin below.
+        const o = (e.C & (CONFIG_OPTIMISTIC | CONFIG_DERIVED_OVERRIDE)) !== 0 && e.o?.we !== NOT_PENDING && e.o?.we !== undefined;
+    const r = !!(e.S & STATUS_UNINITIALIZED);
+    // Capture both error and pending status before the compute clears them.
+    // A conditional can drop its pending source and recover to an unchanged
+    // value, leaving blocked dependents outside that source’s settle walk.
+        const s = e.S & STATUS_ERROR ? e.o?._ : undefined;
+    const a = (e.S & STATUS_PENDING) !== 0;
+    const f = a ? e.o?.ue : undefined;
+    // Pending SOURCE-hood, captured before the compute clears status: a node
+    // whose own flight parked dependents self-registers in _pendingSources
+    // (notifyStatus, isSource). If this recompute supersedes that flight and
+    // settles synchronously, those dependents settle HERE — asyncWrite's
+    // settlePendingSource walk never runs for a landing that was preempted
+    // (#3181).
+        const c = e.o?.ue?.has(e);
+    // Re-ask classification lives in the verdict module; capture the flag before
+    // the recompute wipes _flags below.
+        const _ = (e.oe & REACTIVE_REASK) !== 0;
+    // Captured before the compute clears it on a sync landing: if that landing
+    // is transition-held below, the window must stay open until the hold
+    // commits (commitPendingNode) — a closed window plus a held value reads as
+    // a pending frame to live observers of the verdict (#2990).
+        const E = e.ye;
+    // Creation-time A29 (see enterStagedRead): a pass outside a flush that is
+    // served a live transaction's staged value records the transaction here
+    // and is staged INTO it below — "born held" — instead of committing.
+        const d = stagedEntry;
+    stagedEntry = null;
+    const N = context;
+    context = e;
+    e.Je = null;
+    e.We++;
+    // REACTIVE_ZOMBIE is position, not scheduling state: it says the node sits
+    // on its owner's `_pendingFirstChild` chain, and `disposeChildren` keys its
+    // parent-chain splice off it. A zombie reruns for mainline writes until the
+    // commit that disposes it (#3463), so the per-pass wipe here — and in the
+    // finally below and in updateIfNecessary — must carry it (#3543): a
+    // de-flagged zombie spliced itself out of the LIVE chain at disposal,
+    // orphaning the owner's current child, which stayed subscribed forever.
+        e.oe = REACTIVE_RECOMPUTING_DEPS | e.oe & REACTIVE_ZOMBIE;
+    e.Ge = clock;
+    let I = e._e === NOT_PENDING ? e.ce : e._e;
+    let T = e.ft;
+    let S = false;
+    let O = tracking;
+    let A = currentOptimisticLane;
+    tracking = true;
+    // A computed's fn establishes its OWN dependencies, so it must never run
+    // inside a latest() read window: read() short-circuits through the
+    // companion path before dependency linking, so a memo created (eagerly
+    // computed) inside latest(fn) came out permanently dependency-less (#2926).
+    // latestRead() already suspends the flag for its pull-recomputes; this
+    // covers creation-time computes and flushes that run inside the window.
+        const C = latestReadActive;
+    latestReadActive = false;
+    // A memo computes under its OWN lane posture, never the puller's (A31,
+    // #3442): its value is one shared slot every reader sees, so a pull from a
+    // lane-carrying reader (a probe effect on its companion lane pulling a sync
+    // memo) must not run it with that lane's read carve-outs — under a lane, a
+    // pending node on no lane serves its committed value instead of throwing,
+    // and the memo then published a stale "settled" value, dropped its
+    // pending status, and stopped holding its transaction. The branches below
+    // re-establish the posture the memo itself owns (OPT-dirty, or adopted
+    // through its deps). Effects keep the ambient lane: their runs are the
+    // lane's own view.
+        if (!n) currentOptimisticLane = null;
+    // Lane posture lives with the engine: OPTIMISTIC_DIRTY is only ever set by
+    // engine-driven paths, and _optimisticNodes is only pushed by
+    // _optimisticWrite, so the hook is installed whenever either gate holds.
+    // (Resolved at the top of this pass, ahead of the parking site.)
+        if (u) currentOptimisticLane = u;
+    const p = n && n !== EFFECT_USER;
+    const R = stale;
+    if (p) stale = true;
+    // An effect recorded for this transaction's commit replay (it once read a
+    // node the transaction held and showed the committed value) and now
+    // recomputing UNDER the transaction sees its staged view: the value this
+    // run produces is applied by the commit itself, and the stale recording
+    // would publish the frame a second time. Drop it; the reads below re-record
+    // if they are served the committed view again (a lane's committed read).
+        if (n && activeTransition !== null && activeTransition.Tt.size) activeTransition.Tt.delete(e);
+    try {
+        if (!false && e.C & CONFIG_SYNC) {
+            I = e.Se(I);
+            if (e.o !== null) e.o.pe = null;
+            e.ye = false;
+        } else {
+            // Snapshot `_inFlight` so we can detect whether `_fn` self-registered an async
+            // subscription (e.g. `createProjection` calls `handleAsync` from inside its body
+            // with a setter callback). In that case, the outer `handleAsync` call below would
+            // clobber the fresh subscription, so we skip it and let the internally-registered
+            // iteration drive updates.
+            const t = e.o?.pe;
+            const n = e.Se(I);
+            const i = typeof n === "object" && n !== null;
+            const l = e.o?.pe !== t;
+            I = l || !i ? n : handleAsync(e, n);
+            if (!l && !i) {
+                if (e.o !== null) e.o.pe = null;
+                // A sync (non-object) return is the first real answer; async-shaped
+                // results clear inside handleAsync at their own landing points, and a
+                // self-registered flight (inFlightChanged — projections) clears when
+                // its internal handleAsync lands.
+                                e.ye = false;
+            }
+        }
+        // On a status-free node clearStatus is a guaranteed no-op: every field
+        // its body gates on is either _statusFlags or lives in the cold
+        // extension — no extension, no status to clear. (_x from an unrelated
+        // installer just makes clearStatus a cheap re-verified no-op.)
+        // A node born held keeps STATUS_UNINITIALIZED until its transaction
+        // commits: it has no committed value, and read() holds its readers.
+                if (e.S !== 0 || e.o !== null) clearStatus(e, t && stagedEntry === null);
+        // _optimisticLane is only ever assigned by engine paths (CONFIG_HAS_LANE
+        // is their sticky presence mark).
+                if (e.C & CONFIG_HAS_LANE && e.o?.Ce) GlobalQueue.St(e);
+    } catch (t) {
+        const n = t instanceof NotReadyError;
+        if (n && e.ye) {
+            // Loading window with an unready sync dependency: register for the
+            // source's settle (the settlePendingSource walk runs off
+            // _pendingSources + _blocked alone) but take NO read-visible pending
+            // status, no downstream propagation, no transition, no lane
+            // registration — the committed loading value keeps serving. If the
+            // node is currently errored the error stays the answer until this
+            // retry can actually run.
+            parkLoadingWindow(e, t);
+        } else {
+            // Track pending async in the lane (not the lane's source — it creates the lane
+            // but doesn't belong to it). Set lane BEFORE notifyStatus for downstream propagation.
+            if (n && currentOptimisticLane) GlobalQueue.Ot(e);
+            let i = false;
+            if (n) {
+                ext(e).Oe = true;
+                if (GlobalQueue.At !== null) i = GlobalQueue.At(e, _);
+            }
+            notifyStatus(e, n ? STATUS_PENDING : STATUS_ERROR, t, undefined, n ? e.o?.Ce : undefined);
+            // The replacement source is fully propagated now. If no new flight
+            // re-owned self, retire the superseded flight and its dependent copies.
+                        if (n && c && !e.o?.pe) settlePendingSource(e);
+            // A re-park drops what the earlier pass carried (#3456): a source this
+            // pass no longer reaches — its branch switched, or a fresh flight
+            // replaced the inputs' pending with its own — stays copied onto
+            // dependents that reached it only through here, and its landing walk
+            // stops at this node (nothing left to retire) before it finds them. A
+            // dependent then waits forever on a flight it has no path to. The
+            // re-park twin of the unchanged-value recovery sweep below; dependents
+            // with another path keep the source (retryReaches).
+                        if (n && f) for (const t of f) if (t !== e && !e.o?.ue?.has(t)) settlePendingSource(e, t);
+            if (i) GlobalQueue.k(e);
+        }
+    } finally {
+        tracking = O;
+        latestReadActive = C;
+        if (p) stale = R;
+        // Consume the missed-wake latch (#3037, set by insertSubs): a dep write
+        // landed beneath this pass on a link it had already validated. The wipe
+        // below must not key off DIRTY/CHECK — the read-time pull protocol
+        // (markNode(c) in read()) marks the running node as part of ordinary
+        // bookkeeping, and those marks are correctly discarded here.
+                S = (e.oe & REACTIVE_MISSED_WAKE) !== 0;
+        // REACTIVE_DISPOSED survives too (#3621): the pass may have disposed its
+        // own owner (a memo calling its root's `dispose()`, a cleanup doing so
+        // #3601/#3606), and `disposeChildren` set the flag on this node
+        // reentrantly. Dropped, the node read as live — `refresh()` re-ran it
+        // and `isDisposed()` lied.
+                e.oe = e.oe & (REACTIVE_ZOMBIE | REACTIVE_DISPOSED) | (t ? e.oe & REACTIVE_SNAPSHOT_STALE : 0);
+        context = N;
+    }
+    // The cast re-widens: TS narrowed `stagedEntry` to `null` at the reset
+    // above and does not invalidate that across the compute call that
+    // `enterStagedRead` runs under. No emitted code.
+        const G = stagedEntry;
+    stagedEntry = d;
+    // A node that died during its own pass (#3621) is dead at the end of it,
+    // and the pass is void. Its owner's teardown already unlinked its deps,
+    // removed it from the heap and ran its cleanups; what remains is what the
+    // body did AFTER the `dispose()` call: reads that re-linked the dead node
+    // to its sources (unlinked here — the leak that kept it re-running in a
+    // torn-down tree), a flight it may have started (retired: the landing
+    // checks `_inFlight` identity), and the value it returned. That value is
+    // NOT published: a dead node freezes at its last committed value (#3024),
+    // so nothing is staged, committed, or propagated to subscribers, and an
+    // effect's run is not enqueued (runEffect would refuse it anyway). The
+    // attribution frame opened at the top is still closed.
+        if (e.oe & REACTIVE_DISPOSED) {
+        clearDeps(e);
+        if (e.o !== null) e.o.pe = null;
+        if (attrHooks !== null) attrHooks.recomputeEnd(e, t, false, false, false, false);
+        currentOptimisticLane = A;
+        return;
+    }
+    if (!e.o?._) {
+        // Observe-tier fan-in (HUGE_FAN_IN): the validated prefix [_deps.._depsTail]
+        // IS this pass's distinct sources — count it here rather than per link. A
+        // begin/end bracket around the pass plus a per-link increment measured
+        // -5.8% on createRenderEffects:create1to1 (CodSpeed, dev tier) and cost
+        // several points of the shape wins elsewhere; this walk is a fraction of
+        // the reads that built the list and keeps no module state, so nested
+        // pulls need no save/restore. (The stale tail is trimmed at the end of
+        // this pass, or at its commit — see recompute's tail.)
+        {
+            let t = 0;
+            for (let n = e.Ee; n !== null; n = n.Ie) {
+                t++;
+                if (n === e.Je) break;
+            }
+            if (t >= GRAPH_SIZE_WARN_AT) noteFanIn(e, t);
+        }
+        // INV-11 (#3330): the equality gate compares against the slot this run
+        // publishes to. An override-covered node publishes the override; a lane
+        // recompute (OPT-dirty) direct-commits `_value` — the lane's own reveal
+        // schedule — so a transaction-held `_pendingValue` that already equals
+        // the new result is not "unchanged": the screen still shows `_value`.
+        // Only a transaction-staged run compares against `_pendingValue`.
+                const u = o ? unwrapOverride(e.o?.we) : l || e._e === NOT_PENDING ? e.ce : e._e;
+        let a = false;
+        try {
+            a = !n && r || !e.Fe || !e.Fe(u, I);
+        } catch (t) {
+            // A throwing user comparator is an error of this node's computation.
+            // Route it through the same status path as a compute-phase throw so
+            // error boundaries contain it; otherwise it unwinds the scheduler
+            // flush, bypassing every boundary and wedging the queue (#2837).
+            notifyStatus(e, STATUS_ERROR, t);
+        }
+        // A committed derived change becomes a cause for this node's subscribers,
+        // chaining their attribution through this node to the root write.
+                if (attrHooks !== null) {
+            i = a && !e.o?._;
+            if (i && !n && !t) attrHooks.derivedChanged(e);
+        }
+        // Effects use `_equals: false` (no per-effect closure). The side effects that
+        // the equals closure used to perform — flagging the effect dirty and enqueueing
+        // its runner — happen here instead. `!create` matches the previous `initialized`
+        // gate: the explicit recompute(node, true) inside effect() does not enqueue, so
+        // effect() can call its runner synchronously for the first run.
+                if (n && a) {
+            e.it = !e.o?._;
+            // Reuse one bound runner per effect — runEffect no-ops on a stale
+            // `_modified`, so re-enqueueing the same function is harmless.
+                        if (!t) {
+                e.T.enqueue(n, e.Ct ??= GlobalQueue.Rt.bind(null, e));
+                // Contested effect (#3322). Effects don't entangle transactions (a
+                // shared effect is not shared state), yet they have one value slot:
+                // when this write commits under a different transaction
+                // (`activeTransition`, null = mainline) than the one that produced the
+                // previous value (`_valueTransition`), that view is gone. Rather than
+                // merge the two, each commit re-derives the effect against its own
+                // committed world (Transition._contested, re-dirtied by
+                // finalizePureQueue ahead of the heap run). The previous owner always
+                // needs it if still live: its commit is silent (staging already
+                // notified) and the value it computed is gone. The new owner needs it
+                // too, for the same reason, unless it is mainline — mainline publishes
+                // what it computes. A previous owner that was mainline, or already
+                // committed, left nothing to protect.
+                                let t = e.Gt;
+                if (t !== activeTransition) {
+                    e.Gt = activeTransition;
+                    if (t !== null && (t = currentTransition(t)) !== activeTransition && !t.Dt) {
+                        (t.ht ??= []).push(e);
+                        if (activeTransition !== null) (activeTransition.ht ??= []).push(e);
+                    }
+                }
+            }
+        }
+        if (e.o?._) ; else if (a) {
+            const i = o ? e.o?.we : undefined;
+            if (t && G === null || 
+            // Plain sync flush (no transition on either side) commits effect
+            // values directly — the pending round-trip (queuePendingNode +
+            // commitPendingNodes) exists to sequence transition reveals, and
+            // paying it per effect on the plain path is pure overhead.
+            // DIRECT_COMMIT effects (resolve/until) commit directly even under
+            // their own held transition: their applies deliver on a microtask,
+            // not the stashed queues, so a staged value would hand the immediate
+            // apply stale state — see CONFIG_DIRECT_COMMIT.
+            n && G === null && (activeTransition !== e.me || activeTransition === null || e.C & CONFIG_DIRECT_COMMIT) || l) {
+                // Lanes stage (#3479): a lane pass on a memo publishes its speculative
+                // result as an OVERRIDE — `_value` stays the committed truth, so a
+                // reader off the lane (A17's committed view, #3460) sees a whole
+                // committed frame: the source's shadow and its derivations together,
+                // never a committed shadow beside a speculative memo. The lane's own
+                // readers and untracked reads see the override (A17); the revert
+                // drops it and re-derives (resolveOptimisticNodes). Effects keep the
+                // direct commit — their `_value` is a run result the lane's queues
+                // already sequence. Either way the lane pass drops any superseded
+                // older hold so its queued commit can't clobber the fresh frame: a
+                // hold staged on an earlier, lane-free pass of the SAME transaction
+                // is superseded — left in place, the commit published the older
+                // frame over the fresh one (#3377). A reversion pass (OPT-dirty with
+                // no lane — the override dropped) commits directly: it IS the truth.
+                if (l && !n && currentOptimisticLane !== null) GlobalQueue.je(e, I, currentOptimisticLane); else e.ce = I;
+                if (l) e._e = NOT_PENDING;
+            } else {
+                e._e = I;
+                if (G !== null) {
+                    // Born held: the pass ran from mainline and derived from this
+                    // transaction's staged world (enterStagedRead). Its value is the
+                    // transaction's — staged into it directly, stamped, and committed
+                    // with it; mainline's batch never sees it. A born-held effect
+                    // skips its synchronous first run (effect()) and is replayed by
+                    // the commit like a stale reader that showed the committed frame.
+                    // (A re-pass under a fresh boundary — the in-flush form — is
+                    // already the transaction's: restaged, not re-queued.)
+                    if (e.me !== G) {
+                        e.me = G;
+                        G.In.push(e);
+                    }
+                    if (n) G.Tt.add(e);
+                    // Under a boundary that has not revealed, a held first value is
+                    // something not ready under it (#3540): the boundary collects this
+                    // node as a source and shows its fallback; the source stays
+                    // collected while born held (CollectionQueue._checkSources) and
+                    // is released by the commit that initializes it. A boundary that
+                    // already shows content is not told — it holds like any reader.
+                                        if (underFreshLoadingBoundary(e)) e.T.notify(e, STATUS_PENDING, STATUS_PENDING, new NotReadyError(e));
+                }
+                // A window landing that gets held re-opens the window until the hold
+                // commits — the verdict's held-value branch is window-gated (#2990).
+                                if (E) e.ye = true;
+                // A staged sync recompute is a write path like setSignal/asyncWrite,
+                // so sync derivations of held sources stay visible to isPending()/latest()
+                // (#2831). Both companion writes are transition-scoped (optimistic) and
+                // auto-revert/re-derive at commit. Not gated on an active transition:
+                // a plain flush can still become a hold after this recompute — an
+                // async memo downstream pends and the batch is adopted into a
+                // transaction (scheduler.enterTransition) — and nothing re-derives
+                // the companion at adoption, so a memo held that way read
+                // isPending() false while its held source read true (#3413).
+                                if (e.C & CONFIG_HAS_COMPANIONS && GlobalQueue.Le !== null) GlobalQueue.Le(e, I);
+            }
+            // insertSubs only walks _subs (no scheduling of its own), so a
+            // subscriber-less node has nothing to notify.
+                        if (e.u !== null && (!o || l || e.o?.we !== i)) insertSubs(e, l || o);
+            // A18 supersession, sync twin of asyncWrite's override branch (#3331):
+            // this pass published truth that differs from the override (the gate
+            // compared against it) into the transaction-held slot. Whether the
+            // source is this node's own async or an upstream node it derives from
+            // synchronously makes no difference — "the source recomputed". The
+            // override stays displayed until the commit; the graph moves to the
+            // staged truth now (plain channel, lane demoted). Ordering: "a new
+            // value from the source" postdates the override — an override written
+            // in this same tick (optimisticWrite stamps `_overrideTime`) is the
+            // newer intent over whatever this pass derives from the batch's staged
+            // inputs, and is not superseded by it.
+             else if (o && !l && e.o.Pt !== clock) GlobalQueue.Qe(e, I);
+        } else if (o) {
+            // Unchanged value (equals the override) recomputed while the override
+            // is active: _value may still be stale, so hold the authoritative value
+            // for commit on its own transition's schedule — invisibly (A17/A18).
+            if (e._e === NOT_PENDING) queuePendingNode(e);
+            e._e = I;
+            if (E) e.ye = true;
+ // see the held branch above (#2990)
+            // A confirmation after a supersession restores the override as the
+            // graph's value and notifies; a plain confirmation wakes only an
+            // authoritative-view reader (until()'s predicate, refresh()'s waiter)
+            // that observed this node past its override — "authoritative arrival
+            // equal to the override" is exactly the acknowledgment it waits for;
+            // A17 silence holds for every ordinary subscriber. Both live in the
+            // engine's supersedeOverride.
+                        GlobalQueue.Qe(e, I);
+        } else if (e.ft != T) {
+            for (let t = e.u; t !== null; t = t.Pe) {
+                insertIntoHeapHeight(t.Ae, queueFor(t.Ae));
+            }
+        }
+        // Silent recovery: errored → unchanged value fires no notification, but
+        // dependents still holding the propagated error consumed their dirty flag
+        // in an errored run and may sit on stale commits (#2949). Changed-value
+        // recoveries ride insertSubs above; a comparator throw re-errored the node
+        // (el._x?._error re-set), so this only runs on a genuinely clean recovery.
+                if (!a && !e.o?._) {
+            if (s !== undefined) settleErroredDependents(e, s);
+            // Self-registration (this node's own superseded flight) is the #3181
+            // sweep's business below — retiring it here too would walk twice.
+                        if (f) for (const t of f) if (t !== e) settlePendingSource(e, t);
+        }
+        // #3181: a synchronous settle supersedes the old landing callback, so
+        // recompute owns its pending-source sweep. An uninitialized node without
+        // a replacement source still has no truth to reveal and must stay parked.
+                if (c && !(e.S & (STATUS_PENDING | STATUS_UNINITIALIZED))) {
+            settlePendingSource(e);
+            // The superseded flight will not enter its waiting transactions at
+            // landing. Recheck them now so an unchanged value cannot leave their
+            // writes parked after the last pending source has settled.
+                        wakeParked();
+        }
+    }
+    // A REPORTER whose pass stopped reading a source it reported on (a gate
+    // closed) stops counting for the transaction waiting on it (A15 / #3426:
+    // the hold lasts while a live reporter observes the flight). Nothing else
+    // re-judges a parked transaction (see wokenTransitions; the disposal
+    // (#3372) and boundary (#3375) twins of this site), so the writes it held
+    // stayed staged for as long as the flight stayed up — forever, for one
+    // that never lands (fuzzer #3446 P1, spec O3). The event is "this pass
+    // dropped a dep" — deps past `_depsTail` (trimmed below, or kept by A30
+    // for a staged pass), or a pass that read nothing — not "recovered from
+    // pending": a reporter registered by the stale-reader carve-out
+    // (heldFromStale, an INITIALIZED source refetching) displays the committed
+    // value and is never pending (fuzzer case 79). Every parked transaction,
+    // not the reporter's stamp: the transaction waiting on it registered it
+    // without stamping it. One idle pass per parked transaction; done ones
+    // return at re-entry. Effects only — reporters register from render-effect
+    // notification (INV-3).
+    // (`_depsTail` was reset at the top of the pass; TS keeps that narrowing.)
+        const D = e.Je;
+    if (n && (a && !(e.S & STATUS_PENDING) || (D === null ? e.Ee !== null : D.Ie !== null))) wakeParked();
+    // Dependencies are the committed frame's until it is replaced (A30, #3410; the
+    // deps twin of the held children above): a pass that staged its value
+    // leaves the previous pass's tail linked for `commitPendingNode` to trim,
+    // so a write to a dependency the committed value still derives from
+    // reaches this node — and joins its hold if the stage is transaction-held
+    // by then (a plain flush decides nothing here: the transaction that holds
+    // the pass may open later in the same flush). A pass that published
+    // directly (a first pass, a lane's own reveal) trims now. An errored pass
+    // (a throw, NotReady included, or a comparator throw above) keeps its full
+    // list as before — `_depsTail` marks where it stopped — and the commit
+    // skips it by the same `_error`. An effect's frame is the run its value is
+    // applied by, not the value slot (#3438): a direct-committed pass that
+    // still owes a run (`_modified`) has not replaced what the last run
+    // published — the same flush may stash that run into a transaction it
+    // opens later — so its tail waits for `runEffect` to trim once the run
+    // applies. A pass that changed nothing replaced nothing either (#3469): the
+    // same flush may park with its inputs held, and the committed frame still
+    // derives from the tail — its trim waits on the flush's verdict (heldTrims).
+    // A tracked effect's pass IS its run (it bypasses the heap and runs after
+    // the commit): the frame is replaced, trim now.
+        if (!e.o?._ && e._e === NOT_PENDING && !(n && e.it)) {
+        if (t || l || n === EFFECT_TRACKED) trimStaleDeps(e); else if (e.Je?.Ie ?? e.Ee) heldTrims.push(e);
+    }
+    // Attribution hook: fired before the lane restore so `currentOptimisticLane`
+    // still reflects THIS run's posture. The facts distinguish an overlay
+    // recompute (optimistic lane, transition replay, transition-held commit)
+    // from a plain committed one — the engine must not blame overlay runs as
+    // waste or double-count them against plain aggregates.
+        if (attrHooks !== null) attrHooks.recomputeEnd(e, t, i, l || currentOptimisticLane !== null, activeTransition !== null || e.me !== null, e._e !== NOT_PENDING);
+    currentOptimisticLane = A;
+    // A parked LANE frame is not a hold (#3662): its drain is the effect's own
+    // run, not a commit — the node is neither queued nor stamped for it, and
+    // the release below (for transaction zombies) leaves it parked.
+        const h = (e.C & CONFIG_LANE_FRAME) !== 0;
+    const P = e._e !== NOT_PENDING || !h && e.o !== null && (e.o.dt !== null || e.o.Et !== null) || (e.S & (STATUS_PENDING | STATUS_UNINITIALIZED)) !== 0;
+    // Override-covered holds (hasOverride) always queue: their commit belongs
+    // to their own transition's schedule (A18 re-rule) and is unobservable
+    // under the override (A17). Revert no longer commits anything, so an
+    // unqueued covered hold would leak (INV-7) once the revert clears
+    // _transition.
+    
+    // While a pass's result waits on a commit, its children (and `_disposal`)
+    // wait with it, and a re-run may tear them down immediately
+    // (CONFIG_HELD_CHILDREN, #3404). One exception: a transaction-owned effect
+    // recomputed mainline (contested, #3322) published its value directly — a
+    // `_pendingValue` left from an earlier held pass is that transaction's,
+    // not this one's — so this pass's children are the frame's, and any
+    // zombies deferred at the top are superseded on that same frame. Its
+    // commit rides the transaction, not this flush: release them here rather
+    // than let two generations render at once. A LANE pass is the same case
+    // (#3662): it direct-commits ahead of its transaction, so its children are
+    // the frame's too. Left flagged for an OLDER frame's zombies, the stamped
+    // transaction's same-flush re-run below disposed the lane's freshly built
+    // children (an insert's inner effect, its run still queued in the lane)
+    // and held their replacements for a commit that never came.
+        let F = P && (!t || G !== null || (e.S & STATUS_PENDING) !== 0);
+    if (F && (!e.me || o)) queuePendingNode(e); else if (F && (activeTransition === null || l) && !(e.S & (STATUS_PENDING | STATUS_UNINITIALIZED))) {
+        F = false;
+        if (!h) disposeChildren(e, false, true);
+    }
+    if (F) e.C |= CONFIG_HELD_CHILDREN; else e.C &= ~CONFIG_HELD_CHILDREN;
+    // (A born-held pass IS the transaction's staged view — nothing to refresh.)
+        if (e.me && n && activeTransition !== e.me && G === null) {
+        // The re-run refreshes the transaction's STAGED view (_pendingValue); the
+        // value this pass published in _value belongs to the run that just
+        // finished. Keep that ownership, or the effect phase parks a
+        // mainline-computed value with the transaction (#3412).
+        const t = e.Gt;
+        runInTransition(e.me, () => recompute(e));
+        e.Gt = t;
+    }
+    // Missed-wake reschedule (see the finally above): values this pass read
+    // before the nested commit are stale, so run again now that the heap will
+    // accept the node. Equality gates stop same-value landings from cascading,
+    // and a re-run only latches again if another nested commit changes a dep
+    // beneath it — convergent unless deps genuinely keep changing.
+        if (S) {
+        enqueueSub(e);
+        schedule();
+    }
+}
+
+function updateIfNecessary(e) {
+    // Never re-enter a node that is currently computing: its dep bookkeeping
+    // (_depsTail/_depGen) is live, and a nested recompute would corrupt it.
+    // A mid-pass mark stays latched for recompute's own tail to reschedule
+    // (#3037); readers meanwhile serve the values the pass has so far.
+    // Never recompute a DISPOSED node either: recompute rewrites _flags and
+    // would resurrect it (#2983) — readers serve its last value.
+    if (e.oe & (REACTIVE_RECOMPUTING_DEPS | REACTIVE_DISPOSED)) return;
+    if (e.oe & REACTIVE_CHECK) {
+        for (let t = e.Ee; t; t = t.Ie) {
+            const n = t.De;
+            const i = n.Ne || n;
+            if (i.Se) {
+                updateIfNecessary(i);
+            }
+            if (e.oe & REACTIVE_DIRTY) {
+                break;
+            }
+        }
+    }
+    if (e.oe & (REACTIVE_DIRTY | REACTIVE_OPTIMISTIC_DIRTY) || e.o?._ && e.Ge < clock && !e.o?.pe) {
+        recompute(e);
+    }
+    // The guard above refused an already-disposed node; the recompute it just
+    // ran may have disposed it (#3621) — carry the flag, or it comes back alive.
+    // The manual-write mask is state, not scheduling (#3612): it says the
+    // node's staging is a PROPOSAL, and only the commit (or a later-tick
+    // refresh, #3026) lifts it — a pull that recomputed nothing must not, or
+    // an isPending()/latest() probe decided whether a later write to a held
+    // node was a second proposal or the derivation's `prev`.
+        e.oe = e.oe & (REACTIVE_SNAPSHOT_STALE | REACTIVE_IN_HEAP | REACTIVE_IN_HEAP_HEIGHT | REACTIVE_ZOMBIE | REACTIVE_DISPOSED | REACTIVE_MANUAL_WRITE);
+}
+
+function computed(e, t) {
+    const n = t?.transparent ?? false;
+    // `in` (not `!== undefined`): an explicit `loadingValue: undefined` on a
+    // `T | undefined` node is a real commit #0. The typeof guard tolerates
+    // non-object option values that older call shapes force through `as any`.
+        const i = t !== null && typeof t === "object" && "loadingValue" in t;
+    // Two literals, one per tier, selected at build time (the observe flag is
+    // a literal after replacement; the untaken branch is dead code). The observe
+    // literal is the prod literal plus its `_name` slot — a slot in the
+    // boilerplate, because a post-construction `self._name = …` forces a
+    // hidden-class transition and an out-of-object property store on EVERY node
+    // (measured: the whole of the observe tier's creation overhead). Keep the
+    // two in sync — the dist artifact test pins observe's key set to prod's
+    // plus `_name`.
+        const l = {
+        id: inheritId(t, n, context),
+        C: (n ? CONFIG_TRANSPARENT : 0) | (t?.ownedWrite ? CONFIG_OWNED_WRITE : 0) | (!context || t?.lazy ? CONFIG_AUTO_DISPOSE : 0) | (t?.sync ? CONFIG_SYNC : 0) | (t?.Z ? CONFIG_NO_SNAPSHOT : 0) | (
+        // Plumbing is an observe-tier notion (a name and records to
+        // withhold); the prod literal never carries the bit.
+        t?.Ft ? CONFIG_PLUMBING : 0) | (snapshotCaptureActive && ownerInSnapshotScope(context) ? CONFIG_IN_SNAPSHOT_SCOPE : 0),
+        Fe: t?.equals ?? isEqual,
+        Ve: null,
+        T: context?.T ?? globalQueue,
+        rt: context?.rt ?? defaultContext,
+        Nt: 0,
+        Se: e,
+        ce: i ? t.loadingValue : undefined,
+        ft: 0,
+        gt: undefined,
+        vt: null,
+        Ee: null,
+        Je: null,
+        We: 0,
+        u: null,
+        Ht: null,
+        _parent: context,
+        Fn: null,
+        kt: null,
+        _n: null,
+        oe: t?.lazy ? REACTIVE_LAZY : REACTIVE_NONE,
+        S: i ? 0 : STATUS_UNINITIALIZED,
+        Ge: clock,
+        _e: NOT_PENDING,
+        me: null,
+        Lt: -1,
+        ye: i,
+        o: null,
+        // The slot is always present (hidden class); plumbing leaves it
+        // unset, which `ownerPath` skips.
+        _name: t?.Ft ? undefined : t?.name ?? "computed"
+    };
+    if (t?.unobserved) ext(l).bt = t.unobserved;
+    setupComputedNode(l, t);
+    return l;
+}
+
+/** Lazily allocate a node's cold extension (ONE shape for signals and
+ * computeds — `_x` access stays monomorphic). Installers write through
+ * this; hot paths read `el._x?._field` gated by the _config presence bits.
+ * Never call ext() just to store a field's default. */ function ext(e) {
+    return e.o ??= {
+        we: undefined,
+        Vt: undefined,
+        Pt: 0,
+        Ut: NOT_PENDING,
+        yt: 0,
+        Ce: undefined,
+        xe: undefined,
+        Me: undefined,
+        $n: undefined,
+        t: 0,
+        pe: null,
+        be: null,
+        _: undefined,
+        Oe: undefined,
+        ue: undefined,
+        h: undefined,
+        ve: false,
+        i: null,
+        bt: undefined,
+        ct: undefined,
+        Et: null,
+        dt: null,
+        xt: undefined
+    };
+}
+
+/**
+ * Build an Effect node with all effect-specific fields baked into a single object literal,
+ * so V8 sees the full hidden class shape at construction time. Effects always run in lazy
+ * mode (recompute is called explicitly by `effect()`), so we hardcode the lazy bits and skip
+ * the auto-dispose CONFIG bit (effect() previously cleared it post-construction).
+ */ function createEffectNode(e, t, n, i, l) {
+    const u = l?.transparent ?? false;
+    // Prod and observe boilerplates — see computed() for why the observe tier
+    // gets its `_name` as a literal slot rather than a write after the fact.
+    // The default label is the node kind (tracked effects relabel their computed
+    // in trackedEffect); the wrappers in signals.ts no longer spread a name into
+    // the options to get it.
+        const o = {
+        id: inheritId(l, u, context),
+        C: (u ? CONFIG_TRANSPARENT : 0) | (l?.ownedWrite ? CONFIG_OWNED_WRITE : 0) | (l?.sync ? CONFIG_SYNC : 0) | (l?.Qt ?? 0) | (snapshotCaptureActive && ownerInSnapshotScope(context) ? CONFIG_IN_SNAPSHOT_SCOPE : 0),
+        Fe: false,
+        Ve: null,
+        T: context?.T ?? globalQueue,
+        rt: context?.rt ?? defaultContext,
+        Nt: 0,
+        Se: e,
+        ce: undefined,
+        ft: 0,
+        gt: undefined,
+        vt: null,
+        Ee: null,
+        Je: null,
+        We: 0,
+        u: null,
+        Ht: null,
+        _parent: context,
+        Fn: null,
+        kt: null,
+        _n: null,
+        oe: REACTIVE_LAZY,
+        S: STATUS_UNINITIALIZED,
+        Ge: clock,
+        _e: NOT_PENDING,
+        me: null,
+        Lt: -1,
+        ye: false,
+        it: false,
+        Mt: undefined,
+        wt: t,
+        Wt: n,
+        qt: undefined,
+        He: i,
+        Gt: null,
+        o: null,
+        _name: l?.name ?? "effect"
+    };
+    // Effects dispatch status through the SHARED notifier (statusNotifierOf,
+    // keyed off _type) — storing it per node forced a full NodeExtension
+    // allocation on EVERY effect at creation (an alloc + 19 field stores,
+    // +23% effect creation, caught by the creation benches). Only genuinely
+    // per-node channels (boundaries) live on _x.
+        if (l?.unobserved) ext(o).bt = l.unobserved;
+    setupComputedNode(o, lazyOptions);
+    return o;
+}
+
+/**
+ * The shared status notifier for effect nodes, installed once by effect.ts
+ * at module evaluation (`this`-dispatched — one function serves every
+ * effect, so nodes never store it). Boundary computeds keep their own
+ * per-node channel on `_x._notifyStatus`, which takes precedence.
+ */ let effectStatusNotify = null;
+
+function setEffectStatusNotify(e) {
+    effectStatusNotify = e;
+}
+
+/** Resolve a node's status notifier: an own `_x` channel (boundaries) wins;
+ * effect nodes (`_type` — EFFECT_PURE is 0, and only effect literals carry
+ * the field) fall back to the shared notifier. Presence doubles as the
+ * "display consumer" membership test in the status walks, exactly as the
+ * per-node field did when every effect carried one. */ function statusNotifierOf(e) {
+    const t = e.o?.h;
+    if (t !== undefined) return t;
+    return e.He ? effectStatusNotify ?? undefined : undefined;
+}
+
+const lazyOptions = {
+    lazy: true
+};
+
+function setupComputedNode(e, t) {
+    e.vt = e;
+    const n = context?.Zt ? context.Yt : context;
+    if (context) linkChild(context, e);
+    if (n) e.ft = n.ft + 1;
+    if (GlobalQueue.Bt !== null) GlobalQueue.Bt(e);
+    !t?.lazy && recompute(e, true);
+    if (snapshotCaptureActive && !t?.lazy) {
+        if (!(e.S & STATUS_PENDING) && !(e.C & CONFIG_NO_SNAPSHOT)) {
+            ext(e).ct = e.ce === undefined ? NO_SNAPSHOT : e.ce;
+            e.C |= CONFIG_HAS_SNAPSHOT;
+            snapshotSources.add(e);
+        }
+    }
+}
+
+function signal(e, t, n = null) {
+    // Prod and observe boilerplates — see computed(). The observe literal adds
+    // `_name` and `_owner` (the creating owner, stamped by registerGraph for
+    // createSignal nodes so ownerPath can locate signal subjects; null here,
+    // and staying null on internal signals — one shape either way).
+    const i = {
+        Fe: t?.equals ?? isEqual,
+        C: (t?.ownedWrite ? CONFIG_OWNED_WRITE : 0) | (t?.Z ? CONFIG_NO_SNAPSHOT : 0),
+        ce: e,
+        u: null,
+        Ht: null,
+        Ge: clock,
+        Ne: n,
+        ge: n?.o?.i || null,
+        jt: null,
+        _e: NOT_PENDING,
+        me: null,
+        Lt: -1,
+        o: null,
+        _name: t?.name ?? "signal",
+        ie: null
+    };
+    if (t?.unobserved) ext(i).bt = t.unobserved;
+    if (n) linkFirewallChild(n, i);
+    if (snapshotCaptureActive && !(i.C & CONFIG_NO_SNAPSHOT) && !((n?.S ?? 0) & STATUS_PENDING)) {
+        ext(i).ct = e === undefined ? NO_SNAPSHOT : e;
+        i.C |= CONFIG_HAS_SNAPSHOT;
+        snapshotSources.add(i);
+    }
+    return i;
+}
+
+// ---------------------------------------------------------------------------
+// SLOT SIGNALS (store leaves) — the create-floor diet. Store mounts
+// materialize one signal per touched leaf (~13 × rows on dbmon), so per-node
+// allocations are mount bytes: the generic path costs an options object, an
+// equals closure, an unobserved closure, a NodeExtension to hold it, and
+// three post-construction expandos (acc/px/pxv → hidden-class transitions).
+// slotSignal bakes everything into ONE literal: `_host`/`_key` backrefs
+// replace the closures (equals is a method call — `this` is the node; the
+// unobserved sweep dispatches CONFIG_SLOT_NODE to one shared hook), and the
+// store's wrap-cache fields are pre-shaped.
+/** The shared slot-node unobserved handler — a live binding read directly by
+ * the sweep sites (no wrapper frame, no null check: a CONFIG_SLOT_NODE node
+ * existing implies the store module loaded and registered the hook). */ let slotUnobservedHook;
+
+/** Install the shared slot-node unobserved handler (store module, once). */ function setSlotUnobserved(e) {
+    slotUnobservedHook = e;
+}
+
+/** Push a new node onto its firewall's child chain (the literal already
+ * points `_nextChild` at the old head). Doubly linked so a released leaf
+ * unlinks in O(1) — the chain is walked per mark of the projection and
+ * would otherwise grow by one node per leaf ever read (#3351). */ function linkFirewallChild(e, t) {
+    const n = t.ge;
+    if (n !== null) n.jt = t;
+    ext(e).i = t;
+    e.C |= CONFIG_FW_CHILDREN;
+}
+
+/** Release a firewall child the store no longer addresses (unobserved sweep
+ * dropped it from its target's cache): unlink it from the chain so the
+ * projection stops retaining it and its last value. The node keeps its own
+ * `_nextChild` so a walk that is mid-chain on it still terminates. It also
+ * leaves `_companionChildren` (#3503): the companions themselves are
+ * permanent on the node, but the node is unreachable through the store, so
+ * the set would only retain it and its last value. */ function unlinkFirewallChild(e) {
+    const t = e;
+    const n = t.Ne;
+    if (!n) return;
+    const i = t.jt;
+    const l = t.ge;
+    if (i !== null) i.ge = l; else if (n.o.i === t) n.o.i = l;
+    if (l !== null) l.jt = i;
+    t.jt = null;
+    // #3503: the companion set is the last projection-side reference to a
+    // released leaf; leaving it there kept the leaf and its last value alive
+    // for the projection's lifetime after every latest()/isPending() reader
+    // was disposed.
+        n.o.xt?.delete(t);
+}
+
+function slotSignal(e, t, n, i, l, u = null) {
+    // Prod and observe boilerplates — see computed(). The store relabels the
+    // observe slot (`store.<key>`) when the attribution engine is installed.
+    const o = {
+        Fe: t,
+        C: CONFIG_OWNED_WRITE | CONFIG_SLOT_NODE,
+        ce: e,
+        u: null,
+        Ht: null,
+        Ge: clock,
+        Ne: u,
+        ge: u?.o?.i || null,
+        jt: null,
+        _e: NOT_PENDING,
+        me: null,
+        Lt: -1,
+        o: null,
+        Kt: n,
+        zt: i,
+        acc: l,
+        px: undefined,
+        pxv: undefined,
+        _name: "signal"
+    };
+    if (u) linkFirewallChild(u, o);
+    if (snapshotCaptureActive && !((u?.S ?? 0) & STATUS_PENDING)) {
+        ext(o).ct = e === undefined ? NO_SNAPSHOT : e;
+        o.C |= CONFIG_HAS_SNAPSHOT;
+        snapshotSources.add(o);
+    }
+    return o;
+}
+
+function optimisticSignal(e, t) {
+    const n = signal(e, t);
+    ext(n).we = NOT_PENDING;
+    n.C |= CONFIG_OPTIMISTIC;
+    return n;
+}
+
+function optimisticComputed(e, t) {
+    const n = computed(e, t);
+    ext(n).we = NOT_PENDING;
+    n.C |= CONFIG_OPTIMISTIC;
+    return n;
+}
+
+function isEqual(e, t) {
+    return e === t;
+}
+
+/**
+ * Runs `fn` outside of any reactive tracking — reads inside `fn` will not
+ * subscribe the current scope. Returns whatever `fn` returns.
+ *
+ * Use `untrack` inside a memo or effect when you need to read a signal once
+ * without making the surrounding computation depend on its future changes.
+ *
+ * Pass a `strictReadLabel` string to enable a dev-mode warning: any reactive
+ * read inside `fn` that isn't inside a nested tracking scope will log a
+ * warning naming the label.
+ *
+ * @example
+ * ```ts
+ * createEffect(
+ *   () => trigger(),                 // tracks `trigger` only
+ *   () => {
+ *     const snapshot = untrack(() => state); // read once, untracked
+ *     log(snapshot);
+ *   }
+ * );
+ * ```
+ */ function untrack(e, t) {
+    if (GlobalQueue.Jt === null && !tracking && true) return e();
+    const n = tracking;
+    tracking = false;
+    try {
+        if (GlobalQueue.Jt !== null) return GlobalQueue.Jt(e);
+        return e();
+    } finally {
+        tracking = n;
+    }
+}
+
+/**
+ * Set while runtime bookkeeping reads a node inside another node's pass (a
+ * loading boundary priming its tree at creation, from whatever pass is
+ * mounting it). `context` is that node, but the read is nobody's: the value
+ * is probed, never derived from, so nothing the read would normally record
+ * on `context` may be recorded — not the untracked-pending re-run link
+ * (`read`, `!tracking`; #3528: a boundary's `on` key read this way from
+ * `notify` linked the key's source into an unrelated async memo, a cycle that
+ * never converged), and not a transaction entry (`enterStagedRead`; #3540: a
+ * born-held tree would otherwise pull the mounting pass into the hold).
+ */ let spectating = false;
+
+/**
+ * Evaluates `fn` untracked, recording nothing on the current `context`: no
+ * untracked-pending re-run link, no transaction entry. For bookkeeping reads
+ * made on behalf of no node — see `spectating`.
+ */ function spectate(e) {
+    const t = spectating;
+    spectating = true;
+    try {
+        return untrack(e);
+    } finally {
+        spectating = t;
+    }
+}
+
+/**
+ * Bring a computed to a readable state: lazy/disposed nodes are (re)computed;
+ * an isPending() probe (`refresh`) additionally pulls the node fully up to
+ * date so its status flags reflect the current graph.
+ */ function prepareComputed(e, t) {
+    if (e.oe & REACTIVE_LAZY) {
+        e.oe &= ~REACTIVE_LAZY;
+        recompute(e, true);
+    } else if (e.oe & REACTIVE_DISPOSED) {
+        // Two disposal lifecycles share the flag (#3024). Observation-lifecycle
+        // nodes (CONFIG_AUTO_DISPOSE) are dormant — torn down by unobserved()
+        // when the last subscriber left — and reads reawaken them; that is the
+        // pay-for-use contract. Owner-lifecycle nodes are dead: recomputing would
+        // re-run user code in a torn-down tree (and discard manual writes on
+        // derived-writable signals), so reads return the last committed value.
+        if (e.C & CONFIG_AUTO_DISPOSE) {
+            const t = e._parent;
+            if (t !== null) {
+                // A dormant node was off the chain when its owner died, so the strip
+                // in disposeChildren missed it: freeze here instead (#3024).
+                if (t.oe & REACTIVE_DISPOSED) {
+                    e.C &= ~CONFIG_AUTO_DISPOSE;
+                    return;
+                }
+                // A zombie never left its chain: settleAutodispose releases without
+                // the !ZOMBIE check its siblings have, so this guard is load-bearing.
+                                if (!(e.oe & REACTIVE_ZOMBIE)) linkChild(t, e);
+            }
+            recompute(e, true);
+        }
+    } else if (t) {
+        updateIfNecessary(e);
+    }
+}
+
+/**
+ * Sentinel returned by readNodeFast when the plain-signal fast path does not
+ * apply and the caller must fall back to the full read().
+ */ const READ_SLOW = Symbol("read-slow");
+
+/**
+ * read()'s plain-signal fast path as a standalone entry for hot callers
+ * (store traps). Safe to substitute for read() only because the bail
+ * conditions mirror read()'s prelude and fast-path guard exactly: the
+ * latestRead and pendingCheck windows run side-effectful hooks before the
+ * fast path, `_fn` nodes need prepareComputed, and firewall / override /
+ * snapshot / transition / lane / dev-strictRead state all take the full
+ * resolution. Anything slow returns READ_SLOW; the caller then calls read().
+ */
+/**
+ * Wake only authoritative-view readers (until() predicates) subscribed to `el`.
+ * The A17-silent ack paths — an authoritative arrival equal to the active
+ * override — use this so the predicate re-evaluates without re-firing
+ * ordinary subscribers whose visible (override) value did not change.
+ * Pay-for-use: reached through GlobalQueue._notifyAuthoritativeObservers,
+ * installed at first until() call — apps that never use until() shake it.
+ */ function notifyAuthoritativeObservers(e) {
+    for (let t = e.u; t !== null; t = t.Pe) {
+        const e = t.Ae;
+        if (!(e.C & CONFIG_AUTHORITATIVE_READ)) continue;
+        // Missed-wake latch (#3037), same contract as insertSubs: the reader may
+        // itself have pulled this recompute (updateIfNecessary from its own
+        // read), and the heap refuses RECOMPUTING nodes — latch so recompute's
+        // tail reschedules it with the staged value visible.
+                if (e.oe & REACTIVE_RECOMPUTING_DEPS && t.Ze === e.We && t !== e.Je) e.oe |= REACTIVE_MISSED_WAKE;
+        enqueueSub(e);
+    }
+    schedule();
+}
+
+/** Installs the authoritative-reader wakeup hook. Idempotent; called by every
+ * creator of a CONFIG_AUTHORITATIVE_READ computation — until() and refresh() —
+ * before its first read (same late-binding contract as the optimistic engine;
+ * the gating bit is only ever set by such a read, so the `!` call sites are
+ * safe once every setter installs, #3303). */ function installAuthoritativeRead() {
+    if (GlobalQueue.Xt === null) GlobalQueue.Xt = notifyAuthoritativeObservers;
+}
+
+/**
+ * Stale-reader term of the value selections below: a render effect reading a
+ * node some OTHER live transaction has staged sees the committed value. The
+ * commit is silent — the staging walk was the notification — so a reader
+ * that linked AFTER that walk (an effect created during the hold, a store
+ * key first read under it) would show the old value past the reveal: record
+ * it for the transaction's commit replay (the `_gatedSubs` contract lanes
+ * already use). An effect the transaction itself computed re-derives at its
+ * commit on its own (parked run, or the contested re-derive, #3322) and is
+ * not recorded — replaying it too would publish the frame twice.
+ *
+ * Flight twin (the pending-branch carve-out): the reader is served the
+ * node's committed, pre-flight value and now observes that flight — A15:
+ * async work observed by a reader settles as one unit with the writes that
+ * asked it — so it joins the transaction's reporters for the node. The
+ * reporter the transaction recorded when the flight started may be gone (a
+ * keyed remount disposed it, #3374); a completion check that found no live
+ * reporter committed the writes ahead of the answer, tearing the new
+ * reader's frame (`Count: 1` beside `Details: 0`). Joins an entry the
+ * transaction already holds; a flight nobody had observed yet has none (the
+ * reader is its first observer — a conditional that just revealed it, #3458)
+ * and is notified up the reader's own queue chain under that transaction,
+ * the one sanctioned registration site (INV-3): a collecting boundary above
+ * the reader consumes it as it would any pending, an unboundaried reader
+ * opens the entry — and the transaction, judged complete on its other
+ * flights, revealed the inputs beside the reader's pre-flight value
+ * otherwise (`Count: 1 | A: 1` beside `B: 0`). A staged signal or a settled
+ * node registers nothing. Every reporter dies with its reader
+ * (reporterBlocksSource: the read linked it as a dep). The node's own entry
+ * is the only one that can matter: a chain's intermediate memo is re-pulled
+ * by the read (updateIfNecessary's retry) and enters the transaction, so the
+ * reader holds through the normal path; a node with its own flight that is
+ * also pending on an upstream re-ask blocks through that flight until it
+ * lands, and its landing re-runs the reader into the normal path.
+ */
+/** The replay half of the stale-of-foreign clause (A15 / A26): a stale reader
+ * served the committed value because `txn` holds what it read re-runs at
+ * txn's commit, when the value it was denied becomes the frame — unless its
+ * own last value already came from that transaction. One registration for
+ * the node path (heldFromStale) and the store's backing paths, which have
+ * no node to carry the hold (heldFromReader, the adoption hold view). */ function recordStaleReplay(e, t) {
+    const n = t.Gt;
+    if (n == null || currentTransition(n) !== e) e.Tt.add(t);
+}
+
+/**
+ * The ownership relation (DESIGN-CONSOLIDATION §6, ruled 2026-09-17): is
+ * `hold` part of the running pass's world? A plain reader's world is the
+ * transaction it runs under, through merges. A lane reader's world is its
+ * lane AND the transition that owns the lane — the one asymmetry between a
+ * lane and a separate transaction (a lane sees what lands from its parent as
+ * its own; a separate transaction would wait for the parent to settle) —
+ * see `ownsLane` in lanes.ts, built on this. One relation for the
+ * stale-of-foreign clause (heldFromStale), the lane arm (readsHeldCommitted)
+ * and the store's backing holds (foreignHold); `serve` has no lane arm of
+ * its own, the lane's extra visibility lives here.
+ */ function ownsHold(e) {
+    return activeTransition !== null && currentTransition(e) === currentTransition(activeTransition);
+}
+
+function heldFromStale(e, t) {
+    const n = e.me;
+    if (n === null || ownsHold(n)) return false;
+    const i = currentTransition(n);
+    recordStaleReplay(i, t);
+    const l = i.ae.get(e);
+    if (l) l.add(t); else if (e.S & STATUS_PENDING) runInTransition(i, () => t.T.notify(t, STATUS_PENDING, STATUS_PENDING, e.o._));
+    return true;
+}
+
+/**
+ * A tracked computation about to be served a live transaction's staged value
+ * derives from that transaction's world, so it enters the transaction and
+ * this pass's result is held with it (A29, #3408). The write side (`setSignal`)
+ * and a stamped node's recompute already enter; a reader that only now
+ * starts reading the held node — a conditional memo whose branch flipped —
+ * was the gap: it published a value derived from the staged world into the
+ * mainline frame. Not for a probe (`isPending(() => x())` observes, it does
+ * not derive) and a no-op for the ambient batch (`_transition` null) or the
+ * transaction already active.
+ */
+/** The transaction a pass running OUTSIDE a flush was served a staged value
+ * from (A29, creation-time form). Inside a flush the pass enters through
+ * initTransition; outside one — a memo or effect created from mainline code
+ * while a hold is live — entering would leave `activeTransition` and the
+ * batch pointed at the transaction for the rest of the synchronous block,
+ * so an unrelated write made after the mount was held with someone else's
+ * action. The entry is the pass's alone: recompute stages the node into the
+ * transaction (born held) and mainline is never touched.
+ *
+ * A pass under a FRESH loading boundary takes this path inside a flush too
+ * (#3540). Born held is right for a plain memo or effect: published, its
+ * value would tear the frame. A boundary that has not revealed yet is the
+ * exception by definition — its job is to catch what is not ready under it
+ * rather than let it hold: the pass is staged into the transaction as
+ * above, and the boundary is told (recompute's born-held arm notifies it
+ * with a `NotReadyError` sourced at the node) so it shows its fallback now
+ * and reveals the staged result at the commit.
+ * Entering instead adopted the whole flush into the hold — a `Show` that
+ * flipped mainline to mount a `Loading` over a held value waited for the
+ * hold with it. A boundary that already shows content keeps the entering
+ * path: it has content to keep, and holds like any reader. */ let stagedEntry = null;
+
+/** Is `el` routed to a loading boundary that still shows its fallback — the
+ * nearest pending-collecting queue up its chain is uninitialized? (Measured
+ * 2026-09-18: relocating the walk behind a `GlobalQueue` slot installed by
+ * boundaries.ts saved ~8 B on the core floor and cost boundary-using apps
+ * 50–70 B — the indirection's tokens outweigh the walk. Kept inline.) */ function underFreshLoadingBoundary(e) {
+    for (let t = e.T; t !== null; t = t._parent) if (t.te & STATUS_PENDING) return !t.H;
+    return false;
+}
+
+function enterStagedRead(e, t = e.me) {
+    if (!t || t === activeTransition || pendingCheckActive) return;
+    // A companion (the latest() shadow, the isPending() verdict signal) is the
+    // engine's mirror of the flushed world — reading it, or being it, is an
+    // observation, not a derivation from the hold: latest(x) never enters x's
+    // transaction, and the shadow's own pass never enters either (it would
+    // flip activeTransition under the reader that pulled it). (`el` is null for
+    // a store backing served under a hold — no node, the transaction is the
+    // fold's.)
+        if (e?.o?.$n || context?.o?.$n) return;
+    // A bookkeeping read (`spectate`) is nobody's derivation: it compares or
+    // probes, and enters nothing — the boundary priming read of a born-held
+    // tree must not enter the creator's pass into the hold (#3540).
+        if (spectating) return;
+    // Verdict machinery (GlobalQueue._verdictPull: companion creation and the
+    // latest()/isPending() pulls — the latest() shadow is created before it is
+    // marked optimistic, so the bit alone cannot tell) and optimistic nodes
+    // (own lane posture, A31) read staged truth by design; they keep the
+    // entering path.
+    // (`context` is non-null here: every caller selected a value for a reader.)
+        const n = context;
+    const i = activeTransition === null && !globalQueue.$t;
+    // Verdict pulls are observations, not derivations: a latest() /
+    // isPending() call from mainline must never enter a transaction (it
+    // would capture the rest of the caller's synchronous block).
+        if (i && GlobalQueue.Tn) return;
+    if (n.oe & REACTIVE_RECOMPUTING_DEPS && !(n.C & CONFIG_OPTIMISTIC) && (stagedEntry === null || stagedEntry === t) && (i || !GlobalQueue.Tn && underFreshLoadingBoundary(n))) {
+        stagedEntry = t;
+        return;
+    }
+    globalQueue.initTransition(t);
+}
+
+/**
+ * Rule 1 (value selection), the full arm: does this reader see a STAGED
+ * node's COMMITTED value? One implementation of the rule the fast paths
+ * (readNodeFast, read's fast block) carry as their trivial ternary and that
+ * every slow site — read's tail, the store's backing selection, the lane and
+ * verdict arms — used to restate by hand (docs/DESIGN-CONSOLIDATION.md, move 3b). In order:
+ * - no reader at all (an untracked read) — the committed frame;
+ * - a reader under an optimistic lane the engine says reads committed
+ *   (laneReadsCommitted: another lane's hold, #3460);
+ * - nothing staged;
+ * - a children-forbidden reader (createTrackedEffect / onSettled: the frame,
+ *   never the graph — A32);
+ * - a stale reader (render effect) of a FOREIGN transaction's staged write —
+ *   committed, no entanglement (heldFromStale registers the replay; a node
+ *   born held has no committed frame to fall back to, `noCommitted`);
+ * - HELD truth (#3164, CONFIG_HELD_TRUTH) read by a LANE pass: staged
+ *   confirming truth — fold-staged onto an armed family, or entangle-stolen
+ *   by an awaited until() — is masked from lane passes only, owning
+ *   transaction or not. A lane applies its frame display-ahead at the park,
+ *   so a lane pass served the truth would paint the confirmation beside the
+ *   optimism it confirms (`saving=true` beside the saved row — the #3164
+ *   tear); it keeps committed and is re-run by the reveal's post-revert
+ *   wake. Every other deriving reader falls through to A29 below: the truth
+ *   is a staged value like any other, and the pass that derives from it is
+ *   held with it — including the retaining transaction's own passes, which
+ *   a superseded override already hands the truth (#3568: masking them
+ *   composed the landed `length` with rows still masked to committed).
+ *   latest() and authoritative readers (until()'s predicate) tunnel
+ *   through — the tunnel that keeps the hold deadlock-free.
+ * False means the reader derives from the staged value and enters its
+ * transaction (enterStagedRead, A29).
+ */ function readerSeesCommitted(e, t, n, i) {
+    return !!(!t || currentOptimisticLane !== null && GlobalQueue.On(e, n, t) || e._e === NOT_PENDING || t.C & CONFIG_CHILDREN_FORBIDDEN || stale && !i && heldFromStale(e, t) || e.C & CONFIG_HELD_TRUTH && currentOptimisticLane !== null && !latestReadActive && !(t.C & CONFIG_AUTHORITATIVE_READ));
+}
+
+/** A28 — set when a node is staged (queuePendingNode) or a held node rewritten
+ * (stashHeldRewrite) OUTSIDE a flush; cleared when the next flush begins. The
+ * read sites test this one module boolean instead of `globalQueue._running`:
+ * inside a flush it is false and the A28 arm costs nothing; outside, only a
+ * tick with unflushed writes pays the staged-node check. */ let unflushedStaged = false;
+
+function markUnflushedStaged() {
+    unflushedStaged = true;
+}
+
+/** A28 — a write becomes visible at flush. Outside a flush, a node holding an
+ * AMBIENT staged value (no transaction stamp) was written since the last
+ * flush: ambient staging commits at flush end, so nothing else leaves a node
+ * in this state; a stamped value is the flushed held world, which A28 says
+ * latest() serves. Inside a flush the rule does not apply (A28 (4): promoted
+ * within the round). Structural — no marker on the write path. */ function unflushed(e) {
+    return unflushedValue(e) !== NOT_PENDING;
+}
+
+/** The value an unflushed node serves — the committed value for an ambient
+ * write, the flushed staged value for a rewrite of a held node — or
+ * NOT_PENDING when nothing is unflushed. Exempt: owned-write nodes (A28 (4):
+ * a write issued inside a recompute is promoted at that recompute's end —
+ * boundary and loading machinery, until()'s internals, signals declared for
+ * in-computation writes) and engine companions (the isPending() verdict
+ * signal, the latest() shadow: the system's own writes, made at the source's
+ * write to mirror it, installing eagerly — A28, A8). */ function unflushedValue(e, t = e.ce) {
+    if (globalQueue.$t || e._e === NOT_PENDING || e.C & CONFIG_PROMOTED || e.o?.$n) return NOT_PENDING;
+    // Ambient, or adopted by a transaction before any flush carried the staging
+    // (CONFIG_ADOPTED_UNFLUSHED): nothing flushed is staged — the committed
+    // value answers (the caller's notion of committed: a store node's backing).
+    // A held node: unflushed only if rewritten since the last flush (stash).
+        if (e.me === null || e.C & CONFIG_ADOPTED_UNFLUSHED) return t;
+    return e.o === null ? NOT_PENDING : e.o.Ut;
+}
+
+/** Held nodes rewritten since the last flush (setSignal); the flush clears
+ * their stash — from then on latest() answers with the rewrite. */ const unflushedRewrites = [];
+
+/** Nodes written inside a creation-time recompute (CONFIG_PROMOTED). */ const promotedWrites = [];
+
+/** A28 (5): an optimistic write becomes the ACTIVE override at the flush that
+ * carries it. `_overrideTime` is stamped with `clock` at the write and `clock`
+ * advances after every flush, so "this tick, outside a flush" is unflushed. */ function unflushedOverride(e) {
+    // Companions are optimistic signals written by the engine (see unflushed).
+    return !globalQueue.$t && e.o?.Pt === clock && !e.o?.$n;
+}
+
+/** Active optimistic override on an armed node (an armed slot idles at
+ * NOT_PENDING; undefined = unarmed plain node). The writer's own channels —
+ * the draft, `in`/keys inside the setter — compose on this regardless of
+ * flush state. */ function hasActiveOverride(e) {
+    const t = e.o;
+    return t !== null && t.we !== undefined && t.we !== NOT_PENDING;
+}
+
+/** The override a READER sees: installed, and carried by a flush (A28 (5) —
+ * an optimistic write is a write; until its flush no reader sees it). One
+ * implementation for read()'s override arm, the verdict channels
+ * (latestRead, computePendingState) and the store's selection
+ * (docs/DESIGN-CONSOLIDATION.md, move 3b). */ function visibleOverride(e) {
+    return hasActiveOverride(e) && !unflushedOverride(e);
+}
+
+/** A derivation served the committed value because of an unflushed write
+ * (A28) must run again in the flush that carries it — the late-linker case
+ * (#3337's reason to defer the walk): it linked after the write walked. */ function markLateLinker(e) {
+    // The pass's own tail re-enqueues on this latch (recompute's finally) —
+    // a direct enqueue here would be wiped by the pass's flag reset.
+    e.oe |= REACTIVE_MISSED_WAKE;
+    return true;
+}
+
+/** Companion-bearing nodes written outside a flush (setSignal); the flush
+ * that carries their writes re-syncs their companions (A28). */ const unflushedCompanions = [];
+
+function resyncUnflushedCompanions() {
+    unflushedStaged = false;
+    // Length-guarded: the common flush has nothing here, and must allocate nothing.
+    // Length-guarded: the common flush has nothing here and allocates nothing.
+        if (unflushedRewrites.length !== 0) {
+        for (const e of unflushedRewrites) e.o.Ut = NOT_PENDING;
+        unflushedRewrites.length = 0;
+    }
+    if (promotedWrites.length !== 0) {
+        for (const e of promotedWrites) e.C &= ~CONFIG_PROMOTED;
+        promotedWrites.length = 0;
+    }
+    if (unflushedCompanions.length !== 0) {
+        for (const e of unflushedCompanions) GlobalQueue.Le(e, e._e !== NOT_PENDING ? e._e : e.ce);
+        unflushedCompanions.length = 0;
+    }
+}
+
+function readNodeFast(e) {
+    if (latestReadActive || pendingCheckActive || e.Se || e.Ne || e.o?.we !== undefined || e.o?.ct !== undefined || activeTransition !== null || currentOptimisticLane !== null || snapshotCaptureActive || 
+    // A28: a staged node read outside a flush may serve its committed value;
+    // that arm lives on the slow path so this body stays inlinable.
+    unflushedStaged && e._e !== NOT_PENDING || false) return READ_SLOW;
+    let t = context;
+    if (t?.Zt) t = t.Yt;
+    if (t && tracking) link(e, t);
+    // Children-forbidden readers (createTrackedEffect / onSettled callbacks) get
+    // committed visibility: like the effect half of createEffect and event
+    // handlers, effect-phase code never observes its own unsettled write — the
+    // write lands in the same flush's continuation (#3006).
+    // A stale reader (render effect) recomputing with no transaction active is
+    // mainline: a write staged by a live transaction (`_transition` stamped —
+    // ambient staging never is) stays masked until that transaction's reveal,
+    // the same rule the slow path applies (#3322). Without it a zombie
+    // recompute, or a commit-time re-derive of a contested effect, published
+    // another transaction's uncommitted value.
+        return !t || e._e === NOT_PENDING || t.C & CONFIG_CHILDREN_FORBIDDEN || stale && heldFromStale(e, t) ? e.ce : (enterStagedRead(e), 
+    e._e);
+}
+
+function read(e) {
+    // Handle latest() mode: read from _latestValueComputed
+    // Checked before isPending so that isPending(() => latest(x)) checks
+    // the _pendingSignal of _latestValueComputed (async in flight) rather
+    // than the original node (which stays "pending" while held in a transition).
+    if (latestReadActive) return GlobalQueue.An(e);
+    let t = context;
+    if (t?.Zt) t = t.Yt;
+    const n = e;
+    const i = e.Ne;
+    const l = i || e;
+    // Handle isPending() mode: collect pending state while preserving normal read semantics.
+    // Probe mode is suspended while preparing the node so nested reads during a
+    // recompute don't collect into the probe.
+        if (pendingCheckActive) {
+        GlobalQueue.Cn(e, t, l, i);
+    } else if (typeof n.Se === "function") {
+        prepareComputed(e, false);
+    }
+    if (!n.Se && l === e && e.o?.we === undefined && e.o?.ct === undefined && activeTransition === null && currentOptimisticLane === null && !snapshotCaptureActive && (!unflushedStaged || e._e === NOT_PENDING) && // A28, see readNodeFast
+    true) {
+        if (t && tracking) link(e, t);
+        // Committed visibility for children-forbidden readers and for stale
+        // readers of a foreign transaction's staged write — see readNodeFast.
+                return !t || e._e === NOT_PENDING || t.C & CONFIG_CHILDREN_FORBIDDEN || stale && heldFromStale(e, t) ? e.ce : (enterStagedRead(e), 
+        e._e);
+    }
+    if (t && tracking) {
+        link(e, t, pendingCheckActive);
+        if (l.Se) {
+            const n = queueFor(e);
+            if (l.ft >= n.st) {
+                markNode(t);
+                markHeap(n);
+                updateIfNecessary(l);
+            }
+            // Fresh-pull readers (awaitable refresh's waiter) recompute a dirty
+            // source inline even when the height gate defers to the flush: the
+            // waiter must park on the re-ask's window (or serve its sync answer),
+            // never read the PRE-re-ask value as settled. Self-guarded: a clean
+            // node no-ops and updateIfNecessary refuses disposed nodes (#2983) —
+            // a dead target serves its last value, which is already quiescent.
+             else if (t.C & CONFIG_FRESH_READ) updateIfNecessary(l);
+            const i = l.ft;
+            // parent check is shallow, might need to be recursive
+                        if (i >= t.ft && e._parent !== t) {
+                t.ft = i + 1;
+            }
+        }
+    }
+    if (l.S & STATUS_PENDING) {
+        // A reader landing on a pending node throws — the reveal that discovered
+        // the flight holds on it (A15: observed async settles as one unit) — with
+        // one carve-out: a stale (render) reader of a node pending in some OTHER
+        // transaction keeps showing the node's committed value, no entanglement
+        // (parallel transactions; the reader is recorded for that transaction's
+        // commit replay, `heldFromStale`). The carve-out is sound only while the
+        // committed value is coherent with the visible frame, i.e. while the
+        // flight's inputs are themselves unpublished: the stamp alone does not
+        // say so (it is pending-node bookkeeping), so it is refused when the
+        // inputs are on screen — committed by a batch that left the flight in
+        // the air (CONFIG_INPUTS_PUBLISHED, #3305) or revealed through a lane
+        // (optimistic / latest, #3334) — and the reader holds instead. An
+        // UNINITIALIZED node has no committed value to show and holds too
+        // (firewall-backed store reads always did; plain memos since the #3043
+        // port): falling through served `undefined` as if settled and stranded
+        // the reader outside both transactions, so it never re-ran.
+        if (t && !(stale && !(l.S & STATUS_UNINITIALIZED) && !(l.C & CONFIG_INPUTS_PUBLISHED) && !(l.C & CONFIG_HAS_LANE && GlobalQueue.Rn(l)) && heldFromStale(l, t))) {
+            // Per-lane suspension lives with the engine (a non-null lane implies it
+            // is installed): under a lane, only same-lane pending async without an
+            // active override throws — plus uninitialized sources regardless of
+            // lane (#3276); that check rides laneSuspends so floor bundles don't
+            // pay for it.
+            if (currentOptimisticLane === null || GlobalQueue.Gn(l)) {
+                // An untracked read of a pending node still re-runs its reader when
+                // the node settles — unless nobody is reading (`spectating`, #3528)
+                // or the reader is the node itself.
+                if (!tracking && !spectating && e !== t) link(e, t);
+                throw l.o?._;
+            }
+        } else if (!t && l.S & STATUS_UNINITIALIZED && 
+        // An armed DERIVED override shields the uninitialized state from a
+        // read with no reader identity (#3648 follow-up, A18 (d)): a memo
+        // whose first landing rode its lane (asyncWrite's lane branch →
+        // `laneOverride`) has no `_value` — the flag stays set until the lane's
+        // commit promotes the override — yet the override IS the displayed
+        // value of a pending node (A18 (c), the screen keeps it until the
+        // commit), exactly as an initialized node re-deriving under its override
+        // displays it. "Uninitialized must suspend" (#3276, A19 exception 1) was
+        // written for a node with NOTHING to show; this node has the override.
+        // The arm fires in the body-end window: the source override superseded
+        // (A18 body-end, #3427), the memo re-deriving on the plain channel and
+        // held, its own slot still armed. Fall through to serve()'s override
+        // arm. Readers WITH identity are untouched: a tracked lane reader keeps
+        // the override (A17), an off-lane one suspends (#3651, `overrideRead`).
+        !(hasActiveOverride(e) && e.C & CONFIG_DERIVED_OVERRIDE)) {
+            throw l.o?._;
+        }
+    }
+    // `owner` is the computed itself, or the firewall behind a store node —
+    // firewall-backed reads follow the same rules (memo parity, #2897 ruling):
+    // an errored derive throws for every late reader instead of silently
+    // serving node values (the seed, or last-good data after a failed refetch).
+        if (l.Se && l.S & STATUS_ERROR) {
+        // Only a genuine reactive re-read may retry an errored async source:
+        // - tracking: owned/tracked scope only (never events / `untrack` / effect side-effect phase)
+        // - !pendingCheckActive: an `isPending` probe observes the error, never refetches
+        // - owner._time < clock: only on a later cycle than the one the error was found
+        if (tracking && !pendingCheckActive && l.Ge < clock) {
+            recompute(l);
+            return read(e);
+        } else throw l.o?._;
+    }
+    if (snapshotCaptureActive && t && t.C & CONFIG_IN_SNAPSHOT_SCOPE) {
+        const n = e.o?.ct;
+        if (n !== undefined) {
+            const i = n === NO_SNAPSHOT ? undefined : n;
+            const l = e._e !== NOT_PENDING ? e._e : e.ce;
+            if (l !== i) t.oe |= REACTIVE_SNAPSHOT_STALE;
+            return i;
+        }
+    }
+    const u = serve(e, t, l, e.ce);
+    if (!t && l === e && typeof n.Se === "function" && e.C & CONFIG_AUTO_DISPOSE && !(l.S & STATUS_PENDING) && !e.u && 
+    // An untracked read served a visible override returned before this
+    // sweep registration when the arm was inline; keep that.
+    !visibleOverride(e)) {
+        // Deferred, not inline (#3078): an inline unobserved() here made untracked
+        // reads destructive — dispose on this read, full revival recompute on the
+        // next — so consecutive reads could answer differently with no write in
+        // between (the revival samples the ambient transition/lane context).
+        // The sweep at flush finalization re-validates and reclaims; schedule()
+        // guarantees that flush happens even if nothing else is queued.
+        dormantNodes.add(e);
+        schedule();
+    }
+    return u;
+}
+
+/**
+ * Rule 1, the one slow implementation (DESIGN-CONSOLIDATION move 3b, step
+ * 6c): the value a reader `c` (null = untracked, no pass) is served from
+ * `el`, whose committed value is `committed` — the node's own `_value` for
+ * a signal or memo, the BACKING for a store property node (single-home rule,
+ * O6: committed truth lives in the backing and a node's `_value` is never
+ * served for one). Called by read()'s slow tail and by the store's untracked
+ * node path (nodeValue); the fast paths (readNodeFast, read's fast block)
+ * keep their trivial ternary by design (perf, see the doc). Arms, in order:
+ * - the override (A17), routed through the engine for a tracked reader under
+ *   a lane or a supersession (A18), an authoritative reader marked instead;
+ * - the lane entanglement gate (committed, recorded for replay);
+ * - a node born held has nothing for an untracked reader (A19 exception 1);
+ * - an unflushed write serves committed and re-runs the reader in the
+ *   carrying flush (A28);
+ * - readerSeesCommitted, else the staged value and the transaction (A29).
+ */ function serve(e, t, n, i) {
+    if (hasActiveOverride(e)) {
+        // A17: the override IS the value for every reader — except an authoritative
+        // reader (until()'s predicate carries CONFIG_AUTHORITATIVE_READ): it must
+        // observe independently-arriving truth, and serving it the caller's own
+        // tentative write would trivially satisfy the predicate. The bit is checked
+        // on the reading computation itself, so a shared computed the predicate
+        // pulls recomputes as ITSELF (context = the memo, no bit) under the normal
+        // view. Fall through to normal value selection (staged `_pendingValue` is
+        // authoritative — optimism never lives there); the sticky mark makes the
+        // A17-silent "landing equals override" paths notify this node's subs so
+        // the reader re-runs when truth arrives.
+        // A28 (5): an optimistic write written this tick, outside a flush, is not
+        // yet the active override — fall through to the normal selection (the
+        // authoritative mark below still applies: until() must wake on landing).
+        if (!(t && t.C & CONFIG_AUTHORITATIVE_READ) && !unflushedOverride(e)) {
+            // A tracked read of an override is the engine's selection (a lane or a
+            // supersession implies the engine): a render effect OFF the override's
+            // held lane sees the committed value (#3460, lanes mirror transitions);
+            // a node whose own source answered with a DIFFERENT value hands a
+            // tracked reader the staged truth (A18 supersession, #3331). Untracked
+            // reads display the override.
+            if (t && e.C & (CONFIG_HAS_LANE | CONFIG_OVERRIDE_SUPERSEDED)) return GlobalQueue.Dn(e, t);
+            return unwrapOverride(e.o?.we);
+        }
+        e.C |= CONFIG_AUTHORITATIVE_OBSERVED;
+    }
+    // Entanglement gate: a reader recomputing under an optimistic lane that reads
+    // a pending mid-transition write sees the committed value. Projection-store
+    // manual writes use the firewall's manual-write flag to opt into this path.
+    // Async drivers are not under an optimistic lane and so bypass this, reading
+    // _pendingValue for correct fetching. The sub is recorded for replay at commit
+    // so it re-runs with the new committed view. (Gate details live with the
+    // engine — a non-null lane implies it is installed.)
+        if (currentOptimisticLane !== null && activeTransition !== null && t !== null && GlobalQueue.Pn(e, n, t)) {
+        return i;
+    }
+    // In optimistic lane context, return _value for optimistic/lane-assigned signals
+    // and for regular signals in stale mode (render effects). Non-stale readers (user
+    // effects) see _pendingValue so that latest() and direct reads stay consistent.
+    // (The lane-context clause lives with the engine.) Children-forbidden readers
+    // (createTrackedEffect / onSettled callbacks) get committed visibility — see
+    // readNodeFast (#3006).
+    // A node born held (recompute) has a staged value and no committed one: a
+    // stale reader cannot fall back to the committed frame, so it takes the
+    // staged value and enters like a tracked reader (A29); an untracked reader
+    // has nothing to serve and holds (A19 exception 1) — a bookkeeping read
+    // (`spectate`) likewise: it derives nothing, so it cannot enter, and a
+    // held-only value is simply not ready to it.
+        const l = e._e !== NOT_PENDING && (e.S & STATUS_UNINITIALIZED) !== 0;
+    if (l && (!t || spectating)) throw new NotReadyError(null);
+    const u = t && unflushedStaged ? unflushedValue(e, i) : NOT_PENDING;
+    if (u !== NOT_PENDING) {
+        markLateLinker(t);
+        if (pendingCheckActive) GlobalQueue.Hn(e, u);
+        return u;
+    }
+    const o = readerSeesCommitted(e, t, n, l) ? i : (enterStagedRead(e), e._e);
+    // Record that this isPending() probe observed the fresh pending value, so
+    // the probe doesn't pair "pending" with the new value (#2831).
+        if (pendingCheckActive) GlobalQueue.Hn(e, o);
+    return o;
+}
+
+/** A28 — a rewrite of a HELD node outside a flush keeps the staged value the
+ * last flush left, for latest()/verdicts, until the next flush carries it
+ * (stash, cleared at that flush). Cold: only stamped nodes reach here. */ function stashHeldRewrite(e) {
+    if (globalQueue.$t) return;
+    const t = ext(e);
+    if (t.Ut === NOT_PENDING) {
+        t.Ut = e._e;
+        unflushedRewrites.push(e);
+        unflushedStaged = true;
+    }
+}
+
+/** A28 (4) — written inside a recompute that runs OUTSIDE a flush (a
+ * creation-time compute): promoted at that pass's end, visible to the rest of
+ * the block. Cold: only contextual writes reach here. */ function notePromotedWrite(e) {
+    if (globalQueue.$t || e.C & CONFIG_PROMOTED) return;
+    e.C |= CONFIG_PROMOTED;
+    promotedWrites.push(e);
+}
+
+/** Cold half of setSignal's snapshot arm (see there): first write during
+ * capture to a plain signal without a live snapshot records the pre-write
+ * value. Only plain user signals qualify. Computeds reach setSignal from an
+ * async landing (asyncWrite), and a value arriving from async during the pass
+ * REVEALS — the creation-time arm skips pending computeds for the same
+ * reason. Firewall leaves belong to a projection whose compute is the tree's
+ * own work, captured (or deliberately not) at creation. */ function captureWriteSnapshot(e, t) {
+    if (e.C & CONFIG_NO_SNAPSHOT || e.Se !== undefined || e.Ne || e.o?.ct !== undefined) return;
+    ext(e).ct = t === undefined ? NO_SNAPSHOT : t;
+    e.C |= CONFIG_HAS_SNAPSHOT;
+    snapshotSources.add(e);
+}
+
+function setSignal(e, t) {
+    // A write to a held node is a second proposal on a contested node (A34, #3494):
+    // the writer's tick reveals with the hold, the same value or another. Inside
+    // a flush the round enters now; from mainline the entry waits for the next
+    // flush's start (batchJoins) — entering here left activeTransition set for
+    // the rest of the caller's block, so a memo created after the write became
+    // the transaction's instead of mainline's (A28, A29). Before the equality
+    // gate below: repeating the held value proposes it too — and that repeat
+    // leaves through the gate, so the join schedules its own flush here; left
+    // for the next flush to find, it adopted an unrelated tick (#3519 review).
+    // A held DERIVATION is not a proposal (A34 amendment, #3612): the user
+    // setters of derived nodes (setMemo, the derived store's setter) classify
+    // the write before it reaches here — see heldDerivation — and re-derive
+    // under the hold instead of masking; the join stands either way.
+    if (e.me && activeTransition !== e.me) {
+        if (globalQueue.$t) globalQueue.initTransition(e.me); else {
+            batchJoins.push(e.me);
+ // dupes: a bare return in initTransition
+                        schedule();
+        }
+    }
+    // The optimistic write path lives with the engine: only optimisticSignal /
+    // optimisticComputed callers and optimistic store nodes carry an
+    // _overrideValue slot (flagged by CONFIG_OPTIMISTIC — a masked read of the
+    // always-present config instead of a missing-property probe), and every
+    // module that installs one installs the engine first.
+        if (e.C & CONFIG_OPTIMISTIC) {
+        if (!projectionWriteActive) return GlobalQueue.Ln(e, t);
+        // An authoritative store landing on an override-covered node: the store
+        // twin of asyncWrite's override branch, decided by the engine (#3331).
+                const n = e.o?.we;
+        if (n !== undefined && n !== NOT_PENDING) return GlobalQueue.Vn(e, t);
+    }
+    const n = e._e === NOT_PENDING ? e.ce : e._e;
+    if (typeof t === "function") t = t(n);
+    // Uninitialized check first: the first commit has no previous value, so the
+    // user comparator must not run against `undefined` (matches recompute).
+        const i = !!(e.S & STATUS_UNINITIALIZED) || !e.Fe || !e.Fe(n, t);
+    if (!i) return t;
+    // Attribution hook: this committed write is where a re-run chain begins.
+        if (attrHooks !== null) attrHooks.write(e, n, t);
+    // A write during hydration's snapshot capture to a source that has no
+    // snapshot — created BEFORE capture began (module-level state: an identity
+    // minted from onSettled in the pass, a preference read from storage) —
+    // captures the pre-write value now, so the write is held like any other:
+    // in-scope readers keep serving what the server rendered with and replay
+    // at release. Left uncaptured, the write cascades live through a claim
+    // pass whose DOM writes are skipped, and a component rendered later in the
+    // pass reads a value the server never had. Store leaves written here are
+    // plain signals and qualify the same way.
+        if (snapshotCaptureActive) captureWriteSnapshot(e, n);
+    const l = e._e !== NOT_PENDING;
+    if (!l) queuePendingNode(e);
+    // A28 arms, gated on the loads the write already pays for (a plain ambient
+    // rewrite outside a flush — the write-loop shape — costs `_transition` and
+    // `context` here and nothing else; the arms themselves are cold helpers so
+    // setSignal stays within every setter's inlining budget, ~300 B bytecode).
+     else if (e.me !== null) stashHeldRewrite(e);
+    e._e = t;
+    if (context !== null) notePromotedWrite(e);
+    // syncCompanions only pokes _pendingSignal/_latestValueComputed — with
+    // neither companion present the call is a guaranteed no-op (companions are
+    // only ever created, never removed, and creating one installs the hook and
+    // sets CONFIG_HAS_COMPANIONS — one masked read replaces two optional-field
+    // probes on every write).
+        if (e.C & CONFIG_HAS_COMPANIONS && GlobalQueue.Le !== null) {
+        GlobalQueue.Le(e, t);
+        // A28 (2): the verdict computed here is the PRE-flush one (an unflushed
+        // write is not pending); the flush that carries the write re-syncs so
+        // the companions mirror the flushed world. Only companion nodes pay.
+                if (!globalQueue.$t) unflushedCompanions.push(e);
+    }
+    // _time is a computed-only slot (§12e): writing it on a signal would fork
+    // the lean shape. Every read site is computed-typed.
+        if (e.Se !== undefined) e.Ge = clock;
+    // Staged-rewrite fast path (§12d): a re-write to a node whose subscribers
+    // were already walked — and where nothing has recomputed or linked since
+    // (epoch) — re-stages the value and stops. The walk is idempotent (subs
+    // marked, heap entries flag-guarded, effects queued once); lane and reask
+    // contexts change what a walk MEANS, so they always walk.
+        if (l && e.Lt === notifyEpoch && currentOptimisticLane === null && !reaskArmed) return t;
+    insertSubs(e);
+    schedule();
+    return t;
+}
+
+/**
+ * Suppresses automatic recomputation of `el` until the scheduler drains. Used
+ * when a manual write should win over dependency changes queued in the same
+ * tick. The MANUAL_WRITE flag is cleared by the pending-node drain; projection
+ * computeds don't commit values, but they still need the same end-of-tick
+ * cleanup point.
+ */ function suppressComputedRecompute(e) {
+    deleteFromHeap(e, queueFor(e));
+    if (!(e.oe & REACTIVE_MANUAL_WRITE) && e._e === NOT_PENDING) {
+        queuePendingNode(e);
+        schedule();
+    }
+    e.oe = e.oe & -4 | REACTIVE_MANUAL_WRITE;
+    e.Un = clock;
+}
+
+/** A34 amendment (#3612) — is `el`'s staging a DERIVATION another transaction
+ * holds: stamped by a transaction that is not the writer's, and not a manual
+ * proposal (the mask, on the node or — a store leaf — its firewall)? #2692's
+ * "manual write wins" is a rule for one synchronous frame; across a hold the
+ * held pass result is nobody's proposal, and a user setter reaching it
+ * composes on the committed frame it was written against and becomes `prev`
+ * for the transaction's re-derivation (rederiveHeld) instead of replacing it.
+ * Writes made under the holding transaction masked the node and keep
+ * A34(1)'s last-write-wins. Only the user setters ask; an async landing or a
+ * companion writing a stamped node is the transaction's own work. */ function heldDerivation(e) {
+    return e.me !== null && activeTransition !== e.me && !((e.Ne || e).oe & REACTIVE_MANUAL_WRITE);
+}
+
+/** The held-derivation write's second half: the node re-derives under its
+ * hold with the written staging as the pass's `prev`. Nothing is masked. The
+ * write took the A34 join (setSignal), which scheduled the flush that drains
+ * this; recompute re-enters the stamp. */ function rederiveHeld(e) {
+    e.oe |= REACTIVE_DIRTY;
+ // over CHECK: the heap visit recomputes, not re-checks
+        enqueueSub(e);
+}
+
+/**
+ * User-facing setter for the memo form of `createSignal(fn)`. Behaves like
+ * `setSignal`, but also cancels any pending recompute of the memo so the
+ * manual value wins over a value that would otherwise be produced by an
+ * upstream change in the same tick. Across a hold the write is not a
+ * proposal: a memo another transaction holds as a pass result composes on the
+ * committed value and re-derives under the hold with the write as `prev`
+ * (A34 amendment, #3612; heldDerivation).
+ */ function setMemo(e, t) {
+    const n = heldDerivation(e);
+    if (n && typeof t === "function") t = t(e.ce);
+    const i = setSignal(e, t);
+    n ? rederiveHeld(e) : suppressComputedRecompute(e);
+    return i;
+}
+
+/**
+ * Executes `fn` with the given `owner` set as the current owner. Any reactive
+ * primitives (`createSignal`, `createMemo`, `createEffect`, `onCleanup`,
+ * `cleanup`, etc.) created inside `fn` are attached to that owner, so they
+ * are disposed when the owner is disposed.
+ *
+ * The classic pattern: capture the current owner with `getOwner()` inside a
+ * component, then re-enter it from a callback (event handler, async resolve,
+ * setTimeout) so disposables created in the callback get cleaned up with the
+ * component.
+ *
+ * @example
+ * ```ts
+ * function delayed<T>(ms: number, fn: () => T) {
+ *   const owner = getOwner();
+ *   setTimeout(() => runWithOwner(owner, fn), ms);
+ * }
+ * ```
+ */ function runWithOwner(e, t) {
+    const n = context;
+    const i = tracking;
+    context = e;
+    tracking = false;
+    try {
+        return t();
+    } finally {
+        context = n;
+        tracking = i;
+    }
+}
+
+function staleValues(e, t = true) {
+    const n = stale;
+    stale = t;
+    try {
+        return e();
+    } finally {
+        stale = n;
+    }
+}
+
+/**
+ * Core marking half of `refresh()` (the public wrapper lives in signals.ts —
+ * it validates the target, marks through here, then builds the quiescence
+ * promise on the resolve()/until() effect machinery). Flags the node's next
+ * recompute as a quiet re-ask and schedules it; no-ops for non-derived or
+ * disposed targets and for same-tick manual writes.
+ */ function markRefresh(e) {
+    if (typeof e.Se === "function" && !(e.oe & REACTIVE_DISPOSED)) {
+        if (e.oe & REACTIVE_MANUAL_WRITE) {
+            // A manual write in the CURRENT tick wins over the refresh (#2692).
+            // A mask stamped in an earlier tick only survives because a
+            // transaction (action) is holding the pending drain open; there the
+            // refresh is a later, explicit re-ask and lifts the mask — otherwise
+            // any setStore early in an action silently swallows every refresh()
+            // for the rest of the transaction (#3026).
+            if (e.Un === clock) return;
+            e.oe &= ~REACTIVE_MANUAL_WRITE;
+            // The lift falls through to the re-ask classification below. The held
+            // write's value change already rides the transaction; the refetch it
+            // asks for is the same question with unchanged inputs. Skipping the
+            // mark here classified that refetch as a NEW question, which pends
+            // every leaf (3.1) — an action doing setStore + yield + refresh(store)
+            // lit up every sibling row, and affects() could not narrow it (a mark
+            // only turns pending on). Same-question motion stays silent (3.4);
+            // the written slot and any declared mark carry the pending instead.
+                }
+        // A refresh with no value-change dirt already queued is a re-ask of the
+        // same question: mark it so the recompute classifies any resulting
+        // pending window as quiet (not pending). If the node is already dirty
+        // from a real input change, the question changed — don't mark.
+        // REACTIVE_IN_HEAP counts as dirt: insertSubs schedules subscribers by
+        // heap insertion alone (no DIRTY/CHECK flag), so a same-batch value
+        // change followed by refresh() must not be laundered into a quiet re-ask.
+                if (!(e.oe & (REACTIVE_DIRTY | REACTIVE_CHECK | REACTIVE_IN_HEAP))) {
+            e.oe |= REACTIVE_REASK;
+            armReaskClear();
+        }
+        e.oe = e.oe & ~REACTIVE_CHECK | REACTIVE_DIRTY;
+        // A refresh() self-invalidation is a root cause too — the target's next
+        // run has no changed dep to point at, so it points here instead.
+                if (attrHooks !== null) attrHooks.refreshed(e);
+        insertIntoHeap(e, queueFor(e));
+        schedule();
+    }
+}
+
+export { READ_SLOW, clearSnapshots, computed, context, createEffectNode, currentOptimisticLane, effectStatusNotify, enterStagedRead, ext, hasActiveOverride, heldDerivation, installAuthoritativeRead, isEqual, latestReadActive, markLateLinker, markRefresh, markSnapshotScope, markUnflushedStaged, notifyAuthoritativeObservers, notifyOnLane, optimisticComputed, optimisticSignal, ownsHold, pendingCheckActive, prepareComputed, read, readNodeFast, readerSeesCommitted, recompute, recordStaleReplay, rederiveHeld, releaseSnapshotScope, resyncUnflushedCompanions, runWithOwner, serve, setContextInternal, setEffectStatusNotify, setLatestReadActive, setMemo, setPendingCheckActive, setSignal, setSlotUnobserved, setSnapshotCapture, signal, slotSignal, slotUnobservedHook, snapshotCaptureActive, snapshotSources, spectate, spectating, stale, staleValues, statusNotifierOf, suppressComputedRecompute, tracking, unflushed, unflushedCompanions, unflushedOverride, unflushedStaged, unflushedValue, unlinkFirewallChild, untrack, visibleOverride };