@solidjs/signals 2.0.0-rc.2 → 2.0.0-rc.20

package/dist/types/core/attribution.d.ts

@@ -1,6 +1,7 @@
-import type { Computed } from "./types.js";
+import type { Computed, Signal } from "./types.js";
 /**
- * Dev-mode "why did this run" attribution.
+ * "Why did this run" attribution — the engine behind
+ * `@solidjs/signals/attribution`.
  *
  * The runtime already knows the full dependency set of every scope; this
  * module surfaces it. Every value commit stamps its node with a ChangeRecord
@@ -15,34 +16,111 @@ import type { Computed } from "./types.js";
  *
  * This module is the attribution ENGINE: all semantics live here, and it is
  * decoupled from the core. `enable()` installs it into the core's narrow
- * dev-only hook points (attribution-hooks.ts); core's only obligation is to
- * call those hooks with true facts. Disabled cost is one null check per hook
- * site; prod builds fold the sites out entirely. The same hook surface is the
- * intended substrate for external consumers (devtools) — one mechanism, two
- * front-ends.
+ * observe-tier hook points (attribution-hooks.ts); core's only obligation is
+ * to call those hooks with true facts. Disabled cost is one null check per
+ * hook site; prod builds fold the sites out entirely. Nothing in the core
+ * imports this module — it is reachable only through the package's
+ * `./attribution` entry, so an observe build that never imports it never
+ * ships it. The same hook surface is the intended substrate for external
+ * consumers (devtools) — one mechanism, two front-ends.
  */
 export type ChangeKind = "write" | "derived" | "async" | "refresh";
+/**
+ * Provenance of a root change: the imperative frame that performed it.
+ *
+ * - `interaction` — a user event handler (the web runtime marks dispatch via
+ *   `withInteraction`). `name` is the event type, `target` the element hit
+ *   (`button#next "Next →"` — the quoted text as `AttributionOptions.values`
+ *   allows), `at` the dispatch time on the `performance.now()` clock — the
+ *   base every feedback-latency number is measured from.
+ * - `effect` — an effect callback (`name` = the effect's name; `run` = the
+ *   compute run whose effect phase performed the write, when that run was
+ *   recorded — so a write can be joined to the re-run that produced it).
+ * - `action` — a step of an `action()` generator (`name` = the generator's
+ *   name, when it has one). Writes after an `await` (not a `yield`) run in a
+ *   bare microtask and stamp `external` — the documented escape.
+ * - `async` — an async landing (`name` = the node whose flight landed).
+ * - `navigation` — a router's navigation, declared via `withOrigin` around
+ *   the location write (`name` = the matched route pattern `/users/:id`;
+ *   `to`/`from` the concrete paths; `params` what the pattern bound; `at`
+ *   when it was requested). The router-agnostic seam: any router that wraps
+ *   its write gets navigations named by route in every hold, re-run and
+ *   verdict, with no per-router knowledge anywhere in the engine.
+ * - `external` — none of the above: timers, sockets, promise callbacks, setup.
+ *
+ * `interaction` on a non-interaction frame is the user event the frame runs
+ * under — an action started by a click, an effect whose run was caused by a
+ * click's write, a landing whose flight a click started, a navigation a link
+ * click performed. It is what lets every downstream cost be keyed by the
+ * interaction that paid for it.
+ */
+export interface ChangeOrigin {
+    kind: "interaction" | "effect" | "action" | "async" | "navigation" | "external";
+    name?: string;
+    target?: string;
+    /**
+     * When the frame opened (`performance.now()` clock). Always set for
+     * `interaction` and `navigation`; set on an `effect` frame only while an
+     * `effect` record listener exists (the callback's timed start).
+     */
+    at?: number;
+    interaction?: ChangeOrigin;
+    /** `effect` only: the `RerunEvent.run` of the compute run this callback belongs to. */
+    run?: number;
+    /** `navigation` only: concrete destination and departure paths, and the bound params. */
+    to?: string;
+    from?: string;
+    params?: Readonly<Record<string, string | undefined>>;
+}
 export interface ChangeRecord {
     /** Global monotonic change sequence — orders causes across the app. */
     seq: number;
     kind: ChangeKind;
     name: string;
-    /** Short previews of the value transition (writes only). */
+    /**
+     * Identity of the node that changed — the signal written, the memo whose
+     * value changed — in the same id space as `RerunEvent.nodeId`, so a
+     * derived cause joins the run that produced it and repeated writes to
+     * one signal join each other where `name` alone would merge every
+     * unnamed `signal`. Stamped on every record the engine makes; optional
+     * for a record built elsewhere (a `HeldWrite`, a deserialized artifact).
+     */
+    nodeId?: number;
+    /**
+     * Short previews of the value transition (writes only) — present under
+     * `AttributionOptions.values: "full"`, never carried otherwise.
+     */
     prev?: string;
     value?: string;
     /** First user frames of the triggering write's stack (opt-in). */
     stack?: string[];
     /** For derived changes: the upstream changes that produced this one. */
     causes?: ChangeRecord[];
+    /** Root changes only: who performed the write. */
+    origin?: ChangeOrigin;
+    /** Root changes only: when the write was stamped (`performance.now()` clock). */
+    at?: number;
 }
 export interface RerunEvent {
     /** Global monotonic run sequence. */
     run: number;
+    /** When the run started (`performance.now()` clock). */
+    at: number;
     /** How many times this node has re-run since attribution was enabled. */
     nodeRuns: number;
     nodeKind: "effect" | "memo";
     nodeName: string;
-    node: Computed<any>;
+    /**
+     * Identity of the scope that ran, stable for the node's lifetime within
+     * the process: every run of one memo/effect carries the same `nodeId`, so
+     * runs join to a scope after the record has left the process (where
+     * `nodeName` alone would merge every unnamed `effect`). The engine's own
+     * per-node id, also what `ChangeOrigin.run` and the cycle/relay checks
+     * key on; not meaningful across processes or sessions. In-process
+     * consumers get the live node as the `live` argument beside the record
+     * (`OBSERVE.records.subscribe("rerun", (event, node) => …)`).
+     */
+    nodeId: number;
     /**
      * The deps that changed since this node's previous run. Empty means the
      * re-run was not triggered by a tracked value change (creation-adjacent
@@ -60,35 +138,246 @@ export interface RerunEvent {
     /** Wall time of this run including nested recomputes (ms). */
     totalMs: number;
     /**
-     * Whether the run committed a changed value. A PLAIN memo run with
+     * Whether the run produced a changed value. A PLAIN memo run with
      * `changed: false` was pure waste — the equality cutoff stopped it from
-     * notifying anyone; an effect run with `changed: false` computed without
-     * firing its effect phase. Summed as `wastedMs` in costs() (plain,
-     * non-held runs only — see `phase`).
+     * notifying anyone. Effects run with `_equals: false` in core (their
+     * effect phase re-fires on every recompute), so the engine derives this
+     * fact itself: an effect run whose compute output is identical to the
+     * previous run's reports `changed: false` — the phase re-fired with the
+     * same input, pure waste. Side-effect-only computes (`undefined` output —
+     * effects, projections, memos that work by writing) report `true`:
+     * identity of `undefined` proves nothing about their work.
+     * Summed as `wastedMs` in costs() (plain, non-held runs only — see
+     * `phase`).
      */
     changed: boolean;
     /**
      * Which posture this run executed under. "optimistic" = under an
-     * optimistic lane (overlay recompute); "transition" = a transition was
-     * active or owns the node (the run may be replayed/settled later);
-     * "plain" = an ordinary committed run. Overlay runs are real work (they
-     * count toward time budgets) but are never blamed as waste, and costs()
-     * reports their time separately as `overlayMs`.
+     * optimistic lane (overlay recompute); "held" = a hold was open or owns
+     * the node (the run may be replayed/settled later); "plain" = an ordinary
+     * committed run. Overlay runs are real work (they count toward time
+     * budgets) but are never blamed as waste, and costs() reports their time
+     * separately as `overlayMs`.
      */
-    phase: "plain" | "transition" | "optimistic";
+    phase: "plain" | "held" | "optimistic";
     /**
-     * The changed value was held in `_pendingValue` (a transition hold) rather
-     * than committed directly; its reveal happens on the transition's own
-     * schedule. Held runs are excluded from waste accounting.
+     * The changed value was parked in `_pendingValue` (held) rather than
+     * committed directly; its reveal happens on the hold's own schedule. Held
+     * runs are excluded from waste accounting.
      */
     held: boolean;
+    /** The user interaction this run traces back to through its causes, if any. */
+    interaction?: ChangeOrigin;
+}
+/**
+ * A computation's creation run — the first run, the one with no causes to
+ * explain (a `RerunEvent` is every run after it). The same measurements as a
+ * re-run, less the causal fields a first run cannot have, and `interaction`
+ * inherited from whatever built the node: the enclosing recompute (a
+ * parent's fn creating children) or the handler/effect frame at the top of
+ * the stack. Creation time is already charged to the interaction record's
+ * `created`; this is the per-node face of that sum.
+ */
+export interface CreateEvent {
+    /** When the run started (`performance.now()` clock). */
+    at: number;
+    nodeKind: "effect" | "memo";
+    nodeName: string;
+    /** Same id space as `RerunEvent.nodeId` — the node's later re-runs join here. */
+    nodeId: number;
+    /** Dependency count after this run. */
+    depCount: number;
+    /** Wall time of this run excluding nested recomputes (ms) — children created inside report their own. */
+    selfMs: number;
+    /** Wall time of this run including nested recomputes (ms). */
+    totalMs: number;
+    /** Posture the run executed under — see `RerunEvent.phase`. */
+    phase: "plain" | "held" | "optimistic";
+    /** The value was parked in `_pendingValue` rather than committed — see `RerunEvent.held`. */
+    held: boolean;
+    /** The interaction whose handler or flush built this node, if any. */
+    interaction?: ChangeOrigin;
+}
+/**
+ * One run of an effect's imperative half — the callback that touches the DOM
+ * or the outside world — timed from entry to exit, cleanup included, whether
+ * or not it threw. The compute half is the `RerunEvent`/`CreateEvent` with
+ * the same `nodeId` that preceded it in the flush; `run` joins the two for a
+ * re-run and is absent for a creation's first callback.
+ */
+export interface EffectRunEvent {
+    /** When the callback started (`performance.now()` clock). */
+    at: number;
+    /** Wall time of the callback, nested runs included (ms). */
+    durationMs: number;
+    nodeId: number;
+    nodeName: string;
+    /** `RerunEvent.run` of the compute run whose effect phase this is; absent for a creation. */
+    run?: number;
+    /** The interaction the preceding compute run traced to, if any. */
+    interaction?: ChangeOrigin;
+}
+/**
+ * One `flush()` drain: from the scheduler picking up scheduled work until
+ * every batch it processed has committed (effects ran) or been parked in a
+ * held transition. The unit React's "Scheduler" track paints; the engine's
+ * `flushEnd` settles interactions and navigations on the same instant.
+ * Counts cover the runs the engine recorded inside the drain (excluded
+ * scopes not counted).
+ */
+export interface FlushEvent {
+    /** When the drain started (`performance.now()` clock). */
+    at: number;
+    /** Wall time of the drain (ms). */
+    durationMs: number;
+    /** Re-runs recorded during the drain. */
+    runs: number;
+    /** Creation runs during the drain. */
+    created: number;
+    /** A transition was judged incomplete during the drain — some of its writes stayed staged. */
+    held: boolean;
+    /**
+     * The interaction every recorded run of the drain traced to, when there
+     * was exactly one; absent when none did, or when runs for several
+     * interactions shared the drain.
+     */
+    interaction?: ChangeOrigin;
+}
+/**
+ * One async flight — a promise or async iterable an async computation
+ * registered — from its origin (the earliest the engine knows: a
+ * `markFlight` preload mark, or first sight at registration) to the moment
+ * it landed or was superseded by the node's next flight (`abandoned` — the
+ * answer will be discarded). A flight whose node is disposed mid-air
+ * produces no record.
+ */
+export interface FlightEvent {
+    nodeId: number;
+    nodeName: string;
+    /** Owner-chain labels of the async node, root first, when the node is owned. */
+    ownerPath?: string[];
+    /** When the flight started (`performance.now()` clock). */
+    at: number;
+    /** Wall time in the air (ms). */
+    durationMs: number;
+    outcome: "landed" | "abandoned";
+    /** The interaction whose write started the flight, if any. */
+    interaction?: ChangeOrigin;
 }
+/**
+ * One showing of a loading boundary's fallback, delivered when it stops
+ * showing. `ownerPath` names the boundary by its subtree's owner chain
+ * (`["<App>", "<Feed>"]`); absent when the subtree never reported a path.
+ * The fold in `feedback().fallbacks` is this record summed per boundary,
+ * with the flash verdict.
+ */
+export interface FallbackEvent {
+    ownerPath?: string[];
+    /** When the fallback appeared (`performance.now()` clock). */
+    at: number;
+    /** How long it stayed (ms). */
+    shownMs: number;
+    /** The interaction whose work the boundary was waiting on, if the engine could tell. */
+    interaction?: ChangeOrigin;
+}
+/**
+ * The live graph's size at a navigation's settle — the moment an app has
+ * finished moving between two screens, so a count that climbs visit after
+ * visit is a root or a subscription the previous screen left behind.
+ * Delivered on `OBSERVE.records.subscribe("graph", …)` per settled
+ * navigation; a walk of the owner tree from the registered top-level roots,
+ * made only when something listens or `graphGrowth` is on.
+ */
+export interface GraphEvent extends GraphSize {
+    /** When the navigation settled (`performance.now()` clock). */
+    at: number;
+    /** The route pattern the navigation matched (`name`), else its `to`. */
+    route?: string;
+    /** The navigation this count belongs to. */
+    navigation: NavigationEvent;
+}
+/**
+ * The live reactive graph's size, as `graphSize()` counts it: the owner tree
+ * from the registered top-level roots, then every computation and signal
+ * reachable from it through dependencies and subscriptions — which is how
+ * an ownerless effect (created with no owner, kept alive by its sources)
+ * is found — and the edges between them.
+ */
+export interface GraphSize {
+    /** Top-level roots alive (`render()`'s, module-scope `createRoot()`s, panels). */
+    roots: number;
+    /** Owners in the tree: roots, component owners, owned computations. */
+    owners: number;
+    /** Computations (memos, effects, boundaries), owned or reached through a subscription. */
+    computations: number;
+    /** Signals reached through a computation's dependencies. */
+    signals: number;
+    /** Dependency links — each computation's sources, counted once. */
+    edges: number;
+}
+/**
+ * What user data the engine's records carry — see `AttributionOptions.values`.
+ * Ordered: `"full"` carries the most, `"none"` the least.
+ */
+export type AttributionValues = "full" | "labels" | "none";
 export interface AttributionOptions {
     /** Pretty-print each re-run to the console (default true). */
     log?: boolean;
+    /**
+     * What user data records carry. Default per build tier: `"full"` in dev
+     * builds, `"none"` in observe builds — a production observability
+     * artifact carries no user data unless a holder asks for it. Records name things —
+     * owner paths, `name` options, store paths, route patterns — and are
+     * otherwise numbers, kinds and outcomes; this option governs the fields
+     * that quote application data: the value previews on a root write
+     * (`ChangeRecord.prev`/`value`, and the `HeldWrite.prev`/`value` copied
+     * from them), the text of the element an interaction hit
+     * (`ChangeOrigin.target` / `InteractionEvent.target` — the `"Next →"` in
+     * `button#next "Next →"`), and every sentence built from those (the
+     * console log, `formatRerun`, `formatOrigin`, the SILENT_HOLD / LONG_HOLD
+     * / STACKED_HOLDS verdicts naming the interaction, OPTIMISTIC_REVERTED's
+     * shown and settled values). Applied where the record is BUILT, so
+     * nothing downstream — a ring buffer, a fold, a listener, an exporter —
+     * ever holds what the level excludes.
+     *
+     * - `"full"` — previews of the written values (strings quoted and cut at
+     *   40 characters, numbers and booleans verbatim, everything else a type
+     *   tag such as `Array(12)`) and the element's text: what makes dev
+     *   output readable.
+     * - `"labels"` — no value previews; element text kept only on a `button`
+     *   or an `a` (the control's caption — what the person pressed is the
+     *   point of an interaction record) and dropped for anything else (the
+     *   text of a `div` or a `td` is content).
+     * - `"none"` — no value previews, no element text: names, numbers, kinds
+     *   and outcomes only. What an observe-tier holder that carries records
+     *   out of the process (an APM adapter) should pass.
+     *
+     * Across holds the LEAST permissive level wins — the one key that
+     * combines the other way from the rest, where a hold can only ask for
+     * more: a holder that must not see user data is not overruled by one
+     * that wants it. A holder that names no level asks for the tier's
+     * default, so in an observe build it tightens to `"none"` beside anyone;
+     * a single holder passing `"full"` there gets `"full"`, and an explicit
+     * `"full"` never loosens what another holder demanded. Governs what is
+     * built from the moment the level is in effect; records already in a
+     * ring buffer keep what they carried. A
+     * navigation's concrete paths and params (`ChangeOrigin.to`/`from`/
+     * `params`) are the router's description and are not governed here.
+     */
+    values?: AttributionValues;
+    /**
+     * Run the cost checks — the thresholded findings over the engine's own
+     * accounting: `hotRuns`, `hotTime`, `wideDeps`, `unstableMemos`,
+     * `fanOut`, `wastedRecompute` (default true). `false` turns all six off at once, whatever
+     * their individual settings, so a consumer that wants records only (an
+     * exporter, a profiler track) pays for none of their per-node bookkeeping.
+     * Hold, long-hold and waterfall tracking are records with verdicts on top,
+     * not checks, and are unaffected; disable those through their own options.
+     */
+    checks?: boolean;
     /** Capture the user stack frame of each write — slow (default false). */
     stacks?: boolean;
-    /** Ring-buffer size for `history()` (default 200). */
+    /** Ring-buffer size for each `history(type)` buffer (default 200). */
     historyLimit?: number;
     /**
      * Hot-scope warning: emit a diagnostic when one scope re-runs `count`
@@ -102,8 +391,11 @@ export interface AttributionOptions {
     } | false;
     /**
      * Wide-scope warning: emit a diagnostic when a scope's dependency count
-     * reaches this (default 30) — the coarse-read / helper-leak signature.
-     * Re-warns only if the count then grows by another 50%. `false` disables.
+     * reaches this (default 200) — the coarse-read / helper-leak signature.
+     * Re-warns only if the count then grows by another 50%. HMR plumbing is
+     * not counted, and in dev builds the renderer's insert child-resolution
+     * pass (which must track every row's resolved child) is not judged;
+     * HUGE_FAN_IN (2000, always on) still covers both. `false` disables.
      */
     wideDeps?: number | false;
     /**
@@ -126,61 +418,499 @@ export interface AttributionOptions {
      */
     unstableMemos?: number | false;
     /**
-     * Written-fan-out warning: emit a diagnostic when a committed root
-     * invalidation (write, refresh, async landing) reaches a node with at
-     * least this many subscribers (default 250). Complements the always-on
-     * HUGE_FAN_OUT graph-size warning, specced against it deliberately:
-     * HUGE_FAN_OUT fires at LINK time from GRAPH_SIZE_WARN_AT (2000) up —
-     * static structure so large it warns even if never written — while this
-     * fires at WRITE time from a much lower bar, because fan-out only costs
-     * anything when the node actually changes. Once per node, re-warning only
-     * on 2x subscriber growth, so the two never spam the same node. `false`
+     * Wasted-recompute warning: emit WASTED_RECOMPUTE when a scope re-ran at
+     * least `minRuns` times within `windowMs` and `ratio` or more of those
+     * runs produced an unchanged value while costing `budgetMs` or more of
+     * compute in all (default 5 runs / 80% / 2ms / 1000ms). The equality gate
+     * closed every time: the scope's inputs changed without changing its
+     * result, so the runs were pure cost — the profiler-shaped fact
+     * `costs().wastedMs` sums, as a finding. Held and overlay runs are not
+     * counted (they may be replayed). Once per window per scope. `false`
      * disables.
      */
-    wideWrites?: number | false;
-}
-export interface ScopeCost {
-    name: string;
-    kind: "effect" | "memo";
-    runs: number;
-    selfMs: number;
+    wastedRecompute?: {
+        minRuns: number;
+        ratio: number;
+        budgetMs: number;
+        windowMs: number;
+    } | false;
+    /**
+     * HUGE_FAN_OUT threshold while the engine is enabled: emit the finding
+     * when a committed root invalidation (write, refresh, async landing)
+     * reaches a node with at least this many subscribers (default 250). The
+     * same code the always-on core check emits from GRAPH_SIZE_WARN_AT (2000)
+     * up, with `data.write` naming the invalidation; the engine hands over to
+     * the core there, so one change never carries two findings. Once per
+     * node, re-warning once the count has grown by another 500. HMR plumbing
+     * subscribers are not counted. `false` leaves only the always-on threshold.
+     */
+    fanOut?: number | false;
+    /**
+     * Async-waterfall warning: emit a diagnostic when an async flight that
+     * could only start after an upstream flight resolved (its recompute's
+     * cause chain reaches the upstream's async landing, and its origin
+     * post-dates that landing) forms a sequential chain of 2+ flights, each
+     * of which took at least `minFlightMs` (default 50ms). The duration gate
+     * is one safety valve for what the graph cannot see: a settled
+     * preload/cache hit resolves fast and never warns. In-flight preloads are
+     * absolved by origin: `markFlight()` stamps (and first-seen identity)
+     * prove work predated the upstream landing — parallel, not sequential.
+     * Chains of 2 emit at `info` severity, structured channel only (a
+     * dependent fetch is sometimes intrinsic, and unmarked external preloads
+     * are invisible); 3+ escalate to `warn` with console output. `false`
+     * disables.
+     */
+    waterfalls?: {
+        minFlightMs: number;
+    } | false;
+    /**
+     * Silent-hold warning: emit a diagnostic when a user's writes were held
+     * behind async work for at least `infoMs` (default 100ms — RAIL's "feels
+     * instant" ceiling) and the screen never acknowledged the wait — no
+     * `isPending()`/`latest()` reader downstream of the held writes or their
+     * blockers, no optimistic value, no `affects()` mark, and no lane effect
+     * painted while held. Below `warnMs`
+     * (default 200ms — the INP "good" ceiling) the event is advisory
+     * (structured channel only); at or above it the console gets the finding.
+     * The engine measures to the commit, not the paint, so every number is a
+     * floor on what the user saw; the thresholds sit at the strict end of the
+     * band on purpose. Holds that staged no root write (initial loads, bare
+     * `refresh()`) are never judged: nothing the user did went unanswered.
+     * `false` disables hold tracking altogether (`longHolds` included).
+     */
+    holds?: {
+        infoMs: number;
+        warnMs: number;
+    } | false;
+    /**
+     * Long-hold warning: emit a diagnostic when a hold's quiescent tail — the
+     * time from the LAST write to join it until it committed — reached
+     * `infoMs` (default 500ms), `warn` from `warnMs` (default 1000ms, where
+     * RAIL says the user loses the thread). Measured from the last join so a
+     * hold that keeps taking input (typing) is judged by each wait, not by its
+     * lifetime. A hold this long is past what a stale screen should carry,
+     * acknowledged or not: the honest UI is a fallback, which a `Loading`
+     * boundary gives only when it has not revealed yet or its `on` prop
+     * changed. Reported as LONG_HOLD when the hold was acknowledged; a silent
+     * long hold stays one SILENT_HOLD with the boundary repair appended.
+     * `false` disables.
+     */
+    longHolds?: {
+        infoMs: number;
+        warnMs: number;
+    } | false;
+    /**
+     * Graph-growth warning: emit GRAPH_GROWTH when the live owner count at the
+     * settle of the same route has climbed on `visits` consecutive visits
+     * (default 3) to `ratio` or more of the first (default 1.25), and so has
+     * the part of it that outlived each of those navigations — a root or a
+     * subscription each visit leaves behind, the leak class a heap snapshot
+     * finds, not content a visit owns and leaving disposes. The count is a walk
+     * when a navigation opens and at its settle, never per node. Navigations
+     * that stay on a route pattern continue its visit and are not compared.
+     * `false` disables.
+     */
+    graphGrowth?: {
+        visits: number;
+        ratio: number;
+    } | false;
     /**
-     * Self-time of PLAIN, non-held runs that produced an unchanged value —
-     * the recoverable number. Overlay runs (optimistic/transition) are never
-     * counted here: an optimistic recompute landing back on the committed
-     * value is the mechanism working, not waste.
+     * Abandoned-flights warning: emit ABANDONED_FLIGHTS when one async source
+     * abandons `count` flights within `windowMs` — each superseded by the
+     * next before it landed (default 3 / 1000ms: the request-per-keystroke
+     * signature, every input asking again and the answers discarded). Once
+     * per window per source. In dev builds a framework node that supersedes
+     * its own flights by design (a frames slot arg re-read by each re-shipped
+     * record) is not judged. `false` disables.
      */
-    wastedMs: number;
-    /** Self-time spent in optimistic/transition (overlay) runs. */
-    overlayMs: number;
+    abandonedFlights?: {
+        count: number;
+        windowMs: number;
+    } | false;
+    /**
+     * Fallback-flash finding: emit FALLBACK_FLASH (`info`) when a `Loading`
+     * boundary's fallback shows for less than `FALLBACK_FLASH_MS` (150ms) — a
+     * spinner that appeared and vanished, too much feedback for too little
+     * wait (default true). `false` disables.
+     */
+    fallbackFlashes?: boolean;
+    /**
+     * Stacked-holds warning: emit STACKED_HOLDS when `count` or more
+     * interactions are waiting in one hold when it commits (default 3) — the
+     * person kept clicking or typing while the first answer was in the air,
+     * and every one of them waited on the same source. `false` disables.
+     */
+    stackedHolds?: {
+        count: number;
+    } | false;
+    /**
+     * Optimistic-revert finding: emit OPTIMISTIC_REVERTED (`info`) when an
+     * optimistic value the screen showed is replaced by a different one —
+     * reverted at settle, or superseded by the truth (default true). The
+     * runtime's own optimistic nodes — `isPending`/`latest` companions and
+     * derived overrides — are never judged: they are the acknowledgement
+     * machinery, not a guess the person saw. `false` disables.
+     */
+    optimisticReverts?: boolean;
 }
-export interface WriteCost {
-    /** Root cause name (a signal write, async landing, or refresh target). */
-    name: string;
-    /** Number of downstream re-runs this root triggered. */
-    runs: number;
-    /** Summed self-time of every downstream re-run it caused. */
-    downstreamMs: number;
+/** A fallback shown for less than this is a flash: feedback for a wait too short to need it. */
+export declare const FALLBACK_FLASH_MS = 150;
+/**
+ * The ring buffers `attribution.history(type)` reads, by type. Four of the
+ * five are the channel's records kept since the window opened; `waterfall`
+ * is a fact the engine keeps but never emits — a graph-provable sequential
+ * flight chain (`WaterfallRecord`), of which the ASYNC_WATERFALL finding is
+ * the thresholded view.
+ */
+export interface HistoryRecords {
+    rerun: RerunEvent;
+    waterfall: WaterfallRecord;
+    hold: HoldEvent;
+    navigation: NavigationEvent;
+    interaction: InteractionEvent;
 }
+export type HistoryType = keyof HistoryRecords;
+/** @internal The engine's clock: `performance.now()` where it exists. */
+export declare const now: () => number;
+/**
+ * @internal The fold tables' seam. The engine emits records; the tables that
+ * fold them — `costs()` (scopes and writes), `feedback()` (sources,
+ * interactions, navigations, flights, fallbacks) — live in their own modules
+ * and register here when imported, so a consumer that only subscribes to
+ * records never ships them. Each hook fires at the moment the engine has the
+ * fact; `reset` on `enable()`/`disable()`. With nothing registered every
+ * site is one empty loop.
+ */
+export interface FoldHooks {
+    rerun?(el: Computed<any>, event: RerunEvent): void;
+    hold?(event: HoldEvent): void;
+    navigation?(event: NavigationEvent): void;
+    /** A flight started; `abandoned` when it superseded one still in the air. */
+    flightStart?(el: Computed<any>, abandoned: boolean): void;
+    flightLanded?(el: Computed<any>, ms: number): void;
+    fallback?(boundary: object, tree: Computed<any> | undefined, shown: boolean): void;
+    reset?(): void;
+}
+/** @internal */
+export declare function registerFold(hooks: FoldHooks): void;
+/** @internal Root cause names of a cause chain — the writes/landings/refreshes the chain bottoms out in. */
+export declare function rootsOf(causes: ChangeRecord[], out: Set<string>): void;
+export declare function nodeName(node: Signal<any> | Computed<any>): string;
+/** `click on button#next "Next →"`, `effect "syncTitle"`, `action "save"`, `navigation to /users/:id`, … */
+export declare function formatOrigin(origin: ChangeOrigin): string;
 export declare function formatRerun(event: RerunEvent): string;
+/**
+ * The engine: turn it on and read its ring buffers. Its records are
+ * delivered on the core's channel — `OBSERVE.records.subscribe("rerun" |
+ * "hold" | "interaction" | "navigation" | "create" | "effect" | "flush" |
+ * "flight" | "fallback" | "graph", (event, live) => …)` — the same place
+ * the runtimes' records arrive, so a consumer has one subscribe. The folds
+ * over those records — `costs()`, `feedback()` — and the point queries —
+ * `why()`, `subscriptions()` — and the formatters — `formatRerun()`,
+ * `formatOrigin()` — are named exports of `@solidjs/signals/attribution`
+ * rather than methods here, so a consumer that only wants records (a
+ * production adapter) does not ship the tables a console or an agent reads;
+ * importing a fold is what turns its accounting on.
+ */
 export interface Attribution {
-    enable(opts?: AttributionOptions): void;
+    /**
+     * Install the engine, or take one more hold on it, and return the release
+     * for that hold. The engine is shared by every consumer in the page — a
+     * profiler track, an APM adapter, a diagnostics capture — so each
+     * `enable()` is a hold: the first installs the hooks and resets
+     * everything; one taken while already enabled opens a fresh window over
+     * the ring buffers and folds (`history(type)`, `costs()`, `feedback()` …
+     * read from here on) without disturbing the live tracking state or
+     * anyone's subscriptions, so a capture begun beside a running consumer
+     * still measures only its own scenario.
+     *
+     * Options combine across holds by the most demanding value per key: a
+     * hold's `opts` say what it wants (the defaults fill what it leaves
+     * unsaid), and the engine does whatever any holder asked for — the log
+     * prints while any holder wants it, a check runs while any holder wants
+     * it and at the most sensitive threshold requested, `historyLimit` is the
+     * largest. A hold can add to what another asked for, never take it away,
+     * so the result does not depend on the order holds were taken; releasing
+     * a hold withdraws its requests — a track enabled with `log: false` beside
+     * a console session never silences it, and a capture with tight
+     * thresholds beside a records-only adapter runs the checks for its own
+     * duration. One key runs the other way: `values` combines to the LEAST
+     * permissive level any holder asked for, so an adapter that must not see
+     * user data (`values: "none"`) is honoured beside a console session that
+     * wants everything. The release is idempotent; the last release uninstalls the
+     * hooks and clears every ring buffer. A consumer that re-`enable()`s to
+     * reopen its window must release both holds (or `disable()`).
+     *
+     * Subscriptions are not the engine's: its records arrive on
+     * `OBSERVE.records`, whose listeners outlive any hold — subscribe before
+     * or after `enable()`, and unsubscribe with the function `subscribe`
+     * returned.
+     */
+    enable(opts?: AttributionOptions): () => void;
+    /**
+     * Tear the engine down whatever holds are outstanding: drops every hold,
+     * uninstalls the hooks and clears every ring buffer. The console's and a
+     * test harness's reset — a consumer sharing the page with others releases
+     * its own hold with the function `enable()` returned instead. Idempotent;
+     * a `disable()` with nothing enabled is a no-op reset. Listeners on
+     * `OBSERVE.records` are untouched: they are the channel's.
+     */
     disable(): void;
-    subscribe(listener: (event: RerunEvent) => void): () => void;
-    history(): readonly RerunEvent[];
-    /** Re-run history for one node — pass a memo/effect accessor or raw node. */
-    why(target: unknown): RerunEvent[];
-    /** Current dependency names of one scope — the devtools subscription view. */
-    subscriptions(target: unknown): string[];
-    /**
-     * Aggregated cost tables since enable(): `scopes` ranked by self-time
-     * (with `wastedMs` = time spent on unchanged-value runs), `writes` ranked
-     * by total downstream re-run time each root write caused.
-     */
-    costs(): {
-        scopes: ScopeCost[];
-        writes: WriteCost[];
-    };
-    format: typeof formatRerun;
+    /**
+     * The ring buffer of `type` since `enable()` — the last `historyLimit`
+     * records (default 200), oldest first; the same objects the channel
+     * delivered. Facts, not verdicts: each is recorded regardless of the
+     * thresholds the findings apply to it.
+     *
+     * - `"rerun"` — every re-run recorded (see `RerunEvent`). Kept only while
+     *   a record had an audience — a `rerun` listener, an imported fold
+     *   (`costs`/`feedback`), or the console log; a records-only consumer
+     *   that wants none of those pays for none, and reads an empty buffer.
+     * - `"waterfall"` — every graph-provable sequential flight chain, warned
+     *   or not: the ASYNC_WATERFALL finding is the duration-gated view.
+     * - `"hold"` — every settled transition hold that staged at least one root
+     *   write, acknowledged or not: SILENT_HOLD / LONG_HOLD are the
+     *   thresholded verdicts (`HoldEvent.silent` / `.long` carry them).
+     * - `"navigation"` — every navigation a router declared via `withOrigin`,
+     *   settled or not: what route, under which interaction, how many writes,
+     *   and — once its writes are through — how long that took and how
+     *   (`committed`, `held` with the `HoldEvent` attached, `superseded`).
+     * - `"interaction"` — every user interaction a runtime declared via
+     *   `withInteraction`, settled or not: what was dispatched, when, what it
+     *   wrote, the re-runs and creations it caused, the holds its writes
+     *   waited in and the navigations it performed — and, once through, how
+     *   long the person waited (`settledMs`) and how it ended. The
+     *   per-dispatch record `feedback().interactions` folds by name.
+     */
+    history<K extends HistoryType>(type: K): readonly HistoryRecords[K][];
+    /**
+     * Cooperative preload declaration: stamp a flight object (promise or async
+     * iterable) with its true kickoff time BEFORE the reactive graph sees it.
+     * A route preloader or query cache calls this on the promise it hands out
+     * (on the WRAPPER it mints, with the original kickoff time — wrapping
+     * defeats identity tracking otherwise); any dependent that later awaits it
+     * is then judged against the real start — work already in the air when its
+     * upstream landed is parallel, never a waterfall link. Callable while
+     * attribution is disabled (marks made at navigation time must survive a
+     * later enable()).
+     */
+    markFlight(flight: object, startedAt?: number): void;
+}
+/** @internal The id the engine's records name `node` by, if it has one yet — read without assigning. */
+export declare function nodeIdOf(node: object): number | undefined;
+/** One landed flight: its node name, wall duration, and upstream chain. */
+export interface FlightLink {
+    name: string;
+    ms: number;
+}
+export interface WaterfallRecord {
+    /** Sequential flights, oldest first, ending at the flight that landed. */
+    chain: FlightLink[];
+    /** Summed wall time of the chain — the serialized cost. */
+    sequentialMs: number;
+}
+export interface HeldWrite {
+    name: string;
+    prev?: string;
+    value?: string;
+    origin?: ChangeOrigin;
+}
+export interface HoldEvent {
+    /**
+     * When the wait began (`performance.now()` clock): the interaction that
+     * performed the held writes when one is known (`interaction.at`) or the
+     * first flush that parked them, whichever is earlier. `at + holdMs` is the
+     * commit.
+     */
+    at: number;
+    /** Wall time the user waited: `at` to the commit. */
+    holdMs: number;
+    /**
+     * The quiescent tail: from the LAST held write to join (the user's final
+     * input) to the commit. Equal to `holdMs` for a single write; shorter when
+     * the hold kept taking input. The LONG_HOLD measure.
+     */
+    tailMs: number;
+    /** The user interaction whose writes were held, when the stamp is known. */
+    interaction?: ChangeOrigin;
+    /**
+     * The declared unit of work the held writes belong to — the `navigation`
+     * a router described via `withOrigin` — when one is known. What names the
+     * hold by route (`navigation to /users/:id`) rather than by signal; the
+     * same object as `NavigationEvent.origin`, so the two join by identity.
+     */
+    origin?: ChangeOrigin;
+    /** Flushes that ended with the hold still open. */
+    flushes: number;
+    /** Root signal writes staged behind the hold (the user's unanswered input). */
+    heldWrites: HeldWrite[];
+    /** Async nodes the hold waited on (union across its parked flushes). */
+    blockers: string[];
+    /**
+     * Feedback the graph provably rendered for this hold — an `isPending()`
+     * reader, a `latest()` shadow, an optimistic value, an `affects()` mark —
+     * one entry per affordance, in the order found. Empty and
+     * `paintedDuringHold === 0` is the SILENT_HOLD signature.
+     */
+    acknowledgements: Acknowledgement[];
+    /**
+     * Effect callbacks that ran inside the hold's parked flushes. Mainline
+     * effects are stashed while a hold is open, so these are lane effects —
+     * readers of optimistic values and of `isPending()`/`latest()` companions,
+     * i.e. the screen changing in response to the hold. An unrelated effect
+     * cannot land here: it waits with everything else.
+     */
+    paintedDuringHold: number;
+    /** The hold was opened (or joined) by an `action()`. */
+    action: boolean;
+    /**
+     * The engine's silent-hold verdict, stamped at settle: no affordance
+     * acknowledged the wait (`acknowledgements` empty) and nothing painted
+     * while it was open (`paintedDuringHold === 0`). Duration-free — the
+     * SILENT_HOLD finding is this above `holds.infoMs` — so a consumer can
+     * flag every silent wait or apply its own floor, and a record that left
+     * the process still carries the verdict.
+     */
+    silent: boolean;
+    /**
+     * The engine's long-hold verdict, stamped at settle: the quiescent tail
+     * (`tailMs`) reached `longHolds.infoMs` under the options in effect;
+     * `false` when long-hold tracking is off. The LONG_HOLD finding is this
+     * verdict, tiered by `warnMs`.
+     */
+    long: boolean;
+}
+/**
+ * One way the screen acknowledged a hold. `reader` is where it was painted —
+ * the owner path of the first effect the census found reading the
+ * affordance (`["<App>", "<Feed>", "effect"]`), so a consumer can say WHICH
+ * screen answered, not only that one did. Absent when the affordance was
+ * registered but the census found no reader through the graph (an optimistic
+ * store: its readers are proxy traps, not nodes). `feedback()` ranks
+ * acknowledgements by `kind:source` (`isPending:posts`).
+ */
+export interface Acknowledgement {
+    kind: "isPending" | "latest" | "optimistic" | "affects";
+    /** The node the affordance hangs on — the async source for `isPending`/`latest`, the optimistic/affected node otherwise. */
+    source: string;
+    reader?: string[];
+}
+/**
+ * The live graph's size — a walk, not a counter: nothing is charged at node
+ * creation, disposal, write or re-run; the engine asks at a navigation's
+ * settle. Two passes. The owner tree from the registered top-level roots
+ * gives `owners` and seeds the computations. Then each computation's
+ * dependency links give `edges` and the `signals` and computations they
+ * reach, and each newly met source's subscriber list gives the computations
+ * that read it — including one created with no owner, which no chain holds
+ * and only its sources keep alive: the subscription leak a heap snapshot
+ * finds. Measured at 5–10 ns per link; a 50k-owner graph walks in under a
+ * millisecond. Dormant nodes are spliced out of their chain and are not
+ * counted unless a subscription still reaches them.
+ */
+export declare function graphSize(): GraphSize;
+/** A destination a navigation abandoned when a redirect sent it elsewhere. */
+export interface NavigationHop {
+    name?: string;
+    to?: string;
+    params?: Readonly<Record<string, string | undefined>>;
+    /** When the redirect away from it was declared (`performance.now()` clock). */
+    at: number;
+}
+export interface NavigationEvent {
+    /** The matched route pattern the router gave — `/users/:id`. After a redirect, the final one. */
+    name?: string;
+    to?: string;
+    from?: string;
+    params?: Readonly<Record<string, string | undefined>>;
+    /** When the navigation was requested (`performance.now()` clock). */
+    at: number;
+    /**
+     * The route the document arrived on, declared by the router at its initial
+     * match (`NavigationRef.initial`): `at` is the document's navigation start
+     * (`0` unless the router passed its own), there is no `from`, `writes` is
+     * `0`, and it settles `committed` when the router's frame closes — the
+     * record names the route the page loaded as; its timing is Navigation
+     * Timing's, not the runtime's.
+     */
+    initial?: true;
+    /** The user interaction it ran under, when known — a link click. */
+    interaction?: ChangeOrigin;
+    /** Root writes the frame performed, redirect hops included. */
+    writes: number;
+    /** Destinations abandoned along the way, in order — present only when a redirect occurred. */
+    redirects?: NavigationHop[];
+    /**
+     * Wall time from the request to settle: the end of the drain that committed
+     * its writes, or the commit of the hold they waited in. `undefined` while
+     * unsettled.
+     */
+    settledMs?: number;
+    outcome?: "committed" | "held" | "superseded";
+    /** The hold its writes waited in, when hold tracking recorded one. */
+    hold?: HoldEvent;
+    /**
+     * The frame object its writes were stamped with — `ChangeRecord.origin` on
+     * each, `HoldEvent.origin` on the hold. Join key, by identity.
+     */
+    origin: ChangeOrigin;
+}
+export interface InteractionEvent {
+    /** Event type — `click`, `keydown`, `input`… */
+    name: string;
+    /** The element hit, as the runtime described it — `button#next "Next →"`. */
+    target?: string;
+    /**
+     * When the interaction began (`performance.now()` clock): the browser event's
+     * own timestamp when the runtime supplied it, else the moment the handler
+     * frame opened. Joins `PerformanceEventTiming.startTime` for the same event.
+     */
+    at: number;
+    /**
+     * Browser event creation to handler entry (ms) — the queueing the browser's
+     * INP counts as input delay. Present only when `at` predates the frame.
+     */
+    inputDelayMs?: number;
+    /**
+     * Wall time of the handler itself, entry to return — for the frames of one
+     * event (`InteractionRef.event`), the first listener's entry to the last
+     * one's return, as Event Timing measures processing.
+     */
+    handlerMs: number;
+    /** Root writes attributed to the frame: the handler's, and those of frames it opened (a navigation). */
+    writes: number;
+    /** Re-runs traced back to this interaction while the record was open. */
+    runs: number;
+    /** Computations created in those runs or in the frame's flushes (the "create 1,000 rows" work). */
+    created: number;
+    /** Summed self-time of `runs` and `created` (ms). Quantized per run; `settledMs` is the wall clock. */
+    runMs: number;
+    /** Holds its writes waited in, in settle order. */
+    holds: HoldEvent[];
+    /** Navigations performed under it, in open order. */
+    navigations: NavigationEvent[];
+    /**
+     * Dispatch to settle: the handler's return when it wrote nothing, the end of
+     * the drain that committed its writes, or the commit of the last hold they
+     * waited in — whichever came last. `undefined` while unsettled.
+     */
+    settledMs?: number;
+    outcome?: "idle" | "committed" | "held";
+    /**
+     * The handler returned a thenable (`async () => { await save(); … }`) and
+     * the record waited for it: handler return → the promise settling (the last
+     * one, when several listeners of one event returned one), in
+     * milliseconds, capped at `ASYNC_HANDLER_CAP_MS`. The continuation runs
+     * with no frame on the stack, so writes it makes are not attributed to
+     * this interaction — only its duration is. Absent when the handler
+     * returned synchronously.
+     */
+    continuationMs?: number;
+    /**
+     * The frame object every downstream fact carries — `ChangeOrigin.interaction`
+     * on writes and frames, `RerunEvent.interaction`, `HoldEvent.interaction`,
+     * `NavigationEvent.interaction`. Join key, by identity.
+     */
+    origin: ChangeOrigin;
 }
 export declare const attribution: Attribution;