@bivelio/savings-layer 0.4.36 → 0.5.0-rc.1

package/dist/index.d.ts

@@ -47,12 +47,6 @@ interface CostModel {
     /** Optional discounted rate for provider-cached (prefix) input tokens. */
     cachedInputPer1k?: number;
 }
-/** Oldest provider counter a repeat can be valued against: one billing period. */
-declare const REFERENCE_MAX_AGE_MS$1: number;
-/** Share of the counted output credited to a repeat: only under deterministic sampling, on a catalog model that does not reason and honors it, when the call that ran carried no hidden reasoning. */
-declare const REFERENCE_OUTPUT_SHARE$1 = 0.95;
-/** Input tokens deducted per piece the layer inserted in the call that ran: a message, the wrapper around a native output schema, or half of the empty parameters given to a tool that had none. */
-declare const REFERENCE_MESSAGE_ALLOWANCE$1 = 8;
 /**
  * Why an avoided call (a cache or coalesced hit) could not be valued by
  * reference, and so stays `estimated`.
@@ -102,185 +96,6 @@ interface ReferenceCredit {
     /** Fraction of the standard price applied: 1, or the discounted tier the call ran on or this request declared. */
     tierFactor: number;
 }
-declare class SavingsLedger$1 {
-    avoidedLlm: number;
-    baselineLlm: number;
-    actualLlm: number;
-    baselineRetrieval: number;
-    actualRetrieval: number;
-    baselineTools: number;
-    actualTools: number;
-    retries: number;
-    baselineInput: number;
-    actualInput: number;
-    eliminatedInput: number;
-    compressedInput: number;
-    baselineOutput: number;
-    actualOutput: number;
-    eliminatedOutput: number;
-    providerCachedInput: number;
-    /** Tokens the provider billed as prefix-cache writes. */
-    cacheWriteInputTokens: number;
-    /** The premium on those writes, already priced in USD. Adds to `actual` only. */
-    cacheWritePremiumUsd: number;
-    /** Provider-initiated cache-write premium, priced in USD. Adds to baseline and `actual` alike. */
-    providerCacheWritePremiumUsd: number;
-    /** Cache-write tokens of the served call, every TTL. Read only for `report.nativeSavings`. */
-    nativeCacheWriteInput: number;
-    /** The one-hour part of `nativeCacheWriteInput`. */
-    nativeCacheWrite1hInput: number;
-    /** Whether the layer placed the cache breakpoint on the served call; `null` when unknown. */
-    nativeCacheMarkedByLayer: boolean | null;
-    /** Cuenta LOCAL del prompt que de verdad se envió al proveedor. */
-    sentInputLocal: number;
-    /** Tokens que la capa quitó de ese prompt, en cuenta LOCAL. */
-    removedInputLocal: number;
-    /**
-     * Tokens (LOCAL ruler) that tool_result ELISION removed from the sent prompt.
-     * They exist to be SUBTRACTED from `removedInputLocal`: live deletion really
-     * lowers `sentInputLocal`, so without this correction the elision saving would
-     * enter the reconstructed baseline and the commission, labelled "estimated".
-     * The doctrine is the opposite — "live count, shadow money": elision is a COUNT
-     * (ledger 2 `eliminatedInput`), its dollar is measured by the shadow lane out
-     * of commission, and `basis` stays "measured". Deliberately distinct from the
-     * tool-selector drop, which DOES count in the baseline (a declared tool that is
-     * not sent is an attributable saving; a superseded tool_result is redundancy
-     * not yet billed as verifiable against the invoice).
-     */
-    elididoLocal: number;
-    /**
-     * Tokens (LOCAL ruler) that tool-output compaction removed from the sent
-     * prompt. Subtracted from `removedInputLocal` for the same reason as
-     * {@link elididoLocal}: the saving is a COUNT in ledger 3 (`compressedInput`)
-     * and never enters the reconstructed baseline or the commission, so `basis`
-     * stays "measured" and the provider's invoice shows the effect on its own.
-     */
-    compactedLocal: number;
-    /**
-     * ¿El estimador que produjo las dos cifras anteriores cuenta con la regla del
-     * proveedor? Lo fija el pipeline al resolver el tokenizador del modelo.
-     * Defaults to NO: exactness is claimed only once it has been verified.
-     */
-    estimatorExact: boolean;
-    /** Salida del reintento de reparación. Es coste NUESTRO, nunca del baseline. */
-    repairOutput: number;
-    /**
-     * INPUT of the repair retry. Mirrors `repairOutput`.
-     *
-     * Both are OUR cost, never the customer's, so both are subtracted from the
-     * reconstructed baseline. The baseline is what the naive path would have
-     * cost; work the layer does to fix its own output has no place in it.
-     */
-    repairInput: number;
-    /**
-     * Billed provider calls in this request that came back WITHOUT a complete
-     * usage counter (`usage.complete === false`). Any value above zero makes the
-     * request `unmeasured`: a figure built on a missing counter is not the
-     * provider's meter.
-     */
-    callsWithoutUsage: number;
-    /**
-     * INPUT of the schema-on-demand discovery probe. Held apart exactly like
-     * {@link repairInput}, and for the same reason.
-     *
-     * The layer ADDS this call; the naive path would never make it. Letting its
-     * tokens into the reconstructed baseline would make the probe pay for itself,
-     * and a request that cost MORE than going without the layer would publish a
-     * positive saving. It still counts in full towards `actualInput` — that is
-     * money genuinely spent — it just cannot inflate what that is compared against.
-     */
-    discoveryInput: number;
-    /** OUTPUT of the discovery probe. Our cost, kept out of the baseline. */
-    discoveryOutput: number;
-    /**
-     * Already-priced cost of a call the model cascade made to a cheaper model
-     * whose answer was NOT served. Added to `actual`, never to the baseline.
-     */
-    extraCostUsd: number;
-    /** Input tokens of that call. Published by the pipeline, not priced here. */
-    cascadeInput: number;
-    /** Output tokens of that call. Published by the pipeline, not priced here. */
-    cascadeOutput: number;
-    /**
-     * Record a call to the cheap model that did not produce the served answer:
-     * the one discarded on an escalation, or a shadow-mode sample.
-     *
-     * The caller raises `actualLlm`, because the call happened and the report owes
-     * the reader that count.
-     *
-     * @param providerWritePremium the multiplier the provider bills on its own
-     *   for the call's cache writes (`usage.cacheWriteInputTokens`); 1 = none.
-     */
-    cascadeAttempt(usage: {
-        inputTokens: number;
-        outputTokens: number;
-        cacheWriteInputTokens?: number;
-    }, cheapPrice: CostModel, providerWritePremium?: number): void;
-    /** Record the premium for `tokens` of cache write billed at `multiplicador`x. */
-    cacheWritePremium(tokens: number, multiplicador: number, precioDelQueMarco: CostModel): void;
-    /**
-     * Lo que de verdad hizo la compresión, para publicarlo en
-     * `SavingsReport.compression`. El campo estaba declarado en el contrato y no lo
-     * rellenaba nadie: salía `undefined` aunque la compresión hubiera corrido, y
-     * `fidelityPassed` se calculaba sin llegar nunca al informe.
-     */
-    compressionReport: {
-        originalTokens: number;
-        compressedTokens: number;
-        protectedSpans: number;
-        fidelityPassed: boolean;
-    } | undefined;
-    /** Ledger 2: tokens removed because they were not needed. */
-    eliminate(tokens: number): void;
-    /** Ledger 3: delta between an original encoding and a cheaper one. */
-    compress(originalTokens: number, compressedTokens: number): void;
-    /** Ledger 4: tokens still present but billed/processed as provider-cached. */
-    providerCached(tokens: number): void;
-    /** Record the served call's cache writes for `report.nativeSavings`. Never priced into `estimateCost`. */
-    nativeCacheWrite(tokens: number, tokens1h: number): void;
-    /** Record the premium for `tokens` of cache write the provider bills at `multiplier`x on its own. */
-    providerCacheWritePremium(tokens: number, multiplier: number, price: CostModel): void;
-    /** Input tokens of calls the provider documents as not billed; already included in `actualInput`. */
-    unbilledInput: number;
-    /** Output tokens of those calls; already included in `actualOutput`. */
-    unbilledOutput: number;
-    /** Provider-cached input tokens of those calls; already included in `providerCachedInput`. */
-    unbilledCachedInput: number;
-    /** True when the answer served is a refusal the provider does not bill: no saving is credited. */
-    servedUnbilledRefusal: boolean;
-    /** Record a call the provider documents as not billed: its tokens stay counted and are priced at 0. */
-    unbilledCall(usage: {
-        inputTokens: number;
-        outputTokens: number;
-        cachedInputTokens?: number;
-    }): void;
-    /** Set when an avoided call is valued against the provider counter of an identical call that ran. */
-    reference: ReferenceCredit | null;
-    /** Set when an avoided cache or coalesced hit could not be valued by reference: why it stays estimated. */
-    referenceRejected: ReferenceRejection | null;
-    /** Record an avoided call valued by reference. */
-    avoidCallByReference(credit: ReferenceCredit): void;
-    /** Record that a whole primary LLM call was avoided (cache / resolver hit). */
-    avoidCall(baselineInput: number, baselineOutput: number): void;
-    /**
-     * Baseline de salida publicable.
-     *
-     * En una llamada EVITADA es el de la respuesta que se habría generado. En una
-     * que se EJECUTÓ es la salida medida menos la de la reparación — nunca el techo
-     * `maxOutputTokens ?? 256`. estimateCost ya lo hacía por dentro, pero el
-     * informe publicaba el techo crudo, y el proyector público lo convertía en
-     * dólares: se anunciaba un ahorro de salida que nunca existió.
-     */
-    get publishedBaselineOutput(): number;
-    estimateCost(model: CostModel): {
-        baselineEstimate: number;
-        actual: number;
-        netSavings: number;
-        netSavingsRatio: number;
-        netSavingsUsd: number;
-        basis: CostBasis;
-    };
-}
 
 /**
  * LIST PRICES PER MODEL.
@@ -308,7 +123,7 @@ declare class SavingsLedger$1 {
  * announced price steps, cache-write premiums, tier rates). Two computations
  * citing the same version were priced with the same content.
  */
-declare const PRECIOS_AL_DIA$1 = "2026-09-29";
+declare const PRECIOS_AL_DIA = "2026-09-29";
 /**
  * Key = the model id as the provider publishes it, without a vendor prefix.
  * Lookup tries the exact id, then the id without its provider prefix, then the
@@ -319,7 +134,7 @@ declare const PRECIOS_AL_DIA$1 = "2026-09-29";
  * provider has already announced is applied by {@link precioDeLista} from its
  * effective date, so a request is always priced at the rate of its own day.
  */
-declare const CATALOGO$1: Readonly<Record<string, CostModel>>;
+declare const CATALOGO: Readonly<Record<string, CostModel>>;
 /**
  * The list price for a model on the day of the request, or `undefined` when
  * the catalogue has no price for it in force that day.
@@ -329,7 +144,7 @@ declare const CATALOGO$1: Readonly<Record<string, CostModel>>;
  *   Prices resolve per UTC day, so a price change the provider has announced
  *   applies from its effective date.
  */
-declare function precioDeLista$1(model: string, at?: Date | number): CostModel | undefined;
+declare function precioDeLista(model: string, at?: Date | number): CostModel | undefined;
 
 /**
  * The version the SDK reports for itself.
@@ -339,17 +154,7 @@ declare function precioDeLista$1(model: string, at?: Date | number): CostModel |
  * on the customer's machine. `test/version-declarada.test.ts` ties this
  * constant to package.json so the two can never drift.
  */
-declare const BV_SALA_VERSION = "0.4.36";
-
-/**
- * Environment-agnostic UUID generation.
- * ---------------------------------------------------------------------------
- * Trace IDs only need to be unique, not cryptographically unpredictable. We
- * prefer the platform `crypto.randomUUID` (present in modern Node, browsers,
- * Deno and edge runtimes) and fall back to a v4-shaped id when it is absent, so
- * BV-SALA runs everywhere without a `node:crypto` import that breaks bundlers.
- */
-declare function randomId$1(): string;
+declare const BV_SALA_VERSION = "0.5.0-rc.1";
 
 /** How the layer served a request that is eligible for a verification check. */
 type VerificationStratum = 
@@ -429,7 +234,12 @@ interface VerificationCheck {
      * was not `missed`. Other pairs are counted and exported, but left out of every figure.
      */
     naive?: VerificationSide & {
-        finishReason?: string;
+        /**
+         * Why the provider stopped the copy, in the same closed set as
+         * `SavingsResult.finishReason` (`"stop"`, `"length"`, `"tool_calls"`,
+         * `"content_filter"`, `"other"`), whatever the provider.
+         */
+        finishReason?: "stop" | "length" | "tool_calls" | "content_filter" | "other";
         /** Whether the copy met your output contract: `none` without a contract or on a tool-call turn. */
         contract: "met" | "missed" | "none";
     };
@@ -671,7 +481,7 @@ interface SavingsToggles {
      * `false` switches it off for this request even when the layer enables it.
      *
      * What it removes is a COUNT in `report.tokens.compressedInput`; it is never
-     * priced into `cost.netSavings` or your commission — your provider's invoice
+     * priced into `cost.netSavingsRatio` or your commission — your provider's invoice
      * shows the effect.
      */
     toolOutputCompaction?: boolean;
@@ -746,7 +556,7 @@ interface SavingsToggles {
      * THE MARGIN IS NOT A GUARANTEE THAT EVERY REQUEST GETS CHEAPER. When the
      * model names no tool in the first call, the request still goes out with every
      * schema and the first call is pure added cost: that request costs MORE, and
-     * the report says so (`cost.netSavings` goes negative). The margin required is
+     * the report says so (`cost.netSavingsRatio` goes negative). The margin required is
      * exactly the cost of that outcome, so the pattern breaks even while it
      * happens on at most one request in two. If your traffic asks for a tool less
      * often than that — a chat assistant with a big catalogue answers most turns
@@ -895,8 +705,6 @@ interface DeferredExecutionReport {
      * all. Present only when {@link credited}.
      */
     discountRate?: number;
-    /** @deprecated Use `tierSaving.amount`; same value. Kept for compatibility. */
-    estimatedTierSaving?: number;
     /**
      * What the served tier was worth on the one call whose tier was read: that call's
      * provider-reported usage at the standard rate card minus the same usage at the rate
@@ -921,7 +729,7 @@ interface DeferredExecutionReport {
      * nothing (3) and is not provider caching (4). It changes the PRICE of the
      * tokens, exactly as the model cascade does, and there is no ledger for that.
      *
-     * `cost.netSavings` therefore does not move: it is a ratio of baseline to
+     * `cost.netSavingsRatio` therefore does not move: it is a ratio of baseline to
      * actual priced from a SINGLE rate card, and a discount that scales both sides
      * cannot change it. Folding this in would require a baseline priced from the
      * other card — a request nobody ran.
@@ -1061,7 +869,7 @@ interface ModelCascadeReport {
      *
      * A COUNTERFACTUAL: the strong model was not run, so this is priced from the
      * measured token counts at both list prices. It is deliberately NOT part of
-     * `cost.netSavings` — that number stays a comparison at a single price, and
+     * `cost.netSavingsRatio` — that number stays a comparison at a single price, and
      * this one would inflate it with a call nobody made. Negative on an escalated
      * request, which is exactly what an escalation costs you.
      */
@@ -1198,29 +1006,36 @@ interface SavingsResult<T = unknown> {
      */
     toolCalls?: ToolCallRequest[];
     /**
-     * Por qué paró de generar el modelo (`"stop"`, `"length"`, `"tool_calls"`, …).
-     * Presente cuando la respuesta viene del proveedor; ausente en los atajos
-     * (caché/resolver), que solo sirven respuestas finales completas.
+     * Why the model stopped generating, the same five values for every provider:
      *
-     * In 0.5 this becomes a CLOSED set — `"stop"`, `"length"`, `"tool_calls"`,
-     * `"content_filter"` or `"other"` — and the provider's own value (for example
-     * `"stop_sequence"`, `"pause_turn"` or `"content_filter: OTHER"`) moves to
-     * `finishReasonDetail`. Compare against those five values only; see the 0.5
-     * migration guide.
+     * - `"stop"`: a complete answer (Anthropic `end_turn` / `stop_sequence`,
+     *   Gemini `STOP`).
+     * - `"length"`: cut off by the output limit (Anthropic `max_tokens`, Gemini
+     *   `MAX_TOKENS`). A truncated answer is never cached.
+     * - `"tool_calls"`: the model asked to call tools (`toolCalls`).
+     * - `"content_filter"`: the provider refused or filtered the answer.
+     * - `"other"`: anything else (for example Anthropic `pause_turn`, Gemini
+     *   `MALFORMED_FUNCTION_CALL`); the provider's value is in
+     *   {@link finishReasonDetail}.
+     *
+     * Present when the answer came from the provider; absent on shortcuts (cache,
+     * local resolver), which only serve complete final answers.
      */
-    finishReason?: string;
+    finishReason?: "stop" | "length" | "tool_calls" | "content_filter" | "other";
     /**
-     * The provider's own reason behind `finishReason`, when it gave one. Present
-     * only on answers the provider returned, and only with a non-empty reason:
+     * The provider's own reason behind `finishReason`, when it adds something.
+     * Present only on answers the provider returned:
      *
-     * - Gemini (OpenAI-compatible endpoint): a filtered answer stops with
-     *   `finish_reason: "content_filter: OTHER"` (or `PROHIBITED_CONTENT`,
-     *   `SAFETY`…); it arrives as `finishReason: "content_filter"` and
-     *   `finishReasonDetail: "OTHER"`.
-     * - Anthropic: a refusal (`stop_reason: "refusal"`) arrives as `finishReason:
-     *   "content_filter"` and `finishReasonDetail` set to `stop_details.category`
-     *   (for example `"cyber"` or `"bio"`).
+     * - A refusal or filter: the finer reason. Gemini's `SAFETY`, `OTHER`,
+     *   `PROHIBITED_CONTENT`… (native or `finish_reason: "content_filter: OTHER"`
+     *   on the OpenAI-compatible endpoint), its `promptFeedback.blockReason`, or
+     *   Anthropic's `stop_details.category` (for example `"cyber"` or `"bio"`).
+     * - A stop reason that is not the provider's plain spelling of one of the
+     *   five values: the provider's literal (`"stop_sequence"`, `"pause_turn"`,
+     *   `"MALFORMED_FUNCTION_CALL"`…).
      *
+     * Absent for the plain spellings (`stop`, `end_turn`, `STOP`, `length`,
+     * `max_tokens`, `MAX_TOKENS`, `tool_calls`, `tool_use`, `content_filter`).
      * OpenAI's Structured Outputs refusal message travels in `refusal` instead.
      * Informational: it changes neither what is billed nor what is credited.
      */
@@ -1419,15 +1234,11 @@ interface QualityTruncationCheck {
     /** True when the provider reported `finishReason: "length"`. A truncated answer is never cached. */
     truncated: boolean;
     /**
-     * Why the provider stopped, as `SavingsResult.finishReason` reports it;
-     * absent when it reported none.
-     *
-     * In 0.5 this becomes the same CLOSED set as `SavingsResult.finishReason` —
-     * `"stop"`, `"length"`, `"tool_calls"`, `"content_filter"` or `"other"` — and
-     * the provider's own value moves to {@link finishReasonDetail}. Compare
-     * against those five values only; see the 0.5 migration guide.
+     * Why the provider stopped, as `SavingsResult.finishReason` reports it (the
+     * same five values); absent when it reported none. The provider's own value
+     * is in {@link finishReasonDetail}.
      */
-    finishReason?: string;
+    finishReason?: "stop" | "length" | "tool_calls" | "content_filter" | "other";
     /**
      * The provider's own reason behind `finishReason`, when it gave one: the same
      * value as `SavingsResult.finishReasonDetail` for the answer that was served.
@@ -1561,7 +1372,7 @@ interface QualityLedger {
  * `cache_control`), route (OpenAI `prompt_cache_key`) or keep the prefix
  * stable, but the discount is billed by the provider and much of it would be
  * there without the layer, so it is never credited as the layer's saving:
- * it is not in `cost.netSavings`, not in `cost.baselineEstimate`, and not in
+ * it is not in `cost.netSavingsRatio`, not in `cost.baselineEstimate`, and not in
  * any commission base.
  */
 interface NativeSavingsReport {
@@ -1728,7 +1539,7 @@ interface SavingsReport {
      * The provider's OWN prompt-cache discount on this call, shown apart. Present
      * only when a call ran and the provider reported cache reads or writes.
      *
-     * It is not the layer's saving and it is never part of `cost.netSavings`,
+     * It is not the layer's saving and it is never part of `cost.netSavingsRatio`,
      * `cost.baselineEstimate`, `basis: "reference"` or the commission base: the
      * discount is the provider's, and `cost` prices cached tokens at the cached
      * rate on both sides for exactly that reason. It exists so the report adds up
@@ -1758,16 +1569,6 @@ interface SavingsReport {
     cost: {
         baselineEstimate: number;
         actual: number;
-        /**
-         * A RATIO, not money: `1 - actual / baselineEstimate`, rounded to six
-         * decimals (0 when nothing was saved, negative when the request cost more).
-         * Do not sum it across requests: it is a ratio, not dollars.
-         *
-         * @deprecated Since 0.4.33, read {@link netSavingsRatio} (the same value) or
-         * {@link netSavingsUsd} (the saving in USD). Removed in 0.5: see the 0.5
-         * migration guide.
-         */
-        netSavings: number;
         /**
          * The saving as a RATIO of the baseline: `1 - actual / baselineEstimate`,
          * rounded to six decimals (0 when nothing was saved or there is no baseline,
@@ -1909,7 +1710,7 @@ interface LicenseGate {
  * went to your provider untouched and nothing was served from a cache.
  *
  * Every figure is HYPOTHETICAL (`basis: "shadow"`): it is never part of
- * `cost.baselineEstimate`, `cost.actual` or `cost.netSavings`, never in any
+ * `cost.baselineEstimate`, `cost.actual` or `cost.netSavingsRatio`, never in any
  * commission base, and never added to measured money.
  */
 interface MeasureReport {
@@ -2105,7 +1906,7 @@ declare class InMemoryExactCache implements ExactCache {
  * requests that mean the same thing must canonicalize identically, and two
  * requests that differ in any way that could change the answer must not.
  */
-interface CanonicalRequest$1 {
+interface CanonicalRequest {
     model: string;
     modelFamily: string;
     messages: ChatMessage[];
@@ -2162,27 +1963,6 @@ interface CanonicalRequest$1 {
     tokenizer: string | undefined;
     original: SavingsRequest;
 }
-/**
- * Normalized form of a turn's text, as it enters the cache key.
- *
- * Folded away: Unicode form, case, whitespace, typographic marks, Latin
- * diacritics and sentence punctuation. Preserved exactly: numbers, decimal and
- * thousands separators, times, dates, units, negations, math operators and any
- * combining mark outside the Latin alphabet.
- */
-declare function normalizeText$1(text: string): string;
-declare function sha256$1(input: string): string;
-/** Stable JSON stringify with sorted keys so hashes are order-independent. */
-declare function stableStringify$1(value: unknown): string;
-/**
- * The version-and-size-independent family of a model id.
- *
- * Exported because the cache key is now built for the model that will actually
- * ANSWER, which is not always the one the caller named. Deriving the family a
- * second time somewhere else is how two rulers start to disagree.
- */
-declare function modelFamilyOf$1(model: string): string;
-declare function canonicalize$1(request: SavingsRequest): CanonicalRequest$1;
 
 /**
  * Thrown by `importCalibration` / `exportCalibration` when a calibration state
@@ -2335,8 +2115,6 @@ interface SemanticCalibrationStatus {
     /** Online only: the label count at which the floor is re-tested next. Each test comes when the labels have grown by half. */
     nextTestAt?: number;
 }
-/** Default {@link AnswerJudge}: the stored answer counts as right only if it is the same answer. */
-declare function sameAnswer$1(stored: CachedAnswer, fresh: CachedAnswer): boolean;
 /**
  * Certifies an acceptance floor from labels you already have (your own traffic,
  * or a labeled set scored with your embedder), with the same test the cache
@@ -2474,7 +2252,7 @@ interface SemanticLookup {
 }
 interface SemanticCacheOptions {
     /**
-     * How text becomes a vector. OPTIONAL: omit it and {@link embebedorLocal} is
+     * How text becomes a vector. OPTIONAL: omit it and {@link localEmbedding} is
      * used — no network, no API key, no per-call cost.
      *
      * It used to be required, and the package shipped NO implementation, so this
@@ -2517,13 +2295,6 @@ interface SemanticCacheOptions {
      */
     calibration?: SemanticCalibrationOptions;
 }
-declare function cosine$1(a: number[], b: number[]): number;
-declare function numericTokens$1(s: string): string[];
-declare function multisetEqual$1(a: string[], b: string[]): boolean;
-declare function words$1(s: string): string[];
-declare function negationCount$1(tokens: string[]): number;
-/** True when two normalized queries agree on the facts an embedding tends to blur. */
-declare function keyTokensMatch$1(a: string, b: string): boolean;
 /**
  * In-memory reference implementation. A production build would back this with a
  * real vector store and metadata filters pushed down
@@ -2557,7 +2328,7 @@ declare class InMemorySemanticCache {
      * and labels nothing: without {@link lookup} and {@link observe} (the layer
      * calls both) the calibration never gets a label and never certifies.
      */
-    find(c: CanonicalRequest$1): Promise<SemanticHit | null>;
+    find(c: CanonicalRequest): Promise<SemanticHit | null>;
     /**
      * The best candidate that passes every hard filter, the similarity floor and
      * the fact guard. With `conAceptacion`, also the confidence floor (the path
@@ -2572,7 +2343,7 @@ declare class InMemorySemanticCache {
      * `agentic` picks the traffic segment: calibrations are kept apart for
      * single-shot and agentic requests.
      */
-    lookup(c: CanonicalRequest$1, opts?: {
+    lookup(c: CanonicalRequest, opts?: {
         agentic?: boolean;
     }): Promise<SemanticLookup>;
     /**
@@ -2620,7 +2391,7 @@ declare class InMemorySemanticCache {
     importCalibration(state: SemanticCalibrationState | string | unknown): void;
     private identidad;
     private estadoDe;
-    store(c: CanonicalRequest$1, answer: CachedAnswer): Promise<void>;
+    store(c: CanonicalRequest, answer: CachedAnswer): Promise<void>;
     /** Drop every record for a knowledge epoch (invalidation on re-index). */
     invalidateEpoch(knowledgeEpoch: string): number;
     get size(): number;
@@ -2723,12 +2494,12 @@ interface ResolverResult {
 }
 interface DeterministicResolver {
     readonly name: string;
-    try(c: CanonicalRequest$1): ResolverResult | null;
+    try(c: CanonicalRequest): ResolverResult | null;
 }
 declare class ResolverRegistry {
     private readonly resolvers;
     register(resolver: DeterministicResolver): this;
-    try(c: CanonicalRequest$1): ResolverResult | null;
+    try(c: CanonicalRequest): ResolverResult | null;
     get size(): number;
 }
 /**
@@ -2737,8 +2508,6 @@ declare class ResolverRegistry {
  * decimals. It never uses eval, and declines anything it does not fully parse.
  */
 declare const arithmeticResolver: DeterministicResolver;
-/** Shunting-yard evaluation of a numeric expression. Returns null on any parse issue. */
-declare function evalArithmetic$1(expr: string): number | null;
 /** A registry seeded with the built-in safe resolvers. */
 declare function createDefaultResolvers(): ResolverRegistry;
 
@@ -2779,7 +2548,7 @@ declare class ExecutionRouter {
         classifier?: LocalClassifier;
         maxRounds?: number;
     });
-    classify(c: CanonicalRequest$1): RouteDecision;
+    classify(c: CanonicalRequest): RouteDecision;
 }
 
 /**
@@ -3005,14 +2774,6 @@ interface ConstraintPlan {
     backend: ConstraintBackendId;
     reason: string;
 }
-/**
- * Resolve a constraint backend from an ordered preference list. Each candidate
- * declares whether it supports the (model, schema) pair; the first supporting
- * backend wins. A conformance suite (see benchmarks/) must confirm that a
- * backend which *claims* support actually enforces every constraint — declared
- * support and empirical coverage are not the same thing.
- */
-declare function resolveConstraintBackend$1(model: string, schema: JsonSchema, backends: ConstraintBackend[]): ConstraintPlan;
 /** A permissive backend used when the provider validates structured output natively. */
 declare function providerNativeBackend(isSupported?: (model: string, schema: JsonSchema) => boolean): ConstraintBackend;
 
@@ -3068,33 +2829,8 @@ interface CompressionPolicyInput {
     /** Do not compress contexts smaller than this. */
     minimumContextTokens?: number;
 }
-/**
- * Decide whether compression is net-positive:
- *   benefit = eliminable - compressorCost - latencyPenalty - qualityRisk
- * Here we model the token side; latency/quality penalties are folded into
- * `compressorCostTokens` and `minimumNetTokens` by the caller.
- */
-declare function shouldCompress$1(input: CompressionPolicyInput): {
-    compress: boolean;
-    reason: string;
-};
-/**
- * Protected-span aware wrapper. Extracts protected spans (replacing them with
- * stable placeholders), runs the inner compressor on the remainder, restores
- * the spans, and verifies every one survived. On any fidelity failure it falls
- * back to the ORIGINAL text — a compressor that drops a negation or a number is
- * worse than no compressor at all.
- */
-declare function withProtectedSpans$1(inner: Compressor): Compressor;
 /** No-op compressor: the safe default until a real compressor sidecar is wired in. */
 declare const noopCompressor: Compressor;
-/**
- * Did the compression keep the facts that must not change? Verifies the RESULT —
- * number/date multiset and negation parity — not the compressor's own claim.
- *
- * @deprecated Internal check of the compression stage; leaves the public API in 0.5.
- */
-declare function compresionFiel(original: string, comprimido: string): boolean;
 
 /** A deterministic, content-only transformation applied to a tool output. */
 type CompactionTransform = "ansi" | "carriage-return" | "json" | "repeated-lines";
@@ -3245,25 +2981,6 @@ interface RetrievalOutcome {
     retrievedCount: number;
     keptCount: number;
 }
-/** Remove exact and near-duplicate docs by normalized text / hash. */
-declare function dedupe$1(docs: EvidenceDoc[]): EvidenceDoc[];
-/**
- * Adaptive-k: choose how many documents to keep based on the score distribution
- * rather than a fixed cutoff. When the top document dominates, keep fewer; when
- * scores are flat, keep up to the cap. This directly trims input tokens.
- */
-declare function adaptiveK$1(docs: EvidenceDoc[], opts?: {
-    maxDocuments: number;
-    minDocuments?: number;
-}): EvidenceDoc[];
-/**
- * Execute a retrieval plan. NO_RETRIEVAL / TOOL_ACTION classes retrieve nothing.
- * Otherwise: retrieve -> dedupe -> rerank -> adaptive-k, bounded by maxRounds.
- */
-declare function executeRetrievalPlan$1(plan: RetrievalPlan, deps: {
-    retriever?: Retriever;
-    reranker?: Reranker;
-}): Promise<RetrievalOutcome>;
 
 /**
  * Provider-neutral generation interface. Every provider (OpenAI-compatible,
@@ -3465,8 +3182,13 @@ interface ProviderResponse {
      * valida, y reintentar con el MISMO tope vuelve a truncarlo — se paga una
      * segunda llamada para obtener el mismo fallo. Sin este campo el layer no
      * podía distinguirlo y hacía exactamente eso.
+     *
+     * A custom provider may return any string here; the layer normalizes it to
+     * `SavingsResult.finishReason`'s five values (`"stop"`, `"length"`,
+     * `"tool_calls"`, `"content_filter"`, `"other"`) and keeps the original in
+     * `finishReasonDetail` when it is not one of them.
      */
-    finishReason?: "stop" | "length" | "content_filter" | "tool_calls" | (string & {});
+    finishReason?: "stop" | "length" | "content_filter" | "tool_calls" | "other" | (string & {});
     /**
      * The provider's own reason behind `finishReason`, when it gave one: Gemini's
      * OpenAI-compatible endpoint stops a filtered answer with
@@ -3596,73 +3318,6 @@ interface Provider {
     generateStream?(req: ProviderRequest, options?: ProviderStreamOptions): AsyncIterable<ProviderStreamEvent>;
 }
 
-/**
- * Offline license verification for the paid SDK.
- *
- * A license is a compact signed token `base64url(header).base64url(payload).
- * base64url(sig)`, the exact format BiVelio's licensing service mints. We verify
- * the Ed25519 signature against BiVelio's public key, decode the payload, and
- * apply the same grace policy the issuer applies. Verification is
- * **browser-safe**: it uses
- * `@noble/curves` (pure JS) + Web primitives (`atob`, `TextEncoder`) rather than
- * `node:crypto`, so the shared entry point still compiles into the browser
- * demo bundle. The private signing key never leaves the control plane; only the
- * public key is embedded here, so a token cannot be forged or tampered with.
- */
-/**
- * `stale` NO es una licencia caducada: es una licencia PAGADA Y VIGENTE que no ha
- * podido confirmar su estado contra el control plane dentro de la ventana de
- * frescura — típicamente porque la salida de red está bloqueada.
- *
- * Existe porque antes ese caso reutilizaba `expired` "para que el union y todos
- * los consumidores siguieran igual", y el ahorro de tipo se pagaba con el peor
- * mensaje posible del producto: a un cliente al día, cuyo único problema era un
- * cortafuegos, se le decía *"license is not active (expired). Renew your
- * subscription"*. Le mandábamos a pagar otra vez algo que ya había pagado, y a
- * mirar donde no estaba el problema.
- *
- * El comportamiento no cambia —sigue siendo fail-closed, `granted()` no lo
- * concede— pero ahora el mensaje puede decir la verdad.
- *
- * `none` means "there is no key": what is left after `normalizeLicenseKey` is
- * EMPTY. It is the same value `LicenseError.status` already carries when nothing
- * was configured, because it is the same problem with the same fix: a
- * `licenseKey: "   "` (an empty environment variable, a blank secret file) used
- * to tell the customer their key "does not verify" — there was no key.
- */
-type LicenseStatus$1 = "active" | "grace" | "expired" | "invalid" | "revoked" | "stale" | "none";
-/** Decoded license token body (mirror of `@bivelio/protocol`'s LicensePayload,
- *  declared locally so the SDK stays free of a control-plane dependency). */
-interface LicenseTokenPayload$1 {
-    jti: string;
-    org_id: string;
-    plan: string;
-    seats: number;
-    features: string[];
-    issued_at: number;
-    expires_at: number;
-    grace_days: number;
-    iss?: string;
-    /** Measure licenses only: the REAL term (see `vigenciaEfectiva` in ./vigencia.ts). */
-    measure_expires_at?: number;
-    measure_grace_days?: number;
-}
-interface ResolvedLicense$1 {
-    status: LicenseStatus$1;
-    payload?: LicenseTokenPayload$1;
-    reason?: string;
-}
-declare const EMBEDDED_PUBLIC_KEY$1: string;
-/** Result of verifying a compact `head.body.sig` token's signature. */
-interface CompactVerification$1 {
-    /** True only if the signature is valid for the given public key. */
-    ok: boolean;
-    /** Decoded header (present when `ok`), e.g. `{ alg, typ, v }`. */
-    header?: Record<string, unknown>;
-    /** Raw payload bytes (present when `ok`); the caller decodes/parses them. */
-    payloadBytes?: Uint8Array;
-    reason?: string;
-}
 /**
  * A pinned set of Ed25519 public keys addressed by key-id (`kid`), enabling signing
  * key ROTATION without breaking already-deployed SDKs: new keys are added as extra
@@ -3676,100 +3331,6 @@ interface LicenseKeyset {
     /** kid used for tokens without a `kid` header (the current/legacy key). */
     defaultKid: string;
 }
-/** Narrow a verification key to a {@link LicenseKeyset} (vs a bare SPKI PEM string). */
-declare function isKeyset$1(k: string | LicenseKeyset): k is LicenseKeyset;
-/**
- * Normalize a license key the way it arrives from the real world — ONCE, at the
- * entry point, never inside the verifier.
- *
- * The README says `export BIVELIO_LICENSE_KEY=…` and dotenv trims, so that path
- * never showed it. But anyone reading the key from a file (`readFileSync`), from
- * a Docker/Kubernetes secret (mounted as a file, trailing newline included) or
- * copying it out of an `Authorization` header handed over a CORRECT key with a
- * `\n` after it, a space before it, quotes around it or a `Bearer ` in front —
- * and the SDK answered "does not verify, check you pasted the WHOLE key". A
- * wrong diagnosis on a good key is the worst possible hint: it sends the
- * customer to look where the problem is not.
- *
- * Worse, it was INCONSISTENT: `tok + "\r\n"` passed and `tok + "\n"` did not,
- * because `atob` ignores ASCII whitespace while the padding was computed from a
- * length that included it, and two extra characters give the same remainder
- * modulo 4 as zero. Behavior that depends on which editor saved the file is not
- * behavior.
- *
- * Only OUTER wrapping is removed — whitespace, one pair of single or double
- * quotes, a `Bearer` prefix in any capitalization — in that order, in one pass.
- * Anything INSIDE the key is left alone: an interior space is not wrapping, it
- * is a broken key, and {@link licenseKeyShape} reports it as `malformed`.
- */
-declare function normalizeLicenseKey$1(raw: string): string;
-/**
- * The SHAPE of an already-normalized key, so the error message can tell apart
- * three problems that used to get the same text:
- *
- *   · `empty`     — there is no key (only whitespace, or nothing at all).
- *   · `malformed` — not even the shape of a token: not three dot-separated
- *                   base64url segments. This is where the `bvk_` service key
- *                   pasted where the license goes lands, as does arbitrary text
- *                   or a token with a line break in the middle.
- *   · `token`     — has the shape; whether it verifies is up to the signature.
- *
- * It is the SAME shape rule {@link verifyCompactSignature} enforces, so that "has
- * the shape" and "verification was attempted" mean the same thing.
- */
-type LicenseKeyShape$1 = "empty" | "malformed" | "token";
-/** Classify a normalized key. See {@link LicenseKeyShape}. */
-declare function licenseKeyShape$1(normalizedKey: string): LicenseKeyShape$1;
-/**
- * Verify the Ed25519 signature of any BV-SALA compact token
- * (`base64url(header).base64url(body).base64url(sig)`, sig over `${head}.${body}`)
- * and return the decoded header + raw payload bytes. This is the single shared
- * crypto primitive: license tokens AND signed revocation lists both go through it,
- * so the browser (`@noble/curves`) and node (`node:crypto`) sides agree byte-for-byte
- * and there is no bespoke per-message verification path to keep in sync. It does NOT
- * interpret the payload or the header `typ` — callers do that.
- *
- * Shape is checked SEGMENT by segment, not just by counting dots. Three parts
- * used to be enough and the rest was left to `atob`, which ignores whitespace:
- * that is how a token with `\r\n` after it verified while the same token with
- * `\n` did not (see {@link normalizeLicenseKey}). The shared primitive must give
- * the same answer to the same input, always; a caller that wants to tolerate
- * wrapping normalizes BEFORE calling.
- */
-declare function verifyCompactSignature$1(token: string, publicKey: string | LicenseKeyset): CompactVerification$1;
-/** Verify a token's signature and decode its payload. Does not check expiry. */
-declare function verifyLicenseToken$1(token: string, publicKey: string | LicenseKeyset): {
-    signatureValid: boolean;
-    payload?: LicenseTokenPayload$1;
-    reason?: string;
-};
-/**
- * Apply the grace + revocation policy to an already-decoded payload → a status.
- * Pure and clock-injected, so a LIVE gate can call it on every request to pick up
- * both revocation (jti in the set) AND expiry transitions without re-verifying the
- * signature. Revocation takes precedence over the time-based status.
- *   jti on the revocation set            → revoked
- *   now <= expires_at                    → active
- *   expires_at < now <= +grace_days      → grace (still entitled)
- *   now > expires_at + grace_days        → expired
- */
-declare function deriveStatus$1(payload: {
-    jti: string;
-    expires_at: number;
-    grace_days: number;
-}, nowSeconds: number, revoked?: ReadonlySet<string>): LicenseStatus$1;
-/**
- * Verify + apply the grace policy → a resolved status. Mirrors what the issuing
- * control plane resolves for the same token. Bad signature / malformed
- * → invalid; otherwise the status comes from {@link deriveStatus}.
- *
- * A public entry point, so it normalizes like the others (gate and refresher):
- * the same key must resolve the same whichever way it comes in.
- */
-declare function resolveLicense$1(token: string, publicKey: string | LicenseKeyset, opts?: {
-    nowSeconds?: number;
-    revoked?: ReadonlySet<string>;
-}): ResolvedLicense$1;
 
 /**
  * Everything the pipeline depends on. All fields except `provider` and
@@ -3801,20 +3362,11 @@ interface SavingsLayerConfig {
      * asks for it nor turns it off gets a one-time console warning (silenced by
      * `silenceWarnings`, or by setting `semanticCacheDefault`, `false` included).
      *
-     * Called `semanticCache` until 0.4.33, so that the store and the per-request
-     * switch (`savings.semanticCache`) no longer share a name. See the 0.5
-     * migration guide.
+     * Called `semanticCache` up to 0.4.x; renamed in 0.5 so that the store and
+     * the per-request switch (`savings.semanticCache`) no longer share a name.
+     * See the 0.5 migration guide.
      */
     semanticCacheStore?: InMemorySemanticCache;
-    /**
-     * @deprecated Renamed {@link semanticCacheStore} in 0.4.33, so that the store
-     * and the per-request switch (`savings.semanticCache`) no longer share a name.
-     * It still works, with a one-time warning (silenced by `silenceWarnings` or
-     * `BVSALA_SILENCE_DEPRECATIONS=1`);
-     * when both are given, `semanticCacheStore` wins. Removed in 0.5: see the 0.5
-     * migration guide.
-     */
-    semanticCache?: InMemorySemanticCache;
     /**
      * Default for `savings.semanticCache` on every request of this layer. A
      * request's own value always wins, `false` included. Unset, the semantic cache
@@ -3926,8 +3478,7 @@ interface SavingsLayerConfig {
     /**
      * Signed BiVelio license token (from your paid subscription). This is the
      * primary way to license the SDK: `createSavingsLayer` verifies it offline
-     * against the embedded public key and builds a {@link licenseGate} from it.
-     * Without a valid, active license (and without {@link devUnlicensed}), the
+     * against the embedded public key. Without a valid, active license the
      * pipeline refuses to run — get a key at https://savings.bivelio.com.
      *
      * Surrounding whitespace (the trailing newline of a file or mounted secret),
@@ -3937,32 +3488,7 @@ interface SavingsLayerConfig {
      */
     licenseKey?: string;
     /**
-     * Override the public key used to verify {@link licenseKey}. An SPKI PEM string,
-     * or a kid-addressed {@link LicenseKeyset} (`{ keys: {kid: pem}, defaultKid }`)
-     * to support signing-key rotation — legacy no-kid tokens use `defaultKid`, and
-     * an unknown kid fails closed. For self-hosting/self-testing; defaults to the
-     * build-embedded BiVelio key.
-     *
-     * @deprecated Has no effect in the published package, which always verifies
-     * against the embedded BiVelio key; passing it there logs a one-time warning
-     * (silenced by `silenceWarnings` or `BVSALA_SILENCE_DEPRECATIONS=1`). Removed
-     * from the public type in 0.5: delete it from your config.
-     */
-    licensePublicKey?: string | LicenseKeyset;
-    /**
-     * Pre-built license gate. **Ignored in the published npm package**, where a
-     * {@link licenseKey} verified against the embedded BiVelio public key is the
-     * only accepted license — passing a gate there logs a warning and changes
-     * nothing. It exists for builds from source (this repo's own tests and
-     * benchmarks), where it takes precedence over {@link licenseKey}.
-     *
-     * @deprecated Ignored in the published package. Removed from the public type
-     * in 0.5; license the layer with {@link licenseKey} or `serviceAccountKey`.
-     */
-    licenseGate?: LicenseGate;
-    /**
-     * OPT-IN online revocation call-home. When set (alongside {@link licenseKey} and
-     * with no explicit {@link licenseGate}), the SDK polls this base's
+     * OPT-IN online revocation call-home. When set (alongside {@link licenseKey}), the SDK polls this base's
      * `GET /licenses/revocations` for a signed revocation list and enforces it via a
      * LIVE gate — a license cancelled mid-term stops working without a restart.
      *
@@ -4067,20 +3593,6 @@ interface SavingsLayerConfig {
     };
     /** Opt-in A/B verification on your own traffic. Off by default. See VerificationSettings. */
     verification?: VerificationSettings;
-    /**
-     * Development-only escape hatch to run ungated, honored ONLY when building from
-     * source (local dev, CI, benchmarks) — also via the `BVSALA_DEV_OPEN=1`
-     * environment variable.
-     *
-     * IMPORTANT: this flag is **disabled in the published npm package**. The
-     * distributed build hard-codes the licensing check on, so setting
-     * `devUnlicensed` (or `BVSALA_DEV_OPEN`) there has NO effect — the layer still
-     * requires a valid `licenseKey`. It is not a way to run the product for free.
-     *
-     * @deprecated No effect in the published package. Removed from the public type
-     * in 0.5: delete it from your config.
-     */
-    devUnlicensed?: boolean;
     /** Per-model cost models; falls back to `defaultCostModel`. */
     costModels?: Record<string, CostModel>;
     defaultCostModel?: CostModel;
@@ -4379,7 +3891,22 @@ interface BatchItemResult<T = unknown> {
     /** Whether the answer met the output contract, when the request had one and no tool was called. */
     contract?: "met" | "missed";
     toolCalls?: ToolCallRequest[];
-    finishReason?: string;
+    /**
+     * Why the provider stopped, as `SavingsResult.finishReason` reports it:
+     * one of the same five values for every provider (`"stop"`, `"length"`,
+     * `"tool_calls"`, `"content_filter"`, `"other"`). A `"length"` answer was cut
+     * off by the output limit. The provider's own value, when it adds something,
+     * is in {@link finishReasonDetail}.
+     */
+    finishReason?: "stop" | "length" | "tool_calls" | "content_filter" | "other";
+    /**
+     * The provider's own reason behind `finishReason`, with the same rule as
+     * `SavingsResult.finishReasonDetail`: the refusal category or the
+     * provider's literal (`"stop_sequence"`, `"model_context_window_exceeded"`,
+     * `"pause_turn"`…), absent for the plain spellings (`end_turn`, `stop`,
+     * `max_tokens`, `tool_use`…).
+     */
+    finishReasonDetail?: string;
     /** The provider's own counter for this request. */
     usage?: ProviderUsage;
     /** The processing tier this result was billed at, as read from the provider. */
@@ -4514,81 +4041,6 @@ declare function registerExactTokenizer(registro: {
 }, opciones?: {
     codificaciones?: readonly string[];
 }): Promise<boolean>;
-/**
- * @deprecated Spanish-named aliases kept for one minor version. Use
- * {@link MODELS_BY_ENCODING} and {@link registerExactTokenizer}. The public
- * surface of this package is English-only; these leaked and are on the way out.
- */
-declare const MODELOS_POR_CODIFICACION: Readonly<Record<string, readonly string[]>>;
-/** @deprecated Use {@link registerExactTokenizer}. */
-declare const registrarTokenizadorExacto: typeof registerExactTokenizer;
-
-/**
- * Exact-cache key.
- *
- * A hash of the user text alone is NOT a safe cache key. Two requests may share
- * text but differ in tenant, permission scope, model version, system prompt,
- * schema, toolset, tool-loop fields (tool calls, the call a tool result answers,
- * its error flag), locale, sampling, knowledge epoch or safety policy — any of
- * which can change the correct answer. The exact key includes all of them.
- *
- * `servedModel` CARRIES THE MODEL THAT WILL ACTUALLY ANSWER, not the one the
- * caller named, and that distinction is what stops the model cascade from
- * poisoning the cache. Routing to a cheap model and storing its answer under the
- * STRONG model's key would hand that answer to the next request — one with the
- * cascade off, or from another process entirely — while the report claims the
- * requested model served it. That is a silent model swap, deferred, which is the
- * worst kind of silent model swap.
- *
- * `modelFamily` is derived from whichever id is used, for the same reason: it is
- * part of the identity of whoever answers.
- */
-declare function buildExactKey$1(c: CanonicalRequest$1, opciones?: {
-    servedModel?: string;
-}): string;
-/**
- * In-flight (singleflight) key. Coalesces concurrent, equivalent requests onto
- * a single execution. It can be slightly looser than the exact key (it does not
- * need freshness bucketing, since the work is happening *now*), but must still
- * isolate tenants and permissions.
- *
- * `schema` is REQUIRED here, even though `contract` is already part of the key:
- * two concurrent calls that differ only in `schemaVersion` are different
- * requests and must never coalesce. All three doors — exact key, in-flight key
- * and semantic filter — key on the same identity, by design.
- *
- * `servedModel` CARRIES THE MODEL THAT WILL ACTUALLY ANSWER, for the same reason
- * {@link buildExactKey} does, and this door is the one that leaks FIRST. The
- * exact key protects the answer AFTER it is stored; this one protects it while
- * it is still being written, and a follower does not re-run anything — it is
- * handed the leader's text verbatim.
- *
- * The leak it closes, measured: two concurrent requests for the same text, one
- * with `savings.modelCascade: "on"` and one with the cascade OFF, coalesced onto
- * one execution. The routed leader called the cheap model, and the caller who
- * never opted in — whose own decision was "do not route" — was served the cheap
- * model's answer. Same shape inside a single uniformly-configured deployment:
- * the automatic shut-off can latch between two concurrent requests, so one is
- * routed and the other is not, and coalescing hands one of them the other's
- * model. Whoever the layer decided must go to the strong model has to actually
- * go there.
- *
- * `deferrable` CLOSES THE SAME HOLE ON THE LATENCY AXIS, which is why it belongs
- * here and NOT in {@link buildExactKey}. The two doors do not risk the same
- * thing: the exact key hands out an answer that is ALREADY WRITTEN, where the
- * processing tier only picked a queue and changes neither the text nor the wait
- * of whoever receives it. This one binds a follower to a call that HAS NOT COME
- * BACK YET, so it imposes the leader's wait on them.
- *
- * A caller who never set `savings.deferrable` must not be made to wait for one
- * who did.
- */
-declare function buildInflightKey$1(c: CanonicalRequest$1, opciones?: {
-    servedModel?: string;
-    deferrable?: boolean;
-    toolDiscovery?: boolean;
-    maxOutputTokens?: number;
-}): string;
 
 /**
  * Policy engine: risk and freshness.
@@ -4650,50 +4102,7 @@ interface Policy {
     sensitiveTopic: "self-harm" | "mental-health-crisis" | "medical-emergency" | null;
     reasons: string[];
 }
-declare function evaluatePolicy$1(c: CanonicalRequest$1): Policy;
 
-/**
- * A LOCAL EMBEDDING FUNCTION — no network, no API key, no per-call cost.
- *
- * The semantic cache needs a way to compare meaning. Wiring it to a hosted
- * embedding endpoint is the obvious answer and a bad default: it makes every
- * miss cost a network round-trip and money, and it forces a provider choice onto
- * a customer who never asked for one. So the cache shipped with no embedder at
- * all, which means it never matched anything.
- *
- * This is the third option: a deterministic bag-of-features vector computed in
- * microseconds from the text itself. It captures LEXICAL closeness — the same
- * question asked with different filler words — which is the overwhelming majority
- * of real repeat traffic. It does NOT capture deep semantics: "what is a CDN"
- * and "explain content delivery networks" share no tokens and will not match.
- *
- * WHY THAT IS SAFE, AND THIS IS THE PART THAT MATTERS. A semantic cache's one
- * unforgivable failure is confidently answering a NEARBY BUT DIFFERENT question.
- * This vector never gets the chance: the cache runs a deterministic precision
- * guard beneath the embedding — matching numbers and dates, negation parity and a
- * content-overlap floor — and every isolation field (tenant, principal, model,
- * schema, attached data) is a hard filter applied before similarity is scored.
- * A weak embedding can only cause a MISS, which costs one real call. It cannot
- * produce a wrong answer.
- *
- * Pass your own `embed` to {@link InMemorySemanticCache} to get true semantics.
- * This one exists so that the feature is worth something with zero setup.
- */
-/**
- * The vector for a piece of text.
- *
- * Features are unigrams plus adjacent bigrams: bigrams are what keep word ORDER
- * from being free. Without them "does A cause B" and "does B cause A" are the
- * same point in space.
- *
- * Term weight is sublinear (`1 + log tf`) so a word repeated forty times does not
- * dominate the direction of the vector, and the result is L2-normalised so cosine
- * similarity is a plain dot product.
- *
- * @deprecated Internal detail of {@link localEmbedding}; leaves the public API in 0.5.
- * Use {@link localEmbedding}, which returns the same vector.
- */
-declare function vectorLocal(texto: string): number[];
 /**
  * The default embedding function for {@link InMemorySemanticCache}.
  *
@@ -4701,8 +4110,6 @@ declare function vectorLocal(texto: string): number[];
  * adds no await and no latency to the request path.
  */
 declare function localEmbedding(text: string): number[];
-/** @deprecated Spanish name, removed in 0.5. Use {@link localEmbedding}. */
-declare function embebedorLocal(texto: string): number[];
 /**
  * The similarity floor this embedder is calibrated for.
  *
@@ -4713,8 +4120,6 @@ declare function embebedorLocal(texto: string): number[];
  * populations stop being separable.
  */
 declare const LOCAL_EMBEDDING_THRESHOLD = 0.72;
-/** @deprecated Spanish name, removed in 0.5. Use {@link LOCAL_EMBEDDING_THRESHOLD}. */
-declare const UMBRAL_LOCAL = 0.72;
 
 /**
  * Conversational memory by deltas.
@@ -4745,17 +4150,6 @@ declare function emptyState(): ConversationState;
  * The merge is deterministic and side-effect-free (returns a new object).
  */
 declare function mergeDelta(state: ConversationState, delta: MemoryDelta): ConversationState;
-/**
- * Project the state + recent turns into the minimal message set to send. Only
- * relevant, non-empty state is serialized, then the last `rawTurns` messages,
- * then the current message.
- */
-declare function projectContext$1(opts: {
-    state: ConversationState;
-    history: ChatMessage[];
-    current: ChatMessage;
-    rawTurns?: number;
-}): ChatMessage[];
 declare function renderState(state: ConversationState): string;
 
 /**
@@ -4787,8 +4181,6 @@ interface CompactRowsOptions {
     indent?: string;
 }
 declare function encodeCompactRows(value: unknown, options?: CompactRowsOptions): string;
-/** Return shared column names iff every element is a flat, scalar-valued object with identical keys. */
-declare function uniformColumns$1(arr: unknown[]): string[] | null;
 /**
  * The instruction a model needs to read the compact encoder. Kept short on purpose — this text
  * is the per-request overhead the selector must amortize before it wins.
@@ -4987,30 +4379,6 @@ interface OpenAICompatibleOptions {
      */
     sendPromptCacheKey?: boolean;
 }
-/**
- * ¿Cumple este esquema las reglas del modo `strict` de OpenAI?
- *
- * Son dos, y aplican a TODO objeto anidado:
- *   1. `additionalProperties: false`.
- *   2. `required` tiene que listar TODAS las claves de `properties` — no solo
- *      las obligatorias. En modo strict no existe el campo opcional.
- *
- * Un esquema perfectamente válido como
- *   { properties: { answer, refs }, required: ["answer"] }
- * incumple la segunda y el proveedor responde 400. No es un fallo del cliente:
- * es una restricción de OpenAI que JSON Schema no tiene.
- */
-declare function isStrictCompatible$1(schema: unknown): boolean;
-/**
- * Decides the id that travels on the wire. Explicit option wins; otherwise the
- * endpoint decides: OpenAI/Azure and Google's own Gemini API get the bare id,
- * everything else keeps what the caller wrote. Exported for the pipeline-level test that pins the README
- * snippet; not part of the public API.
- *
- * @deprecated Not part of the public API; leaves it in 0.5. Use the adapter option
- * `stripProviderPrefix`.
- */
-declare function quitarPrefijoDeProveedor(model: string, explicito: boolean | undefined, baseUrl: string): string;
 declare class OpenAICompatibleProvider implements Provider {
     readonly id: string;
     private readonly opts;
@@ -5257,21 +4625,6 @@ declare class GoogleProvider implements Provider {
     private cerrar;
 }
 
-/**
- * Trace collector: an ordered log of pipeline decisions with elapsed timings.
- * Every stage records why it did what it did, so a SavingsReport can be audited
- * and so ablations can attribute savings to a specific module.
- */
-declare class TraceCollector$1 {
-    private readonly events;
-    private readonly start;
-    private readonly now;
-    constructor(now?: () => number);
-    record(stage: string, outcome: string, detail?: Record<string, unknown>): void;
-    elapsed(): number;
-    all(): TraceEvent[];
-}
-
 /**
  * Thrown when the pipeline is asked to run without a valid, active BiVelio
  * license. This is the fail-closed contract: a paid SDK must refuse to operate
@@ -5295,383 +4648,6 @@ declare class LicenseError extends Error {
     constructor(message: string, status?: string);
 }
 
-/** Decoded revocation-list body (mirror of @bivelio/protocol's RevocationListPayload). */
-interface RevocationListBody$1 {
-    v: number;
-    asOf: number;
-    seq: number;
-    revoked: string[];
-    iss?: string;
-}
-interface RevocationListVerification$1 {
-    /** True only if signature + typ + monotonic-seq + skew all pass. */
-    ok: boolean;
-    /** List revision (monotonic). -1 when not ok. */
-    seq: number;
-    /** Unix seconds the list was minted. 0 when not ok. */
-    asOf: number;
-    /** Revoked jti in this list (empty when not ok). */
-    revoked: Set<string>;
-    reason?: string;
-}
-/**
- * Verify a signed revocation-list token offline and enforce the client-side
- * policy: (1) valid Ed25519 signature, (2) header `typ` is a revocation list
- * (a license token can't be replayed here), (3) `seq >= minSeq` — reject a
- * stale/rolled-back list, (4) `asOf <= now + maxSkew` — reject a future-dated list.
- * ANY failure returns `ok:false` with an empty set; the poller then keeps its
- * sticky last-known-good (fail-open), never dropping known revocations.
- */
-declare function verifyRevocationList$1(token: string, publicKey: string | LicenseKeyset, opts?: {
-    minSeq?: number;
-    nowSeconds?: number;
-    maxSkewSeconds?: number;
-}): RevocationListVerification$1;
-/**
- * The SDK's live, sticky, monotonic view of the revocation set. It only ever
- * GROWS (a jti, once seen revoked, stays revoked for the process lifetime) and
- * only advances on a higher-or-equal `seq` — so a stale, rolled-back, or empty
- * list can never un-revoke a license. The live gate reads {@link snapshot} on
- * every request; the poller feeds it via {@link apply}.
- */
-declare class RevocationSnapshot$1 {
-    private _seq;
-    private _asOf;
-    private _lastContactAt;
-    private readonly set;
-    get seq(): number;
-    get asOf(): number;
-    /** Unix seconds of the last CONFIRMED authorized heartbeat (a 304 or a verified
-     *  list). Feeds the live gate's online staleness check; 0 until the first one. */
-    get lastContactAt(): number;
-    /** Record a confirmed authorized heartbeat from the control plane (monotonic). */
-    markContact(nowSeconds: number): void;
-    has(jti: string): boolean;
-    /** Live view of the revoked set (the same Set instance — read-only to callers). */
-    snapshot(): ReadonlySet<string>;
-    /**
-     * Merge an already-verified list. Accepts only `seq >= current seq`; UNIONS the
-     * jti (never removes); advances seq/asOf monotonically. Returns whether the
-     * revoked set grew or the revision advanced.
-     */
-    apply(list: {
-        seq: number;
-        asOf: number;
-        revoked: Iterable<string>;
-    }): boolean;
-}
-
-/** A resolved gate that also exposes overall status + a refresh horizon. */
-interface ResolvedGate$1 extends LicenseGate {
-    isActive(): boolean;
-    status(): LicenseStatus$1;
-    /**
-     * Why the key does not open, when it does not; `undefined` when it verified.
-     * Under `status() === "invalid"` it tells a key that does NOT EVEN HAVE THE
-     * SHAPE of a token (`"malformed"`: the `bvk_` pasted where the license goes,
-     * arbitrary text) from a well-formed one whose signature fails (the verifier's
-     * reason, e.g. `"bad signature"`). Both are `invalid` — the stable handle for
-     * branching does not change — but the advice given to the customer differs.
-     */
-    reason(): string | undefined;
-    /** The signed plan while the license grants (`"measure"` = shadow mode). */
-    plan?(): string | undefined;
-    /** Unix seconds after which the online status should be re-checked (poller anchor). */
-    refreshAfter?(): number | undefined;
-}
-/**
- * Build a LIVE license gate from a signed key. The token is verified + decoded
- * ONCE (no per-request Ed25519 verify — hot-path safe); each `isActive()` /
- * `status()` / `isEntitled()` then re-derives the status from the cached payload +
- * the LIVE clock + the (mutable, sticky) {@link RevocationSnapshot}. This is what
- * lets a revocation — or an expiry crossed mid-process — take effect without a
- * restart, while the customer request path never touches the network (it reads an
- * in-memory Set).
- *
- * The offline path passes a fresh, never-updated snapshot (live clock, zero
- * network); the online path passes a snapshot a {@link RevocationPoller} feeds.
- */
-declare function liveGateFromKey$1(licenseKey: string, publicKey: string | LicenseKeyset, opts: {
-    revocation: RevocationSnapshot$1;
-    now?: () => number;
-    /** Online staleness ceiling, seconds. If the gate goes this long without a
-     *  confirmed authorized heartbeat, it stops granting (closes the block-egress
-     *  fail-open). Undefined = check disabled (the offline path never sets it). */
-    maxStalenessSeconds?: number;
-    /** Freshness anchor (unix seconds); defaults to construction time (boot). */
-    startedAt?: number;
-}): ResolvedGate$1;
-
-/**
- * Build the SDK's structural {@link import("../contracts/types.js").LicenseGate}
- * from a signed license key — the object `createSavingsLayer({ licenseKey })`
- * injects on the OFFLINE path (no `licenseServerUrl`).
- *
- * It is a LIVE gate over an EMPTY, never-updated revocation snapshot: zero network,
- * but the clock is LIVE, so a license that crosses its expiry (past grace) stops
- * working mid-process instead of only after a restart — closing the frozen-clock
- * hole the old one-shot resolution left open. `isActive()` is true while active or
- * in grace; `isEntitled` additionally requires the feature in the token's grant.
- * Online revocation is added by pairing {@link liveGateFromKey} with a
- * poller-fed snapshot, which `createSavingsLayer` wires up for you.
- */
-declare function gateFromKey$1(licenseKey: string, publicKey: string | LicenseKeyset, now?: () => number): ResolvedGate$1;
-
-type RevocationPollEvent$1 = {
-    type: "applied";
-    seq: number;
-    size: number;
-} | {
-    type: "unchanged";
-    seq: number;
-} | {
-    type: "rejected";
-    reason: string;
-} | {
-    type: "error";
-    error: unknown;
-} | {
-    type: "no-fetch";
-};
-interface RevocationPollerOptions$1 {
-    /** Absolute URL of GET /licenses/revocations. */
-    url: string;
-    /** SPKI PEM (or a kid-addressed {@link LicenseKeyset}) used to verify the signed
-     *  list — the same key/keyset that verifies license tokens. */
-    publicKeyPem: string | LicenseKeyset;
-    /** The sticky snapshot to feed. */
-    snapshot: RevocationSnapshot$1;
-    /** Anchor cadence to the license refresh horizon (unix seconds). */
-    getRefreshAfter?: () => number | undefined;
-    /** Injectable clock (unix seconds). */
-    nowSeconds?: () => number;
-    /** Injectable fetch (defaults to the global). */
-    fetchImpl?: typeof fetch;
-    pollIntervalSeconds?: number;
-    floorSeconds?: number;
-    maxIntervalSeconds?: number;
-    jitterRatio?: number;
-    maxSkewSeconds?: number;
-    requestTimeoutMs?: number;
-    /** Injectable RNG in [0,1) for jitter. Default Math.random. */
-    random?: () => number;
-    /** Optional observer (telemetry/tests). */
-    onEvent?: (e: RevocationPollEvent$1) => void;
-    /**
-     * Silencia los avisos por consola. Por defecto están ENCENDIDOS, y ese defecto
-     * es la corrección: el sondeo es fail-open, así que un call-home roto no rompe
-     * nada visible — la lista de revocaciones simplemente se queda vieja, en
-     * silencio, durante días. Quien tenga observabilidad propia puede apagarlos;
-     * quien no la tenga, que es casi todo el mundo, se entera igual.
-     */
-    silenceWarnings?: boolean;
-    /**
-     * @deprecated Spanish name that leaked into the public options. Use
-     * {@link RevocationPollerOptions.silenceWarnings}. Honoured for one minor
-     * version so nobody gets their warnings back by surprise.
-     */
-    silenciarAvisos?: boolean;
-}
-/**
- * PURE, unit-testable next-delay policy (seconds). Steady state anchors to the
- * license refresh horizon, clamped to [floor, ceiling] and jittered. After a
- * failure it backs off exponentially (a revocation can't land during an outage
- * anyway), capped, still jittered, never below the floor.
- */
-declare function computeNextDelay$1(opts: {
-    refreshAfter?: number;
-    nowSeconds: number;
-    consecutiveFailures: number;
-    floorSeconds: number;
-    pollIntervalSeconds: number;
-    maxIntervalSeconds: number;
-    jitterRatio: number;
-    random?: () => number;
-}): number;
-declare class RevocationPoller$1 {
-    private readonly opts;
-    private timer;
-    private controller;
-    private consecutiveFailures;
-    private stopped;
-    private started;
-    /** ETag of the last 200; sent as If-None-Match so an unchanged list 304s (cheap no-op). */
-    private lastEtag;
-    /** 304 seguidos; al llegar al tope se fuerza una descarga firmada. */
-    private consecutive304;
-    constructor(opts: RevocationPollerOptions$1);
-    private now;
-    private resolveFetch;
-    /**
-     * One poll cycle: fetch → verify → apply. NEVER throws (fail-open). A reachable
-     * CP resets the failure backoff even if the returned list is rejected — we keep
-     * last-known-good. Only a transport/abort/parse error counts as a failure.
-     */
-    pollOnce(): Promise<void>;
-    /** El interruptor, aceptando todavía el nombre viejo. */
-    private callado;
-    /** Un aviso por causa y por proceso, respetando el interruptor. */
-    private avisar;
-    /**
-     * THE CONTROL PLANE HAS BEEN UNREACHABLE FOR A WHILE.
-     *
-     * Polling is fail-open by design: a network outage must never take down the
-     * customer's production traffic. The trade-off is that an unreachable control
-     * plane has no visible symptom — the license keeps working and requests keep
-     * flowing while the revocation list quietly goes stale.
-     *
-     * So the warning goes to the console BY DEFAULT, not only through the optional
-     * `onEvent` observer. It fires on the third consecutive failure and then every
-     * twelfth: a lost request is normal, a week-long outage must keep reminding.
-     * Silence it with `silenceWarnings` if you have your own observability.
-     */
-    private avisarFalloDeRed;
-    /** Fire one non-blocking warm poll, then self-schedule. Idempotent. */
-    start(): void;
-    private schedule;
-    /** Stop polling and abort any in-flight request. Idempotent. */
-    stop(): void;
-}
-
-interface UsageEmitterOptions$1 {
-    /** Full ingest URL, e.g. `https://savings.bivelio.com/api/metering/usage`. */
-    url: string;
-    /** Service-account key (`bvk_…`) sent as a Bearer token; the server derives the org from it. */
-    serviceAccountKey: string;
-    /** Optional org hint; the server always derives the authoritative org from the key. */
-    orgId?: string;
-    /** Flush cadence in ms (default 15000). */
-    flushIntervalMs?: number;
-    /** Records per POST (default 50). */
-    maxBatch?: number;
-    /** Hard in-memory cap; oldest records are dropped past it (default 1000). */
-    maxBuffer?: number;
-    /** Injectable fetch (defaults to the global). */
-    fetchImpl?: typeof fetch;
-    /** Injectable clock. */
-    now?: () => number;
-    /** Whether this process is a CI/ephemeral runner (excluded from billable seats). */
-    ci?: boolean;
-    /**
-     * Which surface owns this process: the local gateway (`bvsala claude` and
-     * friends) or the in-process SDK. Sent with the version in the batch's
-     * `x-bvsala-version` header (`gateway/0.4.16`, `sdk/0.4.16`) so the control
-     * plane can tell a gateway that lacks a lever from an SDK that never had it.
-     * Defaults to `"sdk"`; the gateway sets `"gateway"`.
-     */
-    surface?: "gateway" | "sdk";
-    /**
-     * Se llama cuando el ingest rechaza la credencial (401/403). Existe para que un
-     * despliegue con observabilidad pueda ENTERARSE: el aviso por consola sirve en
-     * un portátil, pero en producción los `console.warn` no despiertan a nadie, y
-     * este fallo concreto es invisible por definición —deja de llegar uso, que es
-     * exactamente lo mismo que se ve cuando no hay tráfico—.
-     */
-    onAuthError?: (status: number) => void;
-    /**
-     * Silence the console warnings. They are ON by default, exactly as in the
-     * revocation poller: metering is best-effort, so a broken ingest breaks
-     * nothing visible — usage simply stops arriving, silently, and the customer's
-     * dashboard stays empty forever.
-     */
-    silenceWarnings?: boolean;
-}
-/**
- * Buffers usage records (counts, cost figures and opaque identifiers, never
- * content) and flushes them to the control plane. DETECTIVE
- * metering: strictly best-effort and NON-BLOCKING. `record()` is O(1) and
- * `flush()` never throws, so a slow, unreachable, or erroring control plane can
- * neither block nor crash the customer's request path. The buffer is hard-capped
- * so a sustained outage can never grow memory without bound.
- */
-declare class UsageEmitter$1 {
-    private readonly options;
-    private readonly now;
-    private readonly flushIntervalMs;
-    private readonly maxBatch;
-    private readonly maxBuffer;
-    private buffer;
-    private seatId;
-    private ci;
-    private timer;
-    private inflight;
-    /**
-     * CONSECUTIVE failed flushes. Warn on the third and then every twelfth: one
-     * lost batch is normal, a whole hour is not. Same cadence as the revocation
-     * poller on purpose — two different rhythms for the same class of failure
-     * only confuse whoever reads the logs.
-     */
-    private fallosSeguidos;
-    constructor(options: UsageEmitterOptions$1);
-    /** Set the resolved seat identity (derivation is async; records emitted before
-     *  it resolves are stamped at flush time, so none are lost). */
-    setSeat(seatId: string | undefined, ci: boolean): void;
-    /** Record one request's usage. Never blocks; O(1). */
-    record(report: SavingsReport): void;
-    start(): void;
-    /** Stop the timer and attempt a final flush. Safe to call more than once.
-     *  Fire-and-forget: for a CLEAN shutdown that must not lose the last batch,
-     *  use `stopAndDrain` and await it. */
-    stop(): void;
-    /** Clean shutdown: stop the timer and DRAIN the buffer, awaiting each POST.
-     *  Unlike `stop()` (fire-and-forget), this resolves only once the final batch
-     *  has actually been sent, so a graceful exit doesn't lose the last session's
-     *  usage. Bounded, so an unreachable control plane can't block shutdown. */
-    stopAndDrain(maxIntentos?: number): Promise<void>;
-    /**
-     * Flush one batch. Never throws. A transient server error (5xx) or network
-     * failure re-queues the batch (capped); a 4xx (bad request / bad key) drops it,
-     * since retrying a rejected payload would loop forever.
-     */
-    flush(): Promise<void>;
-    private hacerFlush;
-    /**
-     * THE INGEST HAS BEEN FAILING FOR A WHILE AND NOBODY NOTICED.
-     *
-     * Metering is best-effort by design — it must never take down the customer's
-     * request — and that turns every failure into SILENT data loss. Without this,
-     * the only symptom is an empty dashboard, which is indistinguishable from
-     * "you haven't sent anything yet": the symptom points the opposite way from
-     * the cause.
-     *
-     * Warns on the third consecutive failure and then every twelfth. Not the
-     * first — a lost request is normal — and not every one, which becomes noise.
-     */
-    private avisarDeFallo;
-    private requeue;
-    private toWire;
-}
-
-/** True when running in a CI / ephemeral environment. */
-declare function detectCi$1(): boolean;
-/**
- * Derive the machine seat id for a service-account key. Resolves to `undefined`
- * when no machine fingerprint is available (e.g. a browser) — the caller then
- * reports no seatId and the server counts the request under an "unknown" seat.
- * The id (`m-<32 hex>`) matches the control plane's identity charset.
- */
-declare function deriveSeatId$1(serviceAccountKey: string): Promise<string | undefined>;
-
-/** Fully-resolved, non-optional dependency set the pipeline runs against. */
-interface ResolvedDeps$1 {
-    exactCache: ExactCache;
-    semanticCache: InMemorySemanticCache | undefined;
-    singleflight: Singleflight;
-    resolvers: ResolverRegistry;
-    router: ExecutionRouter;
-    toolSelector: ToolSelector;
-    formatSelector: FormatSelector;
-    schemaRegistry: SchemaRegistry;
-    constraintBackends: ConstraintBackend[];
-}
-/**
- * Wire defaults for anything the caller did not supply. Every default is an
- * in-memory, dependency-free implementation, so `createSavingsLayer({ provider,
- * tokenizer })` is a working layer, and each piece can be replaced à la carte.
- */
-declare function resolveDependencies$1(config: SavingsLayerConfig): ResolvedDeps$1;
-
-declare const DEFAULT_MAX_STALENESS_SECONDS$1: number;
 /** The public handle returned by {@link createSavingsLayer}. */
 interface SavingsLayer {
     generate<T = unknown>(request: SavingsRequest): Promise<SavingsResult<T>>;
@@ -5734,7 +4710,7 @@ interface SavingsLayer {
  *   messages: [{ role: "user", content: "What is 2 + 2 * 3?" }],
  * });
  */
-declare function createSavingsLayer(config: Omit<SavingsLayerConfig, "tokenizer"> & Partial<Pick<SavingsLayerConfig, "tokenizer">>): SavingsLayer;
+declare function createSavingsLayer(configPublica: Omit<SavingsLayerConfig, "tokenizer"> & Partial<Pick<SavingsLayerConfig, "tokenizer">>): SavingsLayer;
 
 /**
  * Rejection of `SavingsStream.result` when the stream was aborted before the
@@ -6132,143 +5108,6 @@ declare function toCsv(rows: readonly Readonly<Record<string, CsvValue>>[], opti
     columns?: readonly string[];
 }): string;
 
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-type CanonicalRequest = CanonicalRequest$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const canonicalize: typeof canonicalize$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const modelFamilyOf: typeof modelFamilyOf$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const normalizeText: typeof normalizeText$1;
-/** @deprecated Internal helper, removed from the public API in 0.5. Use `createHash("sha256").update(text).digest("hex")` from `node:crypto`. */
-declare const sha256: typeof sha256$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const stableStringify: typeof stableStringify$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const buildExactKey: typeof buildExactKey$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const buildInflightKey: typeof buildInflightKey$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-type CompactVerification = CompactVerification$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-declare const EMBEDDED_PUBLIC_KEY: typeof EMBEDDED_PUBLIC_KEY$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-type LicenseKeyShape = LicenseKeyShape$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-type LicenseStatus = LicenseStatus$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-type LicenseTokenPayload = LicenseTokenPayload$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-type ResolvedGate = ResolvedGate$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-type ResolvedLicense = ResolvedLicense$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-type RevocationListBody = RevocationListBody$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-type RevocationListVerification = RevocationListVerification$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-type RevocationPollEvent = RevocationPollEvent$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-declare const RevocationPoller: typeof RevocationPoller$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-type RevocationPoller = RevocationPoller$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-type RevocationPollerOptions = RevocationPollerOptions$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-declare const RevocationSnapshot: typeof RevocationSnapshot$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-type RevocationSnapshot = RevocationSnapshot$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-declare const computeNextDelay: typeof computeNextDelay$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-declare const deriveStatus: typeof deriveStatus$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-declare const gateFromKey: typeof gateFromKey$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-declare const isKeyset: typeof isKeyset$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-declare const licenseKeyShape: typeof licenseKeyShape$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-declare const liveGateFromKey: typeof liveGateFromKey$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-declare const normalizeLicenseKey: typeof normalizeLicenseKey$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-declare const resolveLicense: typeof resolveLicense$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-declare const verifyCompactSignature: typeof verifyCompactSignature$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-declare const verifyLicenseToken: typeof verifyLicenseToken$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-declare const verifyRevocationList: typeof verifyRevocationList$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-declare const UsageEmitter: typeof UsageEmitter$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-type UsageEmitter = UsageEmitter$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-type UsageEmitterOptions = UsageEmitterOptions$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-declare const deriveSeatId: typeof deriveSeatId$1;
-/** @deprecated Internal license machinery, removed from the public API in 0.5: no replacement. In the published package a license enters only through `licenseKey` / `serviceAccountKey`. */
-declare const detectCi: typeof detectCi$1;
-/** @deprecated Internal constant, removed from the public API in 0.5. The default is documented on `licenseRevocation.maxStalenessSeconds`; set that option to change it. */
-declare const DEFAULT_MAX_STALENESS_SECONDS: typeof DEFAULT_MAX_STALENESS_SECONDS$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-type ResolvedDeps = ResolvedDeps$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const resolveDependencies: typeof resolveDependencies$1;
-/** @deprecated Internal constant of the reference-commission rule, removed from the public API in 0.5: no replacement (the rule is described in the documentation, not frozen as API). */
-declare const REFERENCE_MAX_AGE_MS: typeof REFERENCE_MAX_AGE_MS$1;
-/** @deprecated Internal constant of the reference-commission rule, removed from the public API in 0.5: no replacement (the rule is described in the documentation, not frozen as API). */
-declare const REFERENCE_MESSAGE_ALLOWANCE: typeof REFERENCE_MESSAGE_ALLOWANCE$1;
-/** @deprecated Internal constant of the reference-commission rule, removed from the public API in 0.5: no replacement (the rule is described in the documentation, not frozen as API). */
-declare const REFERENCE_OUTPUT_SHARE: typeof REFERENCE_OUTPUT_SHARE$1;
-/** @deprecated Internal ledger of the pipeline, removed from the public API in 0.5. Read the four ledgers from the `SavingsReport` each call returns. */
-declare const SavingsLedger: typeof SavingsLedger$1;
-/** @deprecated Internal ledger of the pipeline, removed from the public API in 0.5. Read the four ledgers from the `SavingsReport` each call returns. */
-type SavingsLedger = SavingsLedger$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5. Read the stage traces from `report.traces`. */
-declare const TraceCollector: typeof TraceCollector$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5. Read the stage traces from `report.traces`. */
-type TraceCollector = TraceCollector$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const adaptiveK: typeof adaptiveK$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const dedupe: typeof dedupe$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const executeRetrievalPlan: typeof executeRetrievalPlan$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const cosine: typeof cosine$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const keyTokensMatch: typeof keyTokensMatch$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const multisetEqual: typeof multisetEqual$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const negationCount: typeof negationCount$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const numericTokens: typeof numericTokens$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const words: typeof words$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const sameAnswer: typeof sameAnswer$1;
-/** @deprecated Internal helper, removed from the public API in 0.5. Arithmetic questions are still answered locally by `arithmeticResolver` / `createDefaultResolvers()`. */
-declare const evalArithmetic: typeof evalArithmetic$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const evaluatePolicy: typeof evaluatePolicy$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const isStrictCompatible: typeof isStrictCompatible$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const projectContext: typeof projectContext$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const resolveConstraintBackend: typeof resolveConstraintBackend$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const shouldCompress: typeof shouldCompress$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const withProtectedSpans: typeof withProtectedSpans$1;
-/** @deprecated Internal pipeline helper, removed from the public API in 0.5: no replacement. */
-declare const uniformColumns: typeof uniformColumns$1;
-/** @deprecated Internal helper, removed from the public API in 0.5. Use `crypto.randomUUID()`. */
-declare const randomId: typeof randomId$1;
-
 /**
  * bivelio-savings-layer (BV-SALA)
  * ---------------------------------------------------------------------------
@@ -6292,19 +5131,13 @@ declare const randomId: typeof randomId$1;
  * The list-price catalogue: model id (without vendor prefix) → USD per 1K
  * tokens. `costModels` on the layer config always wins over it.
  */
-declare const PRICE_CATALOG: typeof CATALOGO$1;
+declare const PRICE_CATALOG: typeof CATALOGO;
 /**
  * The list price for a model on the day of the request, or `undefined` when the
  * catalogue has no price for it in force that day.
  */
-declare const listPrice: typeof precioDeLista$1;
+declare const listPrice: typeof precioDeLista;
 /** When the catalogue prices were last checked against the providers' public rates (YYYY-MM-DD). */
-declare const PRICES_AS_OF: typeof PRECIOS_AL_DIA$1;
-/** @deprecated Spanish name, removed in 0.5. Use {@link PRICE_CATALOG}. */
-declare const CATALOGO: typeof CATALOGO$1;
-/** @deprecated Spanish name, removed in 0.5. Use {@link listPrice}. */
-declare const precioDeLista: typeof precioDeLista$1;
-/** @deprecated Spanish name, removed in 0.5. Use {@link PRICES_AS_OF}. */
-declare const PRECIOS_AL_DIA: typeof PRECIOS_AL_DIA$1;
+declare const PRICES_AS_OF: typeof PRECIOS_AL_DIA;
 
-export { type AnswerJudge, type AnswerSource, type AnthropicOptions, AnthropicProvider, BV_SALA_VERSION, type BatchCounts, type BatchItemResult, type BatchJob, type BatchJobStatus, type BatchRequest, type BatchResults, BatchSubmitError, type BatchSubmitOptions, type BatchSummary, type BivelioHandoff, CATALOGO, COMPACTION_TRANSFORMS, COMPACT_ROWS_INSTRUCTION, type CacheBreakpointStore, type CacheOutcome, type CachePlacement, type CacheTier, type CacheWriteReport, type CachedAnswer, type CalibrationObservation, CalibrationStateError, type CanonicalRequest, type ChatMessage, type CompactRowsDecodeOptions, type CompactRowsOptions, type CompactVerification, type CompactionTransform, type CompressionEconomics, type CompressionPolicyInput, type CompressionRequest, type CompressionResult, type Compressor, type ConstraintBackend, type ConstraintBackendId, type ConstraintPlan, type ConversationState, type CostBasis, type CostModel, type CsvValue, DEFAULT_COST_MODEL, DEFAULT_MAX_STALENESS_SECONDS, DEFAULT_MIN_SAVED_CHARS, type DeferredExecutionReport, type DeterministicResolver, EMBEDDED_PUBLIC_KEY, type EmbeddingFn, type EvidenceDoc, type ExactCache, ExactTokenizer, type ExecutionClass, ExecutionRouter, type FocusExportOptions, type FocusRow, type FormatChoice, FormatSelector, type FormatSelectorOptions, type GoogleOptions, GoogleProvider, type GoogleThinkingConfig, type HandoffFact, HeuristicEstimator, InMemoryCacheBreakpointStore, InMemoryExactCache, InMemorySemanticCache, InMemoryToolSetFreezer, type JsonSchema, LOCAL_EMBEDDING_THRESHOLD, type LangfuseGenerationUpdate, LicenseError, type LicenseGate, type LicenseKeyShape, type LicenseKeyset, type LicenseStatus, type LicenseTokenPayload, type LocalClassifier, MODELOS_POR_CODIFICACION, MODELS_BY_ENCODING, type MeasureReport, type MemoryDelta, type MessageRole, MockProvider, type MockProviderOptions, type ModelCascadeMode, type ModelCascadeReport, type ModelCascadeSettings, type NativeSavingsReport, type OpenAICompatibleOptions, OpenAICompatibleProvider, type OutputContract, PRECIOS_AL_DIA, PRICES_AS_OF, PRICE_CATALOG, type Policy, type PriceLeverSaving, type PriceLeverSettings, type ProtectedSpan, type Provider, type ProviderRequest, type ProviderResponse, type ProviderStreamEvent, type ProviderStreamOptions, type ProviderToolSearch, type ProviderToolSearchUsage, type ProviderUsage, QUALITY_STATUSES, type QualityCacheMatch, type QualityCascadeCheck, type QualityLedger, type QualityRepairCheck, type QualitySchemaCheck, type QualityStatus, type QualityTruncationCheck, REFERENCE_MAX_AGE_MS, REFERENCE_MESSAGE_ALLOWANCE, REFERENCE_OUTPUT_SHARE, type ReasoningEffort, type ReferenceCredit, type ReferenceRejection, type RequestContext, type Reranker, type ResolvedDeps, type ResolvedGate, type ResolvedLicense, ResolverRegistry, type ResolverResult, type RetrievalOptions, type RetrievalOutcome, type RetrievalPlan, type Retriever, type RevocationListBody, type RevocationListVerification, type RevocationPollEvent, RevocationPoller, type RevocationPollerOptions, RevocationSnapshot, type RouteDecision, type SavingsAdapterLevers, type SavingsAdapterOptions, type SavingsAttributeOptions, type SavingsBatches, type SavingsCallOptions, type SavingsLayer, type SavingsLayerConfig, SavingsLedger, type SavingsRecord, type SavingsReport, type SavingsReportForExport, type SavingsRequest, type SavingsResult, type SavingsSeriesOptions, type SavingsSeriesRow, type SavingsSpanAttributes, type SavingsSpanOptions, type SavingsStream, type SavingsStreamEvent, type SavingsStreamOptions, type SavingsToggles, type SavingsTracingOptions, SchemaRegistry, type SemanticCacheOptions, type SemanticCalibrationBound, type SemanticCalibrationOptions, type SemanticCalibrationSegmentState, type SemanticCalibrationState, type SemanticCalibrationStatus, type SemanticHit, type SemanticLookup, type SemanticProbe, type SemanticRecord, type SemanticVerifier, type SerializationFormat, type Serializer, Singleflight, type SpanAttributeValue, type SpanLike, type SpanStartOptionsLike, StreamAbortedError, type StreamingReport, type TokenEstimator, TokenizerRegistry, type ToolCallRequest, type ToolDescriptor, type ToolOutputCompaction, type ToolOutputCompactionSettings, type ToolRisk, type ToolSearchMode, type ToolSearchOptions, type ToolSelection, ToolSelector, type ToolSelectorOptions, type ToolSetFreezer, TraceCollector, type TraceEvent, type TracerLike, type TracerProviderLike, UMBRAL_LOCAL, UsageEmitter, type UsageEmitterOptions, type ValidationError, type ValidationResult, type VerificationCheck, type VerificationExclusion, type VerificationOutcome, type VerificationReport, type VerificationSettings, type VerificationSide, type VerificationStratum, adaptiveK, arithmeticResolver, buildExactKey, buildInflightKey, calibrateSemanticFloor, canonicalize, compactRowsSerializer, compactToolOutput, compresionFiel, computeNextDelay, cosine, createCompactRowsSerializer, createDefaultResolvers, createHandoff, createMockProvider, createSavingsLayer, csvSerializer, decodeCompactRows, dedupe, defaultTokenizerRegistry, deriveSeatId, deriveStatus, detectCi, embebedorLocal, emptyState, encodeCompactRows, evalArithmetic, evaluatePolicy, executeRetrievalPlan, gateFromKey, heuristicEstimator, instrumentSavingsLayer, isKeyset, isStrictCompatible, jsonCompactSerializer, keyTokensMatch, licenseKeyShape, listPrice, liveGateFromKey, localEmbedding, mergeDelta, modelFamilyOf, multisetEqual, negationCount, noopCompressor, normalizeLicenseKey, normalizeText, numericTokens, precioDeLista, projectContext, providerNativeBackend, quitarPrefijoDeProveedor, randomId, recordSavingsOnSpan, registerExactTokenizer, registrarTokenizadorExacto, renderState, resolveConstraintBackend, resolveDependencies, resolveLicense, sameAnswer, savingsSpanAttributes, sha256, shouldCompress, stableStringify, toCsv, toFocusRows, toLangfuseGeneration, toSavingsSeries, uniformColumns, validateAgainstSchema, vectorLocal, verifyCompactSignature, verifyLicenseToken, verifyRevocationList, withProtectedSpans, words };
+export { type AnswerJudge, type AnswerSource, type AnthropicOptions, AnthropicProvider, BV_SALA_VERSION, type BatchCounts, type BatchItemResult, type BatchJob, type BatchJobStatus, type BatchRequest, type BatchResults, BatchSubmitError, type BatchSubmitOptions, type BatchSummary, type BivelioHandoff, COMPACTION_TRANSFORMS, COMPACT_ROWS_INSTRUCTION, type CacheBreakpointStore, type CacheOutcome, type CachePlacement, type CacheTier, type CacheWriteReport, type CachedAnswer, type CalibrationObservation, CalibrationStateError, type ChatMessage, type CompactRowsDecodeOptions, type CompactRowsOptions, type CompactionTransform, type CompressionEconomics, type CompressionPolicyInput, type CompressionRequest, type CompressionResult, type Compressor, type ConstraintBackend, type ConstraintBackendId, type ConstraintPlan, type ConversationState, type CostBasis, type CostModel, type CsvValue, DEFAULT_COST_MODEL, DEFAULT_MIN_SAVED_CHARS, type DeferredExecutionReport, type DeterministicResolver, type EmbeddingFn, type EvidenceDoc, type ExactCache, ExactTokenizer, type ExecutionClass, ExecutionRouter, type FocusExportOptions, type FocusRow, type FormatChoice, FormatSelector, type FormatSelectorOptions, type GoogleOptions, GoogleProvider, type GoogleThinkingConfig, type HandoffFact, HeuristicEstimator, InMemoryCacheBreakpointStore, InMemoryExactCache, InMemorySemanticCache, InMemoryToolSetFreezer, type JsonSchema, LOCAL_EMBEDDING_THRESHOLD, type LangfuseGenerationUpdate, LicenseError, type LicenseGate, type LicenseKeyset, type LocalClassifier, MODELS_BY_ENCODING, type MeasureReport, type MemoryDelta, type MessageRole, MockProvider, type MockProviderOptions, type ModelCascadeMode, type ModelCascadeReport, type ModelCascadeSettings, type NativeSavingsReport, type OpenAICompatibleOptions, OpenAICompatibleProvider, type OutputContract, PRICES_AS_OF, PRICE_CATALOG, type Policy, type PriceLeverSaving, type PriceLeverSettings, type ProtectedSpan, type Provider, type ProviderRequest, type ProviderResponse, type ProviderStreamEvent, type ProviderStreamOptions, type ProviderToolSearch, type ProviderToolSearchUsage, type ProviderUsage, QUALITY_STATUSES, type QualityCacheMatch, type QualityCascadeCheck, type QualityLedger, type QualityRepairCheck, type QualitySchemaCheck, type QualityStatus, type QualityTruncationCheck, type ReasoningEffort, type ReferenceCredit, type ReferenceRejection, type RequestContext, type Reranker, ResolverRegistry, type ResolverResult, type RetrievalOptions, type RetrievalOutcome, type RetrievalPlan, type Retriever, type RouteDecision, type SavingsAdapterLevers, type SavingsAdapterOptions, type SavingsAttributeOptions, type SavingsBatches, type SavingsCallOptions, type SavingsLayer, type SavingsLayerConfig, type SavingsRecord, type SavingsReport, type SavingsReportForExport, type SavingsRequest, type SavingsResult, type SavingsSeriesOptions, type SavingsSeriesRow, type SavingsSpanAttributes, type SavingsSpanOptions, type SavingsStream, type SavingsStreamEvent, type SavingsStreamOptions, type SavingsToggles, type SavingsTracingOptions, SchemaRegistry, type SemanticCacheOptions, type SemanticCalibrationBound, type SemanticCalibrationOptions, type SemanticCalibrationSegmentState, type SemanticCalibrationState, type SemanticCalibrationStatus, type SemanticHit, type SemanticLookup, type SemanticProbe, type SemanticRecord, type SemanticVerifier, type SerializationFormat, type Serializer, Singleflight, type SpanAttributeValue, type SpanLike, type SpanStartOptionsLike, StreamAbortedError, type StreamingReport, type TokenEstimator, TokenizerRegistry, type ToolCallRequest, type ToolDescriptor, type ToolOutputCompaction, type ToolOutputCompactionSettings, type ToolRisk, type ToolSearchMode, type ToolSearchOptions, type ToolSelection, ToolSelector, type ToolSelectorOptions, type ToolSetFreezer, type TraceEvent, type TracerLike, type TracerProviderLike, type ValidationError, type ValidationResult, type VerificationCheck, type VerificationExclusion, type VerificationOutcome, type VerificationReport, type VerificationSettings, type VerificationSide, type VerificationStratum, arithmeticResolver, calibrateSemanticFloor, compactRowsSerializer, compactToolOutput, createCompactRowsSerializer, createDefaultResolvers, createHandoff, createMockProvider, createSavingsLayer, csvSerializer, decodeCompactRows, defaultTokenizerRegistry, emptyState, encodeCompactRows, heuristicEstimator, instrumentSavingsLayer, jsonCompactSerializer, listPrice, localEmbedding, mergeDelta, noopCompressor, providerNativeBackend, recordSavingsOnSpan, registerExactTokenizer, renderState, savingsSpanAttributes, toCsv, toFocusRows, toLangfuseGeneration, toSavingsSeries, validateAgainstSchema };