tempest-react-sdk 0.81.0 → 0.82.0

package/dist/sw.d.ts

@@ -336,7 +336,9 @@ export declare interface InstallBackgroundSyncOptions {
  *
  * A click on an action button opens that action's URL (see
  * {@link PushNotificationAction.url}); a click on the body, or on an action
- * without one, opens the top-level `url`.
+ * without one, opens the top-level `url`. Only a same-origin `http(s)` URL is
+ * opened by default; see {@link InstallNotificationClickHandlerOptions.resolveUrl}
+ * for opting into another origin.
  */
 export declare function installNotificationClickHandler(options?: InstallNotificationClickHandlerOptions): void;
 
@@ -345,6 +347,12 @@ export declare interface InstallNotificationClickHandlerOptions {
      * Resolve the destination URL from the notification data and the clicked
      * action id (`""` or `undefined` for a click on the body). Default: the
      * action's entry in `data.actionUrls`, else `data.url`, else `"/"`.
+     *
+     * The default only ever opens a page of the worker's **own origin**: a URL on
+     * another origin, or with a scheme other than `http:`/`https:`, becomes `"/"`.
+     * A resolver of your own is how an app opts into another origin — what it
+     * returns may point anywhere `http:`/`https:`, and any other scheme still
+     * becomes `"/"`.
      */
     resolveUrl?: (data: unknown, action: string | undefined) => string;
 }
@@ -389,8 +397,9 @@ export declare interface InstallPrecacheOptions {
  * Every `NotificationOptions` member the payload carries is forwarded —
  * `actions`, `requireInteraction`, `vibrate`, `silent`, `renotify`,
  * `timestamp`, `dir` and `lang` alongside `body`, `icon`, `badge`, `image` and
- * `tag`. `data` receives the top-level `url` (default `"/"`), the payload's
- * own `data`, and `actionUrls` mapping each action id to its `url`.
+ * `tag`. `data` receives the payload's own `data`, then the top-level `url`
+ * (default `"/"`) and `actionUrls` mapping each action id to its `url` — in that
+ * order, so `payload.data` cannot override where a click goes.
  */
 export declare function installPushHandler(options?: InstallPushHandlerOptions): void;
 
@@ -414,9 +423,10 @@ export declare interface InstallPushHandlerOptions {
  * `respondWith`), so a later {@link installPrecache} can handle them.
  *
  * A request carrying a credential header (`Authorization`, `X-Api-Key`, … — the
- * route's `credentialHeaders`) is also left untouched unless its route sets
+ * route's `credentialHeaders`), or a CORS `fetch` sent with
+ * `credentials: "include"`, is also left untouched unless its route sets
  * `cacheAuthenticated`, and a response with `Cache-Control: no-store`
- * (or `private`, without that opt-in) is served but never stored. `maxAgeSeconds`
+ * (or `private`, without that opt-in) or `Vary: *` is served but never stored. `maxAgeSeconds`
  * applies to every read, including the offline fallback: an expired entry is a
  * miss, never a last resort. Under `rangeRequests`, the full resource is cached
  * once (the `Range` header stripped) and each range is sliced from it.
@@ -457,7 +467,13 @@ export declare function installSkipWaitingListener(): void;
  *
  * Pass `getPrincipal` to bind each entry to whoever queued it: the worker asks
  * the issuing tab right after queueing, asks every tab again before replaying,
- * and drops the entry when no signed-in tab is its owner.
+ * and drops the entry when no signed-in tab is its owner. Without it every
+ * entry is `"unbound"` and replays with whichever session answers — the next
+ * account's bearer included, unless `clearSessionData()` ran on logout — so a
+ * development build warns once per call (#531). Treating an unbound
+ * credentialed entry as `"unknown"` instead would change nothing: an
+ * `"unknown"` entry replays when no answering tab reports a principal, which is
+ * exactly the page that omits `getPrincipal`.
  *
  * @param getHeaders - Resolves the headers for one queued request at replay.
  * @param options - `getPrincipal` binds entries to their owner.
@@ -548,7 +564,14 @@ export declare interface PushPayload {
     dir?: NotificationDirection;
     /** BCP 47 language tag of the title and body. */
     lang?: string;
-    /** Arbitrary extra data forwarded to `event.notification.data`. */
+    /**
+     * Arbitrary extra data forwarded to `event.notification.data`.
+     *
+     * Its `url` and `actionUrls` keys are overwritten by the top-level `url` and
+     * the actions' `url`s: those two keys decide where a click goes, and a
+     * free-form bag — often filled from the record that triggered the push — must
+     * not be able to redirect it.
+     */
     data?: Record<string, unknown>;
 }
 
@@ -679,8 +702,9 @@ export declare interface RuntimeRoute {
      */
     rangeRequests?: boolean;
     /**
-     * Opt in to caching requests that carry a credential header (see
-     * `credentialHeaders`), and responses marked `Cache-Control: private`.
+     * Opt in to caching credentialed requests — those that carry a credential
+     * header (see `credentialHeaders`) or are a CORS-mode `fetch` sent with
+     * `credentials: "include"` — and responses marked `Cache-Control: private`.
      * Default `false`.
      *
      * Cache Storage is keyed by URL and outlives the session: without this
@@ -697,11 +721,18 @@ export declare interface RuntimeRoute {
      * `clearSessionData()` on logout, which removes the entries themselves.
      * `Cache-Control: no-store` is never cached, opt-in or not.
      *
-     * A cookie session is invisible here: the worker cannot see `Cookie`, and
-     * gating on `request.credentials` would switch off caching for every
-     * same-origin image and font. A cookie-authenticated API stays private only
+     * A cookie session is only partly visible here: the worker cannot see
+     * `Cookie`. What it can see is a `fetch` that asked for cookies explicitly —
+     * `mode: "cors"` with `credentials: "include"`, which is what `ApiClient`
+     * sends under `withCredentials` — and since 0.82.0 that request is treated
+     * as credentialed too (#531). `<img>`, `<audio>` and `<video>` also send
+     * `credentials: "include"`, but in `no-cors` mode, so they keep caching. A
+     * plain same-origin `fetch` (`credentials: "same-origin"`, the default)
+     * carries the cookie without saying so; such an API stays private only
      * through `Cache-Control: private`/`no-store` on its responses and
-     * `clearSessionData()` on logout.
+     * `clearSessionData()` on logout. Opting a cookie request in caches it by
+     * URL — there is no credential header to vary on — so `clearSessionData()`
+     * is then its only isolation.
      */
     cacheAuthenticated?: boolean;
     /**

package/dist/tabular/compact.cjs

@@ -1,2 +1,2 @@
-const e=require("./exceptions.cjs");var t=`TMC1`,n=class n{header;arrays;info;constructor(e,t,n){this.header=e,this.arrays=t,this.info=n}static async create(i,a){let o=await r(i,a),s=new Uint8Array(o);if(String.fromCharCode(...s.subarray(0,4))!==t)throw new e.CompactFormatError(`This is not a compact model file: it does not start with "${t}". A .onnx file goes through TabularPredictor instead.`);let c=new DataView(o).getUint32(4,!0),l=JSON.parse(new TextDecoder().decode(s.subarray(8,8+c)));if(l.schema_version>1)throw new e.CompactFormatError(`This file uses compact layout ${l.schema_version}, and this SDK understands 1. Upgrade tempest-react-sdk before serving it.`);let u={},d=8+c;for(let e of l.sections)u[e.name]=e.dtype===`float32`?new Float32Array(o,d,e.length):new Int32Array(o,d,e.length),d+=e.length*4;return new n(l,u,{kind:l.kind,classes:l.classes,numFeatures:l.n_features,featureNames:l.feature_names,numTrees:l.n_trees??0,isClassifier:l.task===`classification`,estimator:l.estimator})}async predict(e){let t=i(e,this.header.n_features),n=performance.now(),r=[];for(let n of e){let e=this.preprocess(n,t);r.push(this.header.kind===`linear`?this.linearScores(e):this.treeScores(e))}let{labels:a,probabilities:o}=this.finish(r);return{labels:a,probabilities:o,numRows:e.length,ms:performance.now()-n}}async dispose(){}preprocess(e,t){let n=this.header.preprocess;if(n===null)return e;let r=Array(t);for(let i=0;i<t;i+=1)r[i]=(e[i]-n.offset[i])/n.scale[i];return r}linearScores(e){let t=this.arrays.coef,n=this.arrays.intercept,r=this.header.n_features,i=this.header.n_outputs,a=Array(i);for(let o=0;o<i;o+=1){let i=n[o],s=o*r;for(let n=0;n<r;n+=1)i+=t[s+n]*e[n];a[o]=i}return a}treeScores(e){let t=this.arrays.node_feature,n=this.arrays.node_threshold,r=this.arrays.node_left,i=this.arrays.node_right,a=this.arrays.leaf_value,o=this.arrays.tree_offset,s=this.header.n_outputs,c=o.length-1,l=Array(s).fill(0);for(let u=0;u<c;u+=1){let c=o[u],d=t[c];for(;d>=0;)c=Math.fround(e[d])<=n[c]?r[c]:i[c],d=t[c];let f=(-1-d)*s;for(let e=0;e<s;e+=1)l[e]+=a[f+e]}for(let e=0;e<s;e+=1)l[e]=l[e]/c;return l}finish(e){let t=this.header.link,n=this.header.classes;if(t===`identity`)return{labels:e.map(e=>e[0]),probabilities:[]};let r=e.map(e=>{if(t===`sigmoid`){let t=1/(1+Math.exp(-e[0]));return[1-t,t]}if(t===`softmax`){let t=Math.max(...e),n=e.map(e=>Math.exp(e-t)),r=n.reduce((e,t)=>e+t,0);return n.map(e=>e/r)}let n=e.reduce((e,t)=>e+t,0);return n===0?[...e]:e.map(e=>e/n)}),i=this.header.class_type!==`str`;return{labels:r.map(e=>{let t=0;for(let n=1;n<e.length;n+=1)e[n]>e[t]&&(t=n);let r=n[t];return r===void 0?t:i?Number(r):r}),probabilities:r}}};async function r(t,n){if(typeof t==`string`){let r;try{r=await fetch(t,n)}catch(n){throw new e.ModelFetchError(`Could not download the model: ${t}`,{cause:n})}if(!r.ok)throw new e.ModelFetchError(`Could not download the model: ${r.status} ${r.statusText}`);return await r.arrayBuffer()}return t instanceof Uint8Array?t.buffer.slice(t.byteOffset,t.byteOffset+t.byteLength):t}function i(t,n){if(!Array.isArray(t)||t.length===0)throw new e.FeatureShapeError(`predict() needs at least one row, shaped [[f1, f2, ...]].`);let r=t[0]?.length??0,i=t.findIndex(e=>e.length!==r);if(i!==-1)throw new e.FeatureShapeError(`All rows must have the same width; row ${i} has ${t[i]?.length} values, expected ${r}.`);if(r!==n)throw new e.FeatureShapeError(`The model expects ${n} features per row, got ${r}.`);return r}exports.CompactPredictor=n;
+const e=require("./exceptions.cjs");var t=`TMC1`,n=class n{header;arrays;info;constructor(e,t,n){this.header=e,this.arrays=t,this.info=n}static async create(a,o){let s=await r(a,o),c=new Uint8Array(s);if(String.fromCharCode(...c.subarray(0,4))!==t)throw new e.CompactFormatError(`This is not a compact model file: it does not start with "${t}". A .onnx file goes through TabularPredictor instead.`);let l=new DataView(s).getUint32(4,!0),u=JSON.parse(new TextDecoder().decode(c.subarray(8,8+l)));if(u.schema_version>1)throw new e.CompactFormatError(`This file uses compact layout ${u.schema_version}, and this SDK understands 1. Upgrade tempest-react-sdk before serving it.`);let d={},f=8+l;for(let e of u.sections)d[e.name]=e.dtype===`float32`?new Float32Array(s,f,e.length):new Int32Array(s,f,e.length),f+=e.length*4;return u.kind===`tree_ensemble`&&i(u,d),new n(u,d,{kind:u.kind,classes:u.classes,numFeatures:u.n_features,featureNames:u.feature_names,numTrees:u.n_trees??0,isClassifier:u.task===`classification`,estimator:u.estimator})}async predict(e){let t=a(e,this.header.n_features),n=performance.now(),r=[];for(let n of e){let e=this.preprocess(n,t);r.push(this.header.kind===`linear`?this.linearScores(e):this.treeScores(e))}let{labels:i,probabilities:o}=this.finish(r);return{labels:i,probabilities:o,numRows:e.length,ms:performance.now()-n}}async dispose(){}preprocess(e,t){let n=this.header.preprocess;if(n===null)return e;let r=Array(t);for(let i=0;i<t;i+=1)r[i]=(e[i]-n.offset[i])/n.scale[i];return r}linearScores(e){let t=this.arrays.coef,n=this.arrays.intercept,r=this.header.n_features,i=this.header.n_outputs,a=Array(i);for(let o=0;o<i;o+=1){let i=n[o],s=o*r;for(let n=0;n<r;n+=1)i+=t[s+n]*e[n];a[o]=i}return a}treeScores(e){let t=this.arrays.node_feature,n=this.arrays.node_threshold,r=this.arrays.node_left,i=this.arrays.node_right,a=this.arrays.leaf_value,o=this.arrays.tree_offset,s=this.header.n_outputs,c=o.length-1,l=Array(s).fill(0);for(let u=0;u<c;u+=1){let c=o[u],d=t[c];for(;d>=0;)c=Math.fround(e[d])<=n[c]?r[c]:i[c],d=t[c];let f=(-1-d)*s;for(let e=0;e<s;e+=1)l[e]+=a[f+e]}for(let e=0;e<s;e+=1)l[e]=l[e]/c;return l}finish(e){let t=this.header.link,n=this.header.classes;if(t===`identity`)return{labels:e.map(e=>e[0]),probabilities:[]};let r=e.map(e=>{if(t===`sigmoid`){let t=1/(1+Math.exp(-e[0]));return[1-t,t]}if(t===`softmax`){let t=Math.max(...e),n=e.map(e=>Math.exp(e-t)),r=n.reduce((e,t)=>e+t,0);return n.map(e=>e/r)}let n=e.reduce((e,t)=>e+t,0);return n===0?[...e]:e.map(e=>e/n)}),i=this.header.class_type!==`str`;return{labels:r.map(e=>{let t=0;for(let n=1;n<e.length;n+=1)e[n]>e[t]&&(t=n);let r=n[t];return r===void 0?t:i?Number(r):r}),probabilities:r}}};async function r(t,n){if(typeof t==`string`){let r;try{r=await fetch(t,n)}catch(n){throw new e.ModelFetchError(`Could not download the model: ${t}`,{cause:n})}if(!r.ok)throw new e.ModelFetchError(`Could not download the model: ${r.status} ${r.statusText}`);return await r.arrayBuffer()}return t instanceof Uint8Array?t.buffer.slice(t.byteOffset,t.byteOffset+t.byteLength):t}function i(t,n){let r=n.node_feature,i=n.node_left,a=n.node_right,o=n.leaf_value,s=n.tree_offset;if(!r||!i||!a||!o||!s||s.length<2)throw new e.CompactFormatError(`This tree model is missing its node or offset sections.`);let c=r.length,l=t.n_outputs,u=t=>new e.CompactFormatError(`This tree model is malformed: ${t}. Re-export it.`);if(i.length!==c||a.length!==c||s[0]!==0)throw u(`the node sections disagree on the node count`);if(s[s.length-1]!==c)throw u(`the last tree does not end at the last node`);for(let e=0;e<s.length-1;e+=1){let n=s[e],c=s[e+1];if(c<=n)throw u(`tree ${e} has no nodes`);for(let e=n;e<c;e+=1){let n=r[e];if(n>=0){let r=i[e],o=a[e];if(n>=t.n_features)throw u(`node ${e} splits on feature ${n}`);if(r<=e||o<=e||r>=c||o>=c)throw u(`node ${e} points outside its tree or back up it`)}else if((-1-n)*l+l>o.length)throw u(`leaf ${e} reads past the leaf values`)}}}function a(t,n){if(!Array.isArray(t)||t.length===0)throw new e.FeatureShapeError(`predict() needs at least one row, shaped [[f1, f2, ...]].`);let r=t[0]?.length??0,i=t.findIndex(e=>e.length!==r);if(i!==-1)throw new e.FeatureShapeError(`All rows must have the same width; row ${i} has ${t[i]?.length} values, expected ${r}.`);if(r!==n)throw new e.FeatureShapeError(`The model expects ${n} features per row, got ${r}.`);return r}exports.CompactPredictor=n;
 //# sourceMappingURL=compact.cjs.map

package/dist/tabular/compact.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"compact.cjs","names":["CompactFormatError","ModelFetchError","FeatureShapeError","require_exceptions","MAGIC","CompactPredictor","header","arrays","info","buffer","toBuffer","source","requestInit","bytes","headerLength","cursor","section","width","validateRows","rows","started","scores","row","prepared","preprocess","scaled","index","coef","intercept","features","outputs","output","total","base","feature","threshold","left","right","leaf","offsets","trees","totals","tree","node","column","slot","link","classes","probabilities","positive","highest","exponentiated","value","sum","numeric","best","label","response","error","ragged","expected"],"sources":["../../src/tabular/compact.ts"],"sourcesContent":["/**\n * @tempest-limits file-lines — the compact model format: the header, the column\n * dictionary, the quantised weights and the decoder that reads them back. Writer and\n * reader in one file because a format whose two halves live apart drifts.\n */\n/**\n * Running a model with no inference runtime at all.\n *\n * ONNX in the browser costs a **25.6 MB WebAssembly runtime** (6.0 MB\n * gzipped) before the first prediction, while the model itself is a few\n * hundred kilobytes. For an app whose only model is tabular, that runtime\n * *is* the download — so this reader replaces it with 1.49 KB of\n * arithmetic (measured, brotli).\n *\n * A linear model is a dot product. A tree is a chain of comparisons. That\n * is the whole implementation; there is nothing here that a WebAssembly\n * kernel would do better at this size.\n *\n * The file it reads (`.tmc`, magic `TMC1`) is written by\n * `tempest-fastapi-sdk`'s `export_sklearn_to_compact`, which verifies the\n * bytes against scikit-learn's own predictions and refuses to write a file\n * that disagrees. **It is data, not code**: no `eval`, no generated\n * JavaScript, nothing a strict CSP forbids.\n *\n * The trade against {@link TabularPredictor}: ONNX covers every estimator,\n * this covers linear models and tree ensembles. Pick by what your app\n * already ships — a page that loads ONNX for vision pays nothing extra for\n * tabular ONNX.\n */\n\nimport { CompactFormatError, FeatureShapeError, ModelFetchError } from \"./exceptions\";\nimport type { FeatureRow, PredictedLabel, TabularPrediction } from \"./types\";\n\n/** Magic bytes opening every compact file. */\nconst MAGIC = \"TMC1\";\n\n/** Layout version this reader understands. */\nexport const SUPPORTED_COMPACT_SCHEMA = 1;\n\n/** What the file holds. */\nexport type CompactKind = \"linear\" | \"tree_ensemble\";\n\n/** How raw scores become probabilities. */\ntype CompactLink = \"softmax\" | \"sigmoid\" | \"normalize\" | \"identity\";\n\n/** One array stored after the header. */\ninterface CompactSection {\n    readonly name: string;\n    readonly dtype: \"float32\" | \"int32\";\n    readonly length: number;\n}\n\n/** The parsed header. */\ninterface CompactHeader {\n    readonly schema_version: number;\n    readonly kind: CompactKind;\n    readonly task: \"classification\" | \"regression\";\n    readonly link: CompactLink;\n    readonly classes: readonly string[];\n    /** How scikit-learn typed those labels: ``int``, ``float`` or ``str``. */\n    readonly class_type?: \"int\" | \"float\" | \"str\";\n    readonly n_features: number;\n    readonly n_outputs: number;\n    readonly n_trees?: number;\n    readonly estimator: string;\n    readonly feature_names: readonly string[];\n    readonly preprocess: { readonly offset: number[]; readonly scale: number[] } | null;\n    readonly sections: readonly CompactSection[];\n}\n\n/** What a loaded compact model is. */\nexport interface CompactPredictorInfo {\n    /** Which reader path the file uses. */\n    readonly kind: CompactKind;\n    /** Class labels in score-column order; empty for a regressor. */\n    readonly classes: readonly string[];\n    /** Values expected per row. */\n    readonly numFeatures: number;\n    /** Column order recorded at training time, when the export had one. */\n    readonly featureNames: readonly string[];\n    /** Trees in the ensemble; `0` for a linear model. */\n    readonly numTrees: number;\n    /** Whether the model produces class scores. */\n    readonly isClassifier: boolean;\n    /** Class name of the exported estimator. */\n    readonly estimator: string;\n}\n\n/**\n * A compact model, loaded and ready to answer.\n *\n * @example\n * ```ts\n * const predictor = await CompactPredictor.create(\"/models/risk.tmc\");\n * const { labels, probabilities } = await predictor.predict([[5.1, 3.5, 1.4, 0.2]]);\n * ```\n */\nexport class CompactPredictor {\n    private constructor(\n        private readonly header: CompactHeader,\n        private readonly arrays: Record<string, Float32Array | Int32Array>,\n        /** What is loaded. */\n        public readonly info: CompactPredictorInfo,\n    ) {}\n\n    /**\n     * Load a `.tmc` file.\n     *\n     * @param source A URL, or the bytes when the app already has them.\n     * @param requestInit `fetch` options, when `source` is a URL.\n     * @returns The loaded predictor.\n     * @throws {@link ModelFetchError} when a URL cannot be read.\n     * @throws {@link CompactFormatError} when the bytes are not a compact\n     *   model, or use a layout newer than this reader.\n     */\n    static async create(\n        source: string | ArrayBuffer | Uint8Array,\n        requestInit?: RequestInit,\n    ): Promise<CompactPredictor> {\n        const buffer = await toBuffer(source, requestInit);\n        const bytes = new Uint8Array(buffer);\n\n        if (String.fromCharCode(...bytes.subarray(0, 4)) !== MAGIC) {\n            throw new CompactFormatError(\n                \"This is not a compact model file: it does not start with \" +\n                    `\"${MAGIC}\". A .onnx file goes through TabularPredictor instead.`,\n            );\n        }\n\n        const view = new DataView(buffer);\n        const headerLength = view.getUint32(4, true);\n        const header = JSON.parse(\n            new TextDecoder().decode(bytes.subarray(8, 8 + headerLength)),\n        ) as CompactHeader;\n\n        if (header.schema_version > SUPPORTED_COMPACT_SCHEMA) {\n            throw new CompactFormatError(\n                `This file uses compact layout ${header.schema_version}, and this ` +\n                    `SDK understands ${SUPPORTED_COMPACT_SCHEMA}. Upgrade ` +\n                    \"tempest-react-sdk before serving it.\",\n            );\n        }\n\n        const arrays: Record<string, Float32Array | Int32Array> = {};\n        let cursor = 8 + headerLength;\n        for (const section of header.sections) {\n            arrays[section.name] =\n                section.dtype === \"float32\"\n                    ? new Float32Array(buffer, cursor, section.length)\n                    : new Int32Array(buffer, cursor, section.length);\n            cursor += section.length * 4;\n        }\n\n        return new CompactPredictor(header, arrays, {\n            kind: header.kind,\n            classes: header.classes,\n            numFeatures: header.n_features,\n            featureNames: header.feature_names,\n            numTrees: header.n_trees ?? 0,\n            isClassifier: header.task === \"classification\",\n            estimator: header.estimator,\n        });\n    }\n\n    /**\n     * Predict for a batch of rows.\n     *\n     * @param rows One array of feature values per row, in training column\n     *   order. A single row is still wrapped: `[[...]]`.\n     * @returns Labels, class scores when the model is a classifier, and the\n     *   call's duration.\n     * @throws {@link FeatureShapeError} when the batch is empty, ragged, or\n     *   the wrong width.\n     */\n    async predict(rows: readonly FeatureRow[]): Promise<TabularPrediction> {\n        const width = validateRows(rows, this.header.n_features);\n        const started = performance.now();\n\n        const scores: number[][] = [];\n        for (const row of rows) {\n            const prepared = this.preprocess(row, width);\n            scores.push(\n                this.header.kind === \"linear\"\n                    ? this.linearScores(prepared)\n                    : this.treeScores(prepared),\n            );\n        }\n\n        const { labels, probabilities } = this.finish(scores);\n        return {\n            labels,\n            probabilities,\n            numRows: rows.length,\n            ms: performance.now() - started,\n        };\n    }\n\n    /** Releasing nothing, so callers can swap predictors without branching. */\n    async dispose(): Promise<void> {\n        /* no runtime, nothing to release */\n    }\n\n    /**\n     * Apply the folded scaler, when the export had one.\n     *\n     * @param row The raw feature values.\n     * @param width How many there are.\n     * @returns The values the model was trained on.\n     */\n    private preprocess(row: FeatureRow, width: number): number[] {\n        const preprocess = this.header.preprocess;\n        if (preprocess === null) return row as number[];\n        const scaled = new Array<number>(width);\n        for (let index = 0; index < width; index += 1) {\n            scaled[index] =\n                ((row[index] as number) - (preprocess.offset[index] as number)) /\n                (preprocess.scale[index] as number);\n        }\n        return scaled;\n    }\n\n    /**\n     * Score one row against the coefficient matrix.\n     *\n     * @param row The prepared feature values.\n     * @returns One raw score per output.\n     */\n    private linearScores(row: readonly number[]): number[] {\n        const coef = this.arrays.coef as Float32Array;\n        const intercept = this.arrays.intercept as Float32Array;\n        const features = this.header.n_features;\n        const outputs = this.header.n_outputs;\n\n        const scores = new Array<number>(outputs);\n        for (let output = 0; output < outputs; output += 1) {\n            let total = intercept[output] as number;\n            const base = output * features;\n            for (let index = 0; index < features; index += 1) {\n                total += (coef[base + index] as number) * (row[index] as number);\n            }\n            scores[output] = total;\n        }\n        return scores;\n    }\n\n    /**\n     * Walk every tree and average what the leaves hold.\n     *\n     * A leaf is marked by a negative `feature` entry, which also carries\n     * its slot in the value array — so the walk needs no second lookup and\n     * the file stores values only for leaves.\n     *\n     * **The comparison runs in float32** (`Math.fround`), because that is\n     * what scikit-learn does: it casts its input to float32 before\n     * traversing, so a threshold like 5.099999904632568 — a float32 value\n     * widened for storage — and an input of 5.1 compare *equal* and go\n     * left. Comparing in float64 sends that row right instead, which on an\n     * iris forest changed one tree's vote and moved a probability by 0.05.\n     *\n     * @param row The prepared feature values.\n     * @returns One averaged score per output.\n     */\n    private treeScores(row: readonly number[]): number[] {\n        const feature = this.arrays.node_feature as Int32Array;\n        const threshold = this.arrays.node_threshold as Float32Array;\n        const left = this.arrays.node_left as Int32Array;\n        const right = this.arrays.node_right as Int32Array;\n        const leaf = this.arrays.leaf_value as Float32Array;\n        const offsets = this.arrays.tree_offset as Int32Array;\n        const outputs = this.header.n_outputs;\n        const trees = offsets.length - 1;\n\n        const totals = new Array<number>(outputs).fill(0);\n        for (let tree = 0; tree < trees; tree += 1) {\n            let node = offsets[tree] as number;\n            let column = feature[node] as number;\n            while (column >= 0) {\n                node =\n                    Math.fround(row[column] as number) <= (threshold[node] as number)\n                        ? (left[node] as number)\n                        : (right[node] as number);\n                column = feature[node] as number;\n            }\n            const slot = (-1 - column) * outputs;\n            for (let output = 0; output < outputs; output += 1) {\n                totals[output] += leaf[slot + output] as number;\n            }\n        }\n        for (let output = 0; output < outputs; output += 1) {\n            totals[output] = (totals[output] as number) / trees;\n        }\n        return totals;\n    }\n\n    /**\n     * Turn raw scores into labels and probabilities.\n     *\n     * @param scores One score array per row.\n     * @returns Labels and probabilities in the shape the ONNX route uses,\n     *   so an app can swap runtimes without touching its own code. That\n     *   includes the label's **type**: an integer class comes back as a\n     *   number here exactly as ONNX returns it, because two routes over one\n     *   model that disagree on `0` versus `\"0\"` break the day someone\n     *   switches.\n     */\n    private finish(scores: readonly number[][]): {\n        labels: PredictedLabel[];\n        probabilities: number[][];\n    } {\n        const link = this.header.link;\n        const classes = this.header.classes;\n\n        if (link === \"identity\") {\n            return { labels: scores.map((row) => row[0] as number), probabilities: [] };\n        }\n\n        const probabilities = scores.map((row) => {\n            if (link === \"sigmoid\") {\n                const positive = 1 / (1 + Math.exp(-(row[0] as number)));\n                return [1 - positive, positive];\n            }\n            if (link === \"softmax\") {\n                const highest = Math.max(...row);\n                const exponentiated = row.map((value) => Math.exp(value - highest));\n                const total = exponentiated.reduce((sum, value) => sum + value, 0);\n                return exponentiated.map((value) => value / total);\n            }\n            const total = row.reduce((sum, value) => sum + value, 0);\n            return total === 0 ? [...row] : row.map((value) => value / total);\n        });\n\n        const numeric = this.header.class_type !== \"str\";\n        const labels = probabilities.map((row) => {\n            let best = 0;\n            for (let index = 1; index < row.length; index += 1) {\n                if ((row[index] as number) > (row[best] as number)) best = index;\n            }\n            const label = classes[best];\n            if (label === undefined) return best;\n            return numeric ? Number(label) : label;\n        });\n\n        return { labels, probabilities };\n    }\n}\n\n/**\n * Read a source into an `ArrayBuffer`.\n *\n * @param source A URL or the bytes.\n * @param requestInit `fetch` options.\n * @returns The bytes.\n * @throws {@link ModelFetchError} when the URL cannot be read.\n */\nasync function toBuffer(\n    source: string | ArrayBuffer | Uint8Array,\n    requestInit?: RequestInit,\n): Promise<ArrayBuffer> {\n    if (typeof source === \"string\") {\n        let response: Response;\n        try {\n            response = await fetch(source, requestInit);\n        } catch (error) {\n            throw new ModelFetchError(`Could not download the model: ${source}`, {\n                cause: error,\n            });\n        }\n        if (!response.ok) {\n            throw new ModelFetchError(\n                `Could not download the model: ${response.status} ${response.statusText}`,\n            );\n        }\n        return await response.arrayBuffer();\n    }\n    if (source instanceof Uint8Array) {\n        return source.buffer.slice(\n            source.byteOffset,\n            source.byteOffset + source.byteLength,\n        ) as ArrayBuffer;\n    }\n    return source;\n}\n\n/**\n * Check a batch before predicting on it.\n *\n * @param rows The batch.\n * @param expected Features the model wants per row.\n * @returns The batch width.\n * @throws {@link FeatureShapeError} when the batch cannot be predicted on.\n */\nfunction validateRows(rows: readonly FeatureRow[], expected: number): number {\n    if (!Array.isArray(rows) || rows.length === 0) {\n        throw new FeatureShapeError(\"predict() needs at least one row, shaped [[f1, f2, ...]].\");\n    }\n    const width = rows[0]?.length ?? 0;\n    const ragged = rows.findIndex((row) => row.length !== width);\n    if (ragged !== -1) {\n        throw new FeatureShapeError(\n            `All rows must have the same width; row ${ragged} has ` +\n                `${rows[ragged]?.length} values, expected ${width}.`,\n        );\n    }\n    if (width !== expected) {\n        throw new FeatureShapeError(\n            `The model expects ${expected} features per row, got ${width}.`,\n        );\n    }\n    return width;\n}\n"],"mappings":"A,oCAkCA,IAAMI,EAAQ,OA+DDC,EAAb,MAAaA,CAAiB,CAEL,OACA,OAED,KAJpB,YACI,EACA,EAEA,EACF,CAJmB,KAAA,OAAAC,EACA,KAAA,OAAAC,EAED,KAAA,KAAAC,CACjB,CAYH,aAAa,OACT,EACA,EACyB,CACzB,IAAMC,EAAS,MAAMC,EAASC,EAAQC,CAAW,EAC3CC,EAAQ,IAAI,WAAWJ,CAAM,EAEnC,GAAI,OAAO,aAAa,GAAGI,EAAM,SAAS,EAAG,CAAC,CAAC,IAAMT,EACjD,MAAM,IAAIJ,EAAAA,mBACN,6DACQI,EAAM,uDAClB,EAIJ,IAAMU,EAAe,IADJ,SAASL,CACL,CAAA,CAAK,UAAU,EAAG,EAAI,EACrCH,EAAS,KAAK,MAChB,IAAI,YAAY,CAAC,CAAC,OAAOO,EAAM,SAAS,EAAG,EAAIC,CAAY,CAAC,CAChE,EAEA,GAAIR,EAAO,eAAA,EACP,MAAM,IAAIN,EAAAA,mBACN,iCAAiCM,EAAO,eAAe,2EAG3D,EAGJ,IAAMC,EAAoD,CAAC,EACvDQ,EAAS,EAAID,EACjB,IAAK,IAAME,KAAWV,EAAO,SACzB,EAAOU,EAAQ,MACXA,EAAQ,QAAU,UACZ,IAAI,aAAaP,EAAQM,EAAQC,EAAQ,MAAM,EAC/C,IAAI,WAAWP,EAAQM,EAAQC,EAAQ,MAAM,EACvD,GAAUA,EAAQ,OAAS,EAG/B,OAAO,IAAIX,EAAiBC,EAAQC,EAAQ,CACxC,KAAMD,EAAO,KACb,QAASA,EAAO,QAChB,YAAaA,EAAO,WACpB,aAAcA,EAAO,cACrB,SAAUA,EAAO,SAAW,EAC5B,aAAcA,EAAO,OAAS,iBAC9B,UAAWA,EAAO,SACtB,CAAC,CACL,CAYA,MAAM,QAAQ,EAAyD,CACnE,IAAMW,EAAQC,EAAaC,EAAM,KAAK,OAAO,UAAU,EACjDC,EAAU,YAAY,IAAI,EAE1BC,EAAqB,CAAC,EAC5B,IAAK,IAAMC,KAAOH,EAAM,CACpB,IAAMI,EAAW,KAAK,WAAWD,EAAKL,CAAK,EAC3CI,EAAO,KACH,KAAK,OAAO,OAAS,SACf,KAAK,aAAaE,CAAQ,EAC1B,KAAK,WAAWA,CAAQ,CAClC,CACJ,CAEA,GAAM,CAAE,SAAQ,iBAAkB,KAAK,OAAOF,CAAM,EACpD,MAAO,CACH,SACA,gBACA,QAASF,EAAK,OACd,GAAI,YAAY,IAAI,EAAIC,CAC5B,CACJ,CAGA,MAAM,SAAyB,CAE/B,CASA,WAAmB,EAAiB,EAAyB,CACzD,IAAMI,EAAa,KAAK,OAAO,WAC/B,GAAIA,IAAe,KAAM,OAAOF,EAChC,IAAMG,EAAa,MAAcR,CAAK,EACtC,IAAK,IAAIS,EAAQ,EAAGA,EAAQT,EAAO,GAAS,EACxC,EAAOS,IACDJ,EAAII,GAAqBF,EAAW,OAAOE,IAC5CF,EAAW,MAAME,GAE1B,OAAOD,CACX,CAQA,aAAqB,EAAkC,CACnD,IAAME,EAAO,KAAK,OAAO,KACnBC,EAAY,KAAK,OAAO,UACxBC,EAAW,KAAK,OAAO,WACvBC,EAAU,KAAK,OAAO,UAEtBT,EAAa,MAAcS,CAAO,EACxC,IAAK,IAAIC,EAAS,EAAGA,EAASD,EAAS,GAAU,EAAG,CAChD,IAAIE,EAAQJ,EAAUG,GAChBE,EAAOF,EAASF,EACtB,IAAK,IAAIH,EAAQ,EAAGA,EAAQG,EAAU,GAAS,EAC3C,GAAUF,EAAKM,EAAOP,GAAqBJ,EAAII,GAEnD,EAAOK,GAAUC,CACrB,CACA,OAAOX,CACX,CAmBA,WAAmB,EAAkC,CACjD,IAAMa,EAAU,KAAK,OAAO,aACtBC,EAAY,KAAK,OAAO,eACxBC,EAAO,KAAK,OAAO,UACnBC,EAAQ,KAAK,OAAO,WACpBC,EAAO,KAAK,OAAO,WACnBC,EAAU,KAAK,OAAO,YACtBT,EAAU,KAAK,OAAO,UACtBU,EAAQD,EAAQ,OAAS,EAEzBE,EAAa,MAAcX,CAAO,CAAC,CAAC,KAAK,CAAC,EAChD,IAAK,IAAIY,EAAO,EAAGA,EAAOF,EAAO,GAAQ,EAAG,CACxC,IAAIG,EAAOJ,EAAQG,GACfE,EAASV,EAAQS,GACrB,KAAOC,GAAU,GACb,EACI,KAAK,OAAOtB,EAAIsB,EAAiB,GAAMT,EAAUQ,GAC1CP,EAAKO,GACLN,EAAMM,GACjB,EAAST,EAAQS,GAErB,IAAME,GAAQ,GAAKD,GAAUd,EAC7B,IAAK,IAAIC,EAAS,EAAGA,EAASD,EAAS,GAAU,EAC7C,EAAOC,IAAWO,EAAKO,EAAOd,EAEtC,CACA,IAAK,IAAIA,EAAS,EAAGA,EAASD,EAAS,GAAU,EAC7C,EAAOC,GAAWU,EAAOV,GAAqBS,EAElD,OAAOC,CACX,CAaA,OAAe,EAGb,CACE,IAAMK,EAAO,KAAK,OAAO,KACnBC,EAAU,KAAK,OAAO,QAE5B,GAAID,IAAS,WACT,MAAO,CAAE,OAAQzB,EAAO,IAAK,GAAQC,EAAI,EAAY,EAAG,cAAe,CAAC,CAAE,EAG9E,IAAM0B,EAAgB3B,EAAO,IAAK,GAAQ,CACtC,GAAIyB,IAAS,UAAW,CACpB,IAAMG,EAAW,GAAK,EAAI,KAAK,IAAI,CAAE3B,EAAI,EAAa,GACtD,MAAO,CAAC,EAAI2B,EAAUA,CAAQ,CAClC,CACA,GAAIH,IAAS,UAAW,CACpB,IAAMI,EAAU,KAAK,IAAI,GAAG5B,CAAG,EACzB6B,EAAgB7B,EAAI,IAAK,GAAU,KAAK,IAAI8B,EAAQF,CAAO,CAAC,EAC5DlB,EAAQmB,EAAc,QAAQ,EAAK,IAAUE,EAAMD,EAAO,CAAC,EACjE,OAAOD,EAAc,IAAK,GAAUC,EAAQpB,CAAK,CACrD,CACA,IAAMA,EAAQV,EAAI,QAAQ,EAAK,IAAU+B,EAAMD,EAAO,CAAC,EACvD,OAAOpB,IAAU,EAAI,CAAC,GAAGV,CAAG,EAAIA,EAAI,IAAK,GAAU8B,EAAQpB,CAAK,CACpE,CAAC,EAEKsB,EAAU,KAAK,OAAO,aAAe,MAW3C,MAAO,CAAE,OAVMN,EAAc,IAAK,GAAQ,CACtC,IAAIO,EAAO,EACX,IAAK,IAAI7B,EAAQ,EAAGA,EAAQJ,EAAI,OAAQ,GAAS,EACxCA,EAAII,GAAqBJ,EAAIiC,KAAkB,EAAO7B,GAE/D,IAAM8B,EAAQT,EAAQQ,GAEtB,OADIC,IAAU,IAAA,GAAkBD,EACzBD,EAAU,OAAOE,CAAK,EAAIA,CACrC,CAES,EAAQ,eAAc,CACnC,CACJ,EAUA,eAAe9C,EACX,EACA,EACoB,CACpB,GAAI,OAAOC,GAAW,SAAU,CAC5B,IAAI8C,EACJ,GAAI,CACA,EAAW,MAAM,MAAM9C,EAAQC,CAAW,CAC9C,OAAS8C,EAAO,CACZ,MAAM,IAAIzD,EAAAA,gBAAgB,iCAAiCU,IAAU,CACjE,MAAO+C,CACX,CAAC,CACL,CACA,GAAI,CAACD,EAAS,GACV,MAAM,IAAIxD,EAAAA,gBACN,iCAAiCwD,EAAS,OAAO,GAAGA,EAAS,YACjE,EAEJ,OAAO,MAAMA,EAAS,YAAY,CACtC,CAOA,OANI9C,aAAkB,WACXA,EAAO,OAAO,MACjBA,EAAO,WACPA,EAAO,WAAaA,EAAO,UAC/B,EAEGA,CACX,CAUA,SAASO,EAAa,EAA6B,EAA0B,CACzE,GAAI,CAAC,MAAM,QAAQC,CAAI,GAAKA,EAAK,SAAW,EACxC,MAAM,IAAIjB,EAAAA,kBAAkB,2DAA2D,EAE3F,IAAMe,EAAQE,EAAK,EAAE,EAAE,QAAU,EAC3BwC,EAASxC,EAAK,UAAW,GAAQG,EAAI,SAAWL,CAAK,EAC3D,GAAI0C,IAAW,GACX,MAAM,IAAIzD,EAAAA,kBACN,0CAA0CyD,EAAO,OAC1CxC,EAAKwC,EAAO,EAAE,OAAO,oBAAoB1C,EAAM,EAC1D,EAEJ,GAAIA,IAAU2C,EACV,MAAM,IAAI1D,EAAAA,kBACN,qBAAqB0D,EAAS,yBAAyB3C,EAAM,EACjE,EAEJ,OAAOA,CACX,C"}
+{"version":3,"file":"compact.cjs","names":["CompactFormatError","ModelFetchError","FeatureShapeError","require_exceptions","MAGIC","CompactPredictor","header","arrays","info","buffer","toBuffer","source","requestInit","bytes","headerLength","cursor","section","assertTreesTerminate","width","validateRows","rows","started","scores","row","prepared","preprocess","scaled","index","coef","intercept","features","outputs","output","total","base","feature","threshold","left","right","leaf","offsets","trees","totals","tree","node","column","slot","link","classes","probabilities","positive","highest","exponentiated","value","sum","numeric","best","label","response","error","nodes","broken","detail","start","end","lower","upper","ragged","expected"],"sources":["../../src/tabular/compact.ts"],"sourcesContent":["/**\n * @tempest-limits file-lines — the compact model format: the header, the column\n * dictionary, the quantised weights and the decoder that reads them back. Writer and\n * reader in one file because a format whose two halves live apart drifts.\n */\n/**\n * Running a model with no inference runtime at all.\n *\n * ONNX in the browser costs a **25.6 MB WebAssembly runtime** (6.0 MB\n * gzipped) before the first prediction, while the model itself is a few\n * hundred kilobytes. For an app whose only model is tabular, that runtime\n * *is* the download — so this reader replaces it with 1.49 KB of\n * arithmetic (measured, brotli).\n *\n * A linear model is a dot product. A tree is a chain of comparisons. That\n * is the whole implementation; there is nothing here that a WebAssembly\n * kernel would do better at this size.\n *\n * The file it reads (`.tmc`, magic `TMC1`) is written by\n * `tempest-fastapi-sdk`'s `export_sklearn_to_compact`, which verifies the\n * bytes against scikit-learn's own predictions and refuses to write a file\n * that disagrees. **It is data, not code**: no `eval`, no generated\n * JavaScript, nothing a strict CSP forbids.\n *\n * The trade against {@link TabularPredictor}: ONNX covers every estimator,\n * this covers linear models and tree ensembles. Pick by what your app\n * already ships — a page that loads ONNX for vision pays nothing extra for\n * tabular ONNX.\n */\n\nimport { CompactFormatError, FeatureShapeError, ModelFetchError } from \"./exceptions\";\nimport type { FeatureRow, PredictedLabel, TabularPrediction } from \"./types\";\n\n/** Magic bytes opening every compact file. */\nconst MAGIC = \"TMC1\";\n\n/** Layout version this reader understands. */\nexport const SUPPORTED_COMPACT_SCHEMA = 1;\n\n/** What the file holds. */\nexport type CompactKind = \"linear\" | \"tree_ensemble\";\n\n/** How raw scores become probabilities. */\ntype CompactLink = \"softmax\" | \"sigmoid\" | \"normalize\" | \"identity\";\n\n/** One array stored after the header. */\ninterface CompactSection {\n    readonly name: string;\n    readonly dtype: \"float32\" | \"int32\";\n    readonly length: number;\n}\n\n/** The parsed header. */\ninterface CompactHeader {\n    readonly schema_version: number;\n    readonly kind: CompactKind;\n    readonly task: \"classification\" | \"regression\";\n    readonly link: CompactLink;\n    readonly classes: readonly string[];\n    /** How scikit-learn typed those labels: ``int``, ``float`` or ``str``. */\n    readonly class_type?: \"int\" | \"float\" | \"str\";\n    readonly n_features: number;\n    readonly n_outputs: number;\n    readonly n_trees?: number;\n    readonly estimator: string;\n    readonly feature_names: readonly string[];\n    readonly preprocess: { readonly offset: number[]; readonly scale: number[] } | null;\n    readonly sections: readonly CompactSection[];\n}\n\n/** What a loaded compact model is. */\nexport interface CompactPredictorInfo {\n    /** Which reader path the file uses. */\n    readonly kind: CompactKind;\n    /** Class labels in score-column order; empty for a regressor. */\n    readonly classes: readonly string[];\n    /** Values expected per row. */\n    readonly numFeatures: number;\n    /** Column order recorded at training time, when the export had one. */\n    readonly featureNames: readonly string[];\n    /** Trees in the ensemble; `0` for a linear model. */\n    readonly numTrees: number;\n    /** Whether the model produces class scores. */\n    readonly isClassifier: boolean;\n    /** Class name of the exported estimator. */\n    readonly estimator: string;\n}\n\n/**\n * A compact model, loaded and ready to answer.\n *\n * @example\n * ```ts\n * const predictor = await CompactPredictor.create(\"/models/risk.tmc\");\n * const { labels, probabilities } = await predictor.predict([[5.1, 3.5, 1.4, 0.2]]);\n * ```\n */\nexport class CompactPredictor {\n    private constructor(\n        private readonly header: CompactHeader,\n        private readonly arrays: Record<string, Float32Array | Int32Array>,\n        /** What is loaded. */\n        public readonly info: CompactPredictorInfo,\n    ) {}\n\n    /**\n     * Load a `.tmc` file.\n     *\n     * @param source A URL, or the bytes when the app already has them.\n     * @param requestInit `fetch` options, when `source` is a URL.\n     * @returns The loaded predictor.\n     * @throws {@link ModelFetchError} when a URL cannot be read.\n     * @throws {@link CompactFormatError} when the bytes are not a compact\n     *   model, or use a layout newer than this reader.\n     */\n    static async create(\n        source: string | ArrayBuffer | Uint8Array,\n        requestInit?: RequestInit,\n    ): Promise<CompactPredictor> {\n        const buffer = await toBuffer(source, requestInit);\n        const bytes = new Uint8Array(buffer);\n\n        if (String.fromCharCode(...bytes.subarray(0, 4)) !== MAGIC) {\n            throw new CompactFormatError(\n                \"This is not a compact model file: it does not start with \" +\n                    `\"${MAGIC}\". A .onnx file goes through TabularPredictor instead.`,\n            );\n        }\n\n        const view = new DataView(buffer);\n        const headerLength = view.getUint32(4, true);\n        const header = JSON.parse(\n            new TextDecoder().decode(bytes.subarray(8, 8 + headerLength)),\n        ) as CompactHeader;\n\n        if (header.schema_version > SUPPORTED_COMPACT_SCHEMA) {\n            throw new CompactFormatError(\n                `This file uses compact layout ${header.schema_version}, and this ` +\n                    `SDK understands ${SUPPORTED_COMPACT_SCHEMA}. Upgrade ` +\n                    \"tempest-react-sdk before serving it.\",\n            );\n        }\n\n        const arrays: Record<string, Float32Array | Int32Array> = {};\n        let cursor = 8 + headerLength;\n        for (const section of header.sections) {\n            arrays[section.name] =\n                section.dtype === \"float32\"\n                    ? new Float32Array(buffer, cursor, section.length)\n                    : new Int32Array(buffer, cursor, section.length);\n            cursor += section.length * 4;\n        }\n\n        if (header.kind === \"tree_ensemble\") assertTreesTerminate(header, arrays);\n\n        return new CompactPredictor(header, arrays, {\n            kind: header.kind,\n            classes: header.classes,\n            numFeatures: header.n_features,\n            featureNames: header.feature_names,\n            numTrees: header.n_trees ?? 0,\n            isClassifier: header.task === \"classification\",\n            estimator: header.estimator,\n        });\n    }\n\n    /**\n     * Predict for a batch of rows.\n     *\n     * @param rows One array of feature values per row, in training column\n     *   order. A single row is still wrapped: `[[...]]`.\n     * @returns Labels, class scores when the model is a classifier, and the\n     *   call's duration.\n     * @throws {@link FeatureShapeError} when the batch is empty, ragged, or\n     *   the wrong width.\n     */\n    async predict(rows: readonly FeatureRow[]): Promise<TabularPrediction> {\n        const width = validateRows(rows, this.header.n_features);\n        const started = performance.now();\n\n        const scores: number[][] = [];\n        for (const row of rows) {\n            const prepared = this.preprocess(row, width);\n            scores.push(\n                this.header.kind === \"linear\"\n                    ? this.linearScores(prepared)\n                    : this.treeScores(prepared),\n            );\n        }\n\n        const { labels, probabilities } = this.finish(scores);\n        return {\n            labels,\n            probabilities,\n            numRows: rows.length,\n            ms: performance.now() - started,\n        };\n    }\n\n    /** Releasing nothing, so callers can swap predictors without branching. */\n    async dispose(): Promise<void> {\n        /* no runtime, nothing to release */\n    }\n\n    /**\n     * Apply the folded scaler, when the export had one.\n     *\n     * @param row The raw feature values.\n     * @param width How many there are.\n     * @returns The values the model was trained on.\n     */\n    private preprocess(row: FeatureRow, width: number): number[] {\n        const preprocess = this.header.preprocess;\n        if (preprocess === null) return row as number[];\n        const scaled = new Array<number>(width);\n        for (let index = 0; index < width; index += 1) {\n            scaled[index] =\n                ((row[index] as number) - (preprocess.offset[index] as number)) /\n                (preprocess.scale[index] as number);\n        }\n        return scaled;\n    }\n\n    /**\n     * Score one row against the coefficient matrix.\n     *\n     * @param row The prepared feature values.\n     * @returns One raw score per output.\n     */\n    private linearScores(row: readonly number[]): number[] {\n        const coef = this.arrays.coef as Float32Array;\n        const intercept = this.arrays.intercept as Float32Array;\n        const features = this.header.n_features;\n        const outputs = this.header.n_outputs;\n\n        const scores = new Array<number>(outputs);\n        for (let output = 0; output < outputs; output += 1) {\n            let total = intercept[output] as number;\n            const base = output * features;\n            for (let index = 0; index < features; index += 1) {\n                total += (coef[base + index] as number) * (row[index] as number);\n            }\n            scores[output] = total;\n        }\n        return scores;\n    }\n\n    /**\n     * Walk every tree and average what the leaves hold.\n     *\n     * A leaf is marked by a negative `feature` entry, which also carries\n     * its slot in the value array — so the walk needs no second lookup and\n     * the file stores values only for leaves.\n     *\n     * **The comparison runs in float32** (`Math.fround`), because that is\n     * what scikit-learn does: it casts its input to float32 before\n     * traversing, so a threshold like 5.099999904632568 — a float32 value\n     * widened for storage — and an input of 5.1 compare *equal* and go\n     * left. Comparing in float64 sends that row right instead, which on an\n     * iris forest changed one tree's vote and moved a probability by 0.05.\n     *\n     * @param row The prepared feature values.\n     * @returns One averaged score per output.\n     */\n    private treeScores(row: readonly number[]): number[] {\n        const feature = this.arrays.node_feature as Int32Array;\n        const threshold = this.arrays.node_threshold as Float32Array;\n        const left = this.arrays.node_left as Int32Array;\n        const right = this.arrays.node_right as Int32Array;\n        const leaf = this.arrays.leaf_value as Float32Array;\n        const offsets = this.arrays.tree_offset as Int32Array;\n        const outputs = this.header.n_outputs;\n        const trees = offsets.length - 1;\n\n        const totals = new Array<number>(outputs).fill(0);\n        for (let tree = 0; tree < trees; tree += 1) {\n            let node = offsets[tree] as number;\n            let column = feature[node] as number;\n            while (column >= 0) {\n                node =\n                    Math.fround(row[column] as number) <= (threshold[node] as number)\n                        ? (left[node] as number)\n                        : (right[node] as number);\n                column = feature[node] as number;\n            }\n            const slot = (-1 - column) * outputs;\n            for (let output = 0; output < outputs; output += 1) {\n                totals[output] += leaf[slot + output] as number;\n            }\n        }\n        for (let output = 0; output < outputs; output += 1) {\n            totals[output] = (totals[output] as number) / trees;\n        }\n        return totals;\n    }\n\n    /**\n     * Turn raw scores into labels and probabilities.\n     *\n     * @param scores One score array per row.\n     * @returns Labels and probabilities in the shape the ONNX route uses,\n     *   so an app can swap runtimes without touching its own code. That\n     *   includes the label's **type**: an integer class comes back as a\n     *   number here exactly as ONNX returns it, because two routes over one\n     *   model that disagree on `0` versus `\"0\"` break the day someone\n     *   switches.\n     */\n    private finish(scores: readonly number[][]): {\n        labels: PredictedLabel[];\n        probabilities: number[][];\n    } {\n        const link = this.header.link;\n        const classes = this.header.classes;\n\n        if (link === \"identity\") {\n            return { labels: scores.map((row) => row[0] as number), probabilities: [] };\n        }\n\n        const probabilities = scores.map((row) => {\n            if (link === \"sigmoid\") {\n                const positive = 1 / (1 + Math.exp(-(row[0] as number)));\n                return [1 - positive, positive];\n            }\n            if (link === \"softmax\") {\n                const highest = Math.max(...row);\n                const exponentiated = row.map((value) => Math.exp(value - highest));\n                const total = exponentiated.reduce((sum, value) => sum + value, 0);\n                return exponentiated.map((value) => value / total);\n            }\n            const total = row.reduce((sum, value) => sum + value, 0);\n            return total === 0 ? [...row] : row.map((value) => value / total);\n        });\n\n        const numeric = this.header.class_type !== \"str\";\n        const labels = probabilities.map((row) => {\n            let best = 0;\n            for (let index = 1; index < row.length; index += 1) {\n                if ((row[index] as number) > (row[best] as number)) best = index;\n            }\n            const label = classes[best];\n            if (label === undefined) return best;\n            return numeric ? Number(label) : label;\n        });\n\n        return { labels, probabilities };\n    }\n}\n\n/**\n * Read a source into an `ArrayBuffer`.\n *\n * @param source A URL or the bytes.\n * @param requestInit `fetch` options.\n * @returns The bytes.\n * @throws {@link ModelFetchError} when the URL cannot be read.\n */\nasync function toBuffer(\n    source: string | ArrayBuffer | Uint8Array,\n    requestInit?: RequestInit,\n): Promise<ArrayBuffer> {\n    if (typeof source === \"string\") {\n        let response: Response;\n        try {\n            response = await fetch(source, requestInit);\n        } catch (error) {\n            throw new ModelFetchError(`Could not download the model: ${source}`, {\n                cause: error,\n            });\n        }\n        if (!response.ok) {\n            throw new ModelFetchError(\n                `Could not download the model: ${response.status} ${response.statusText}`,\n            );\n        }\n        return await response.arrayBuffer();\n    }\n    if (source instanceof Uint8Array) {\n        return source.buffer.slice(\n            source.byteOffset,\n            source.byteOffset + source.byteLength,\n        ) as ArrayBuffer;\n    }\n    return source;\n}\n\n/**\n * Prove once, at load, that every tree walk ends inside its arrays.\n *\n * The walk in `treeScores` follows `node_left`/`node_right` until it meets a\n * leaf, with no step limit — a limit there would be paid on every node of every\n * row. So a file whose child index points back up the tree spun the tab forever\n * on the first `predict`, and one pointing past the arrays read `undefined` and\n * answered `NaN`. scikit-learn numbers a node before its children, and\n * `export_sklearn_to_compact` keeps that order with each tree's nodes contiguous,\n * so a well-formed file satisfies `node < child < next tree's offset` everywhere\n * — measured on every tree fixture in `__fixtures__` (1 761 nodes, 0 violations).\n * Checking that invariant is linear in the node count and rules out a cycle by\n * construction, since every step strictly increases the index.\n *\n * @param header The parsed header.\n * @param arrays The sections, by name.\n * @throws {@link CompactFormatError} when a tree could loop, leave its arrays,\n *   read a feature the model does not have, or a leaf slot past the values.\n */\nfunction assertTreesTerminate(\n    header: CompactHeader,\n    arrays: Record<string, Float32Array | Int32Array>,\n): void {\n    const feature = arrays.node_feature;\n    const left = arrays.node_left;\n    const right = arrays.node_right;\n    const leaf = arrays.leaf_value;\n    const offsets = arrays.tree_offset;\n    if (!feature || !left || !right || !leaf || !offsets || offsets.length < 2) {\n        throw new CompactFormatError(\"This tree model is missing its node or offset sections.\");\n    }\n    const nodes = feature.length;\n    const outputs = header.n_outputs;\n    const broken = (detail: string): CompactFormatError =>\n        new CompactFormatError(`This tree model is malformed: ${detail}. Re-export it.`);\n\n    if (left.length !== nodes || right.length !== nodes || offsets[0] !== 0) {\n        throw broken(\"the node sections disagree on the node count\");\n    }\n    if (offsets[offsets.length - 1] !== nodes) {\n        throw broken(\"the last tree does not end at the last node\");\n    }\n    for (let tree = 0; tree < offsets.length - 1; tree += 1) {\n        const start = offsets[tree] as number;\n        const end = offsets[tree + 1] as number;\n        if (end <= start) throw broken(`tree ${tree} has no nodes`);\n        for (let node = start; node < end; node += 1) {\n            const column = feature[node] as number;\n            if (column >= 0) {\n                const lower = left[node] as number;\n                const upper = right[node] as number;\n                if (column >= header.n_features) {\n                    throw broken(`node ${node} splits on feature ${column}`);\n                }\n                if (lower <= node || upper <= node || lower >= end || upper >= end) {\n                    throw broken(`node ${node} points outside its tree or back up it`);\n                }\n            } else if ((-1 - column) * outputs + outputs > leaf.length) {\n                throw broken(`leaf ${node} reads past the leaf values`);\n            }\n        }\n    }\n}\n\n/**\n * Check a batch before predicting on it.\n *\n * @param rows The batch.\n * @param expected Features the model wants per row.\n * @returns The batch width.\n * @throws {@link FeatureShapeError} when the batch cannot be predicted on.\n */\nfunction validateRows(rows: readonly FeatureRow[], expected: number): number {\n    if (!Array.isArray(rows) || rows.length === 0) {\n        throw new FeatureShapeError(\"predict() needs at least one row, shaped [[f1, f2, ...]].\");\n    }\n    const width = rows[0]?.length ?? 0;\n    const ragged = rows.findIndex((row) => row.length !== width);\n    if (ragged !== -1) {\n        throw new FeatureShapeError(\n            `All rows must have the same width; row ${ragged} has ` +\n                `${rows[ragged]?.length} values, expected ${width}.`,\n        );\n    }\n    if (width !== expected) {\n        throw new FeatureShapeError(\n            `The model expects ${expected} features per row, got ${width}.`,\n        );\n    }\n    return width;\n}\n"],"mappings":"A,oCAkCA,IAAMI,EAAQ,OA+DDC,EAAb,MAAaA,CAAiB,CAEL,OACA,OAED,KAJpB,YACI,EACA,EAEA,EACF,CAJmB,KAAA,OAAAC,EACA,KAAA,OAAAC,EAED,KAAA,KAAAC,CACjB,CAYH,aAAa,OACT,EACA,EACyB,CACzB,IAAMC,EAAS,MAAMC,EAASC,EAAQC,CAAW,EAC3CC,EAAQ,IAAI,WAAWJ,CAAM,EAEnC,GAAI,OAAO,aAAa,GAAGI,EAAM,SAAS,EAAG,CAAC,CAAC,IAAMT,EACjD,MAAM,IAAIJ,EAAAA,mBACN,6DACQI,EAAM,uDAClB,EAIJ,IAAMU,EAAe,IADJ,SAASL,CACL,CAAA,CAAK,UAAU,EAAG,EAAI,EACrCH,EAAS,KAAK,MAChB,IAAI,YAAY,CAAC,CAAC,OAAOO,EAAM,SAAS,EAAG,EAAIC,CAAY,CAAC,CAChE,EAEA,GAAIR,EAAO,eAAA,EACP,MAAM,IAAIN,EAAAA,mBACN,iCAAiCM,EAAO,eAAe,2EAG3D,EAGJ,IAAMC,EAAoD,CAAC,EACvDQ,EAAS,EAAID,EACjB,IAAK,IAAME,KAAWV,EAAO,SACzB,EAAOU,EAAQ,MACXA,EAAQ,QAAU,UACZ,IAAI,aAAaP,EAAQM,EAAQC,EAAQ,MAAM,EAC/C,IAAI,WAAWP,EAAQM,EAAQC,EAAQ,MAAM,EACvD,GAAUA,EAAQ,OAAS,EAK/B,OAFIV,EAAO,OAAS,iBAAiBW,EAAqBX,EAAQC,CAAM,EAEjE,IAAIF,EAAiBC,EAAQC,EAAQ,CACxC,KAAMD,EAAO,KACb,QAASA,EAAO,QAChB,YAAaA,EAAO,WACpB,aAAcA,EAAO,cACrB,SAAUA,EAAO,SAAW,EAC5B,aAAcA,EAAO,OAAS,iBAC9B,UAAWA,EAAO,SACtB,CAAC,CACL,CAYA,MAAM,QAAQ,EAAyD,CACnE,IAAMY,EAAQC,EAAaC,EAAM,KAAK,OAAO,UAAU,EACjDC,EAAU,YAAY,IAAI,EAE1BC,EAAqB,CAAC,EAC5B,IAAK,IAAMC,KAAOH,EAAM,CACpB,IAAMI,EAAW,KAAK,WAAWD,EAAKL,CAAK,EAC3CI,EAAO,KACH,KAAK,OAAO,OAAS,SACf,KAAK,aAAaE,CAAQ,EAC1B,KAAK,WAAWA,CAAQ,CAClC,CACJ,CAEA,GAAM,CAAE,SAAQ,iBAAkB,KAAK,OAAOF,CAAM,EACpD,MAAO,CACH,SACA,gBACA,QAASF,EAAK,OACd,GAAI,YAAY,IAAI,EAAIC,CAC5B,CACJ,CAGA,MAAM,SAAyB,CAE/B,CASA,WAAmB,EAAiB,EAAyB,CACzD,IAAMI,EAAa,KAAK,OAAO,WAC/B,GAAIA,IAAe,KAAM,OAAOF,EAChC,IAAMG,EAAa,MAAcR,CAAK,EACtC,IAAK,IAAIS,EAAQ,EAAGA,EAAQT,EAAO,GAAS,EACxC,EAAOS,IACDJ,EAAII,GAAqBF,EAAW,OAAOE,IAC5CF,EAAW,MAAME,GAE1B,OAAOD,CACX,CAQA,aAAqB,EAAkC,CACnD,IAAME,EAAO,KAAK,OAAO,KACnBC,EAAY,KAAK,OAAO,UACxBC,EAAW,KAAK,OAAO,WACvBC,EAAU,KAAK,OAAO,UAEtBT,EAAa,MAAcS,CAAO,EACxC,IAAK,IAAIC,EAAS,EAAGA,EAASD,EAAS,GAAU,EAAG,CAChD,IAAIE,EAAQJ,EAAUG,GAChBE,EAAOF,EAASF,EACtB,IAAK,IAAIH,EAAQ,EAAGA,EAAQG,EAAU,GAAS,EAC3C,GAAUF,EAAKM,EAAOP,GAAqBJ,EAAII,GAEnD,EAAOK,GAAUC,CACrB,CACA,OAAOX,CACX,CAmBA,WAAmB,EAAkC,CACjD,IAAMa,EAAU,KAAK,OAAO,aACtBC,EAAY,KAAK,OAAO,eACxBC,EAAO,KAAK,OAAO,UACnBC,EAAQ,KAAK,OAAO,WACpBC,EAAO,KAAK,OAAO,WACnBC,EAAU,KAAK,OAAO,YACtBT,EAAU,KAAK,OAAO,UACtBU,EAAQD,EAAQ,OAAS,EAEzBE,EAAa,MAAcX,CAAO,CAAC,CAAC,KAAK,CAAC,EAChD,IAAK,IAAIY,EAAO,EAAGA,EAAOF,EAAO,GAAQ,EAAG,CACxC,IAAIG,EAAOJ,EAAQG,GACfE,EAASV,EAAQS,GACrB,KAAOC,GAAU,GACb,EACI,KAAK,OAAOtB,EAAIsB,EAAiB,GAAMT,EAAUQ,GAC1CP,EAAKO,GACLN,EAAMM,GACjB,EAAST,EAAQS,GAErB,IAAME,GAAQ,GAAKD,GAAUd,EAC7B,IAAK,IAAIC,EAAS,EAAGA,EAASD,EAAS,GAAU,EAC7C,EAAOC,IAAWO,EAAKO,EAAOd,EAEtC,CACA,IAAK,IAAIA,EAAS,EAAGA,EAASD,EAAS,GAAU,EAC7C,EAAOC,GAAWU,EAAOV,GAAqBS,EAElD,OAAOC,CACX,CAaA,OAAe,EAGb,CACE,IAAMK,EAAO,KAAK,OAAO,KACnBC,EAAU,KAAK,OAAO,QAE5B,GAAID,IAAS,WACT,MAAO,CAAE,OAAQzB,EAAO,IAAK,GAAQC,EAAI,EAAY,EAAG,cAAe,CAAC,CAAE,EAG9E,IAAM0B,EAAgB3B,EAAO,IAAK,GAAQ,CACtC,GAAIyB,IAAS,UAAW,CACpB,IAAMG,EAAW,GAAK,EAAI,KAAK,IAAI,CAAE3B,EAAI,EAAa,GACtD,MAAO,CAAC,EAAI2B,EAAUA,CAAQ,CAClC,CACA,GAAIH,IAAS,UAAW,CACpB,IAAMI,EAAU,KAAK,IAAI,GAAG5B,CAAG,EACzB6B,EAAgB7B,EAAI,IAAK,GAAU,KAAK,IAAI8B,EAAQF,CAAO,CAAC,EAC5DlB,EAAQmB,EAAc,QAAQ,EAAK,IAAUE,EAAMD,EAAO,CAAC,EACjE,OAAOD,EAAc,IAAK,GAAUC,EAAQpB,CAAK,CACrD,CACA,IAAMA,EAAQV,EAAI,QAAQ,EAAK,IAAU+B,EAAMD,EAAO,CAAC,EACvD,OAAOpB,IAAU,EAAI,CAAC,GAAGV,CAAG,EAAIA,EAAI,IAAK,GAAU8B,EAAQpB,CAAK,CACpE,CAAC,EAEKsB,EAAU,KAAK,OAAO,aAAe,MAW3C,MAAO,CAAE,OAVMN,EAAc,IAAK,GAAQ,CACtC,IAAIO,EAAO,EACX,IAAK,IAAI7B,EAAQ,EAAGA,EAAQJ,EAAI,OAAQ,GAAS,EACxCA,EAAII,GAAqBJ,EAAIiC,KAAkB,EAAO7B,GAE/D,IAAM8B,EAAQT,EAAQQ,GAEtB,OADIC,IAAU,IAAA,GAAkBD,EACzBD,EAAU,OAAOE,CAAK,EAAIA,CACrC,CAES,EAAQ,eAAc,CACnC,CACJ,EAUA,eAAe/C,EACX,EACA,EACoB,CACpB,GAAI,OAAOC,GAAW,SAAU,CAC5B,IAAI+C,EACJ,GAAI,CACA,EAAW,MAAM,MAAM/C,EAAQC,CAAW,CAC9C,OAAS+C,EAAO,CACZ,MAAM,IAAI1D,EAAAA,gBAAgB,iCAAiCU,IAAU,CACjE,MAAOgD,CACX,CAAC,CACL,CACA,GAAI,CAACD,EAAS,GACV,MAAM,IAAIzD,EAAAA,gBACN,iCAAiCyD,EAAS,OAAO,GAAGA,EAAS,YACjE,EAEJ,OAAO,MAAMA,EAAS,YAAY,CACtC,CAOA,OANI/C,aAAkB,WACXA,EAAO,OAAO,MACjBA,EAAO,WACPA,EAAO,WAAaA,EAAO,UAC/B,EAEGA,CACX,CAqBA,SAASM,EACL,EACA,EACI,CACJ,IAAMkB,EAAU5B,EAAO,aACjB8B,EAAO9B,EAAO,UACd+B,EAAQ/B,EAAO,WACfgC,EAAOhC,EAAO,WACdiC,EAAUjC,EAAO,YACvB,GAAI,CAAC4B,GAAW,CAACE,GAAQ,CAACC,GAAS,CAACC,GAAQ,CAACC,GAAWA,EAAQ,OAAS,EACrE,MAAM,IAAIxC,EAAAA,mBAAmB,yDAAyD,EAE1F,IAAM4D,EAAQzB,EAAQ,OAChBJ,EAAUzB,EAAO,UACjBuD,EAAU,GACZ,IAAI7D,EAAAA,mBAAmB,iCAAiC8D,EAAO,gBAAgB,EAEnF,GAAIzB,EAAK,SAAWuB,GAAStB,EAAM,SAAWsB,GAASpB,EAAQ,KAAO,EAClE,MAAMqB,EAAO,8CAA8C,EAE/D,GAAIrB,EAAQA,EAAQ,OAAS,KAAOoB,EAChC,MAAMC,EAAO,6CAA6C,EAE9D,IAAK,IAAIlB,EAAO,EAAGA,EAAOH,EAAQ,OAAS,EAAG,GAAQ,EAAG,CACrD,IAAMuB,EAAQvB,EAAQG,GAChBqB,EAAMxB,EAAQG,EAAO,GAC3B,GAAIqB,GAAOD,EAAO,MAAMF,EAAO,QAAQlB,EAAK,cAAc,EAC1D,IAAK,IAAIC,EAAOmB,EAAOnB,EAAOoB,EAAK,GAAQ,EAAG,CAC1C,IAAMnB,EAASV,EAAQS,GACvB,GAAIC,GAAU,EAAG,CACb,IAAMoB,EAAQ5B,EAAKO,GACbsB,EAAQ5B,EAAMM,GACpB,GAAIC,GAAUvC,EAAO,WACjB,MAAMuD,EAAO,QAAQjB,EAAK,qBAAqBC,GAAQ,EAE3D,GAAIoB,GAASrB,GAAQsB,GAAStB,GAAQqB,GAASD,GAAOE,GAASF,EAC3D,MAAMH,EAAO,QAAQjB,EAAK,uCAAuC,CAEzE,MAAO,IAAK,GAAKC,GAAUd,EAAUA,EAAUQ,EAAK,OAChD,MAAMsB,EAAO,QAAQjB,EAAK,4BAA4B,CAE9D,CACJ,CACJ,CAUA,SAASzB,EAAa,EAA6B,EAA0B,CACzE,GAAI,CAAC,MAAM,QAAQC,CAAI,GAAKA,EAAK,SAAW,EACxC,MAAM,IAAIlB,EAAAA,kBAAkB,2DAA2D,EAE3F,IAAMgB,EAAQE,EAAK,EAAE,EAAE,QAAU,EAC3B+C,EAAS/C,EAAK,UAAW,GAAQG,EAAI,SAAWL,CAAK,EAC3D,GAAIiD,IAAW,GACX,MAAM,IAAIjE,EAAAA,kBACN,0CAA0CiE,EAAO,OAC1C/C,EAAK+C,EAAO,EAAE,OAAO,oBAAoBjD,EAAM,EAC1D,EAEJ,GAAIA,IAAUkD,EACV,MAAM,IAAIlE,EAAAA,kBACN,qBAAqBkE,EAAS,yBAAyBlD,EAAM,EACjE,EAEJ,OAAOA,CACX,C"}

package/dist/tabular/compact.js

@@ -8,24 +8,24 @@ var r = "TMC1", i = class t {
 		this.header = e, this.arrays = t, this.info = n;
 	}
 	static async create(n, i) {
-		let o = await a(n, i), s = new Uint8Array(o);
-		if (String.fromCharCode(...s.subarray(0, 4)) !== r) throw new e(`This is not a compact model file: it does not start with "${r}". A .onnx file goes through TabularPredictor instead.`);
-		let c = new DataView(o).getUint32(4, !0), l = JSON.parse(new TextDecoder().decode(s.subarray(8, 8 + c)));
-		if (l.schema_version > 1) throw new e(`This file uses compact layout ${l.schema_version}, and this SDK understands 1. Upgrade tempest-react-sdk before serving it.`);
-		let u = {}, d = 8 + c;
-		for (let e of l.sections) u[e.name] = e.dtype === "float32" ? new Float32Array(o, d, e.length) : new Int32Array(o, d, e.length), d += e.length * 4;
-		return new t(l, u, {
-			kind: l.kind,
-			classes: l.classes,
-			numFeatures: l.n_features,
-			featureNames: l.feature_names,
-			numTrees: l.n_trees ?? 0,
-			isClassifier: l.task === "classification",
-			estimator: l.estimator
+		let s = await a(n, i), c = new Uint8Array(s);
+		if (String.fromCharCode(...c.subarray(0, 4)) !== r) throw new e(`This is not a compact model file: it does not start with "${r}". A .onnx file goes through TabularPredictor instead.`);
+		let l = new DataView(s).getUint32(4, !0), u = JSON.parse(new TextDecoder().decode(c.subarray(8, 8 + l)));
+		if (u.schema_version > 1) throw new e(`This file uses compact layout ${u.schema_version}, and this SDK understands 1. Upgrade tempest-react-sdk before serving it.`);
+		let d = {}, f = 8 + l;
+		for (let e of u.sections) d[e.name] = e.dtype === "float32" ? new Float32Array(s, f, e.length) : new Int32Array(s, f, e.length), f += e.length * 4;
+		return u.kind === "tree_ensemble" && o(u, d), new t(u, d, {
+			kind: u.kind,
+			classes: u.classes,
+			numFeatures: u.n_features,
+			featureNames: u.feature_names,
+			numTrees: u.n_trees ?? 0,
+			isClassifier: u.task === "classification",
+			estimator: u.estimator
 		});
 	}
 	async predict(e) {
-		let t = o(e, this.header.n_features), n = performance.now(), r = [];
+		let t = s(e, this.header.n_features), n = performance.now(), r = [];
 		for (let n of e) {
 			let e = this.preprocess(n, t);
 			r.push(this.header.kind === "linear" ? this.linearScores(e) : this.treeScores(e));
@@ -108,7 +108,26 @@ async function a(e, t) {
 	}
 	return e instanceof Uint8Array ? e.buffer.slice(e.byteOffset, e.byteOffset + e.byteLength) : e;
 }
-function o(e, n) {
+function o(t, n) {
+	let r = n.node_feature, i = n.node_left, a = n.node_right, o = n.leaf_value, s = n.tree_offset;
+	if (!r || !i || !a || !o || !s || s.length < 2) throw new e("This tree model is missing its node or offset sections.");
+	let c = r.length, l = t.n_outputs, u = (t) => new e(`This tree model is malformed: ${t}. Re-export it.`);
+	if (i.length !== c || a.length !== c || s[0] !== 0) throw u("the node sections disagree on the node count");
+	if (s[s.length - 1] !== c) throw u("the last tree does not end at the last node");
+	for (let e = 0; e < s.length - 1; e += 1) {
+		let n = s[e], c = s[e + 1];
+		if (c <= n) throw u(`tree ${e} has no nodes`);
+		for (let e = n; e < c; e += 1) {
+			let n = r[e];
+			if (n >= 0) {
+				let r = i[e], o = a[e];
+				if (n >= t.n_features) throw u(`node ${e} splits on feature ${n}`);
+				if (r <= e || o <= e || r >= c || o >= c) throw u(`node ${e} points outside its tree or back up it`);
+			} else if ((-1 - n) * l + l > o.length) throw u(`leaf ${e} reads past the leaf values`);
+		}
+	}
+}
+function s(e, n) {
 	if (!Array.isArray(e) || e.length === 0) throw new t("predict() needs at least one row, shaped [[f1, f2, ...]].");
 	let r = e[0]?.length ?? 0, i = e.findIndex((e) => e.length !== r);
 	if (i !== -1) throw new t(`All rows must have the same width; row ${i} has ${e[i]?.length} values, expected ${r}.`);

package/dist/tabular/compact.js.map

@@ -1 +1 @@
-{"version":3,"file":"compact.js","names":["MAGIC","CompactPredictor","header","arrays","info","buffer","toBuffer","source","requestInit","bytes","CompactFormatError","headerLength","cursor","section","width","validateRows","rows","started","scores","row","prepared","preprocess","scaled","index","coef","intercept","features","outputs","output","total","base","feature","threshold","left","right","leaf","offsets","trees","totals","tree","node","column","slot","link","classes","probabilities","positive","highest","exponentiated","value","sum","numeric","best","label","response","error","ModelFetchError","FeatureShapeError","ragged","expected"],"sources":["../../src/tabular/compact.ts"],"sourcesContent":["/**\n * @tempest-limits file-lines — the compact model format: the header, the column\n * dictionary, the quantised weights and the decoder that reads them back. Writer and\n * reader in one file because a format whose two halves live apart drifts.\n */\n/**\n * Running a model with no inference runtime at all.\n *\n * ONNX in the browser costs a **25.6 MB WebAssembly runtime** (6.0 MB\n * gzipped) before the first prediction, while the model itself is a few\n * hundred kilobytes. For an app whose only model is tabular, that runtime\n * *is* the download — so this reader replaces it with 1.49 KB of\n * arithmetic (measured, brotli).\n *\n * A linear model is a dot product. A tree is a chain of comparisons. That\n * is the whole implementation; there is nothing here that a WebAssembly\n * kernel would do better at this size.\n *\n * The file it reads (`.tmc`, magic `TMC1`) is written by\n * `tempest-fastapi-sdk`'s `export_sklearn_to_compact`, which verifies the\n * bytes against scikit-learn's own predictions and refuses to write a file\n * that disagrees. **It is data, not code**: no `eval`, no generated\n * JavaScript, nothing a strict CSP forbids.\n *\n * The trade against {@link TabularPredictor}: ONNX covers every estimator,\n * this covers linear models and tree ensembles. Pick by what your app\n * already ships — a page that loads ONNX for vision pays nothing extra for\n * tabular ONNX.\n */\n\nimport { CompactFormatError, FeatureShapeError, ModelFetchError } from \"./exceptions\";\nimport type { FeatureRow, PredictedLabel, TabularPrediction } from \"./types\";\n\n/** Magic bytes opening every compact file. */\nconst MAGIC = \"TMC1\";\n\n/** Layout version this reader understands. */\nexport const SUPPORTED_COMPACT_SCHEMA = 1;\n\n/** What the file holds. */\nexport type CompactKind = \"linear\" | \"tree_ensemble\";\n\n/** How raw scores become probabilities. */\ntype CompactLink = \"softmax\" | \"sigmoid\" | \"normalize\" | \"identity\";\n\n/** One array stored after the header. */\ninterface CompactSection {\n    readonly name: string;\n    readonly dtype: \"float32\" | \"int32\";\n    readonly length: number;\n}\n\n/** The parsed header. */\ninterface CompactHeader {\n    readonly schema_version: number;\n    readonly kind: CompactKind;\n    readonly task: \"classification\" | \"regression\";\n    readonly link: CompactLink;\n    readonly classes: readonly string[];\n    /** How scikit-learn typed those labels: ``int``, ``float`` or ``str``. */\n    readonly class_type?: \"int\" | \"float\" | \"str\";\n    readonly n_features: number;\n    readonly n_outputs: number;\n    readonly n_trees?: number;\n    readonly estimator: string;\n    readonly feature_names: readonly string[];\n    readonly preprocess: { readonly offset: number[]; readonly scale: number[] } | null;\n    readonly sections: readonly CompactSection[];\n}\n\n/** What a loaded compact model is. */\nexport interface CompactPredictorInfo {\n    /** Which reader path the file uses. */\n    readonly kind: CompactKind;\n    /** Class labels in score-column order; empty for a regressor. */\n    readonly classes: readonly string[];\n    /** Values expected per row. */\n    readonly numFeatures: number;\n    /** Column order recorded at training time, when the export had one. */\n    readonly featureNames: readonly string[];\n    /** Trees in the ensemble; `0` for a linear model. */\n    readonly numTrees: number;\n    /** Whether the model produces class scores. */\n    readonly isClassifier: boolean;\n    /** Class name of the exported estimator. */\n    readonly estimator: string;\n}\n\n/**\n * A compact model, loaded and ready to answer.\n *\n * @example\n * ```ts\n * const predictor = await CompactPredictor.create(\"/models/risk.tmc\");\n * const { labels, probabilities } = await predictor.predict([[5.1, 3.5, 1.4, 0.2]]);\n * ```\n */\nexport class CompactPredictor {\n    private constructor(\n        private readonly header: CompactHeader,\n        private readonly arrays: Record<string, Float32Array | Int32Array>,\n        /** What is loaded. */\n        public readonly info: CompactPredictorInfo,\n    ) {}\n\n    /**\n     * Load a `.tmc` file.\n     *\n     * @param source A URL, or the bytes when the app already has them.\n     * @param requestInit `fetch` options, when `source` is a URL.\n     * @returns The loaded predictor.\n     * @throws {@link ModelFetchError} when a URL cannot be read.\n     * @throws {@link CompactFormatError} when the bytes are not a compact\n     *   model, or use a layout newer than this reader.\n     */\n    static async create(\n        source: string | ArrayBuffer | Uint8Array,\n        requestInit?: RequestInit,\n    ): Promise<CompactPredictor> {\n        const buffer = await toBuffer(source, requestInit);\n        const bytes = new Uint8Array(buffer);\n\n        if (String.fromCharCode(...bytes.subarray(0, 4)) !== MAGIC) {\n            throw new CompactFormatError(\n                \"This is not a compact model file: it does not start with \" +\n                    `\"${MAGIC}\". A .onnx file goes through TabularPredictor instead.`,\n            );\n        }\n\n        const view = new DataView(buffer);\n        const headerLength = view.getUint32(4, true);\n        const header = JSON.parse(\n            new TextDecoder().decode(bytes.subarray(8, 8 + headerLength)),\n        ) as CompactHeader;\n\n        if (header.schema_version > SUPPORTED_COMPACT_SCHEMA) {\n            throw new CompactFormatError(\n                `This file uses compact layout ${header.schema_version}, and this ` +\n                    `SDK understands ${SUPPORTED_COMPACT_SCHEMA}. Upgrade ` +\n                    \"tempest-react-sdk before serving it.\",\n            );\n        }\n\n        const arrays: Record<string, Float32Array | Int32Array> = {};\n        let cursor = 8 + headerLength;\n        for (const section of header.sections) {\n            arrays[section.name] =\n                section.dtype === \"float32\"\n                    ? new Float32Array(buffer, cursor, section.length)\n                    : new Int32Array(buffer, cursor, section.length);\n            cursor += section.length * 4;\n        }\n\n        return new CompactPredictor(header, arrays, {\n            kind: header.kind,\n            classes: header.classes,\n            numFeatures: header.n_features,\n            featureNames: header.feature_names,\n            numTrees: header.n_trees ?? 0,\n            isClassifier: header.task === \"classification\",\n            estimator: header.estimator,\n        });\n    }\n\n    /**\n     * Predict for a batch of rows.\n     *\n     * @param rows One array of feature values per row, in training column\n     *   order. A single row is still wrapped: `[[...]]`.\n     * @returns Labels, class scores when the model is a classifier, and the\n     *   call's duration.\n     * @throws {@link FeatureShapeError} when the batch is empty, ragged, or\n     *   the wrong width.\n     */\n    async predict(rows: readonly FeatureRow[]): Promise<TabularPrediction> {\n        const width = validateRows(rows, this.header.n_features);\n        const started = performance.now();\n\n        const scores: number[][] = [];\n        for (const row of rows) {\n            const prepared = this.preprocess(row, width);\n            scores.push(\n                this.header.kind === \"linear\"\n                    ? this.linearScores(prepared)\n                    : this.treeScores(prepared),\n            );\n        }\n\n        const { labels, probabilities } = this.finish(scores);\n        return {\n            labels,\n            probabilities,\n            numRows: rows.length,\n            ms: performance.now() - started,\n        };\n    }\n\n    /** Releasing nothing, so callers can swap predictors without branching. */\n    async dispose(): Promise<void> {\n        /* no runtime, nothing to release */\n    }\n\n    /**\n     * Apply the folded scaler, when the export had one.\n     *\n     * @param row The raw feature values.\n     * @param width How many there are.\n     * @returns The values the model was trained on.\n     */\n    private preprocess(row: FeatureRow, width: number): number[] {\n        const preprocess = this.header.preprocess;\n        if (preprocess === null) return row as number[];\n        const scaled = new Array<number>(width);\n        for (let index = 0; index < width; index += 1) {\n            scaled[index] =\n                ((row[index] as number) - (preprocess.offset[index] as number)) /\n                (preprocess.scale[index] as number);\n        }\n        return scaled;\n    }\n\n    /**\n     * Score one row against the coefficient matrix.\n     *\n     * @param row The prepared feature values.\n     * @returns One raw score per output.\n     */\n    private linearScores(row: readonly number[]): number[] {\n        const coef = this.arrays.coef as Float32Array;\n        const intercept = this.arrays.intercept as Float32Array;\n        const features = this.header.n_features;\n        const outputs = this.header.n_outputs;\n\n        const scores = new Array<number>(outputs);\n        for (let output = 0; output < outputs; output += 1) {\n            let total = intercept[output] as number;\n            const base = output * features;\n            for (let index = 0; index < features; index += 1) {\n                total += (coef[base + index] as number) * (row[index] as number);\n            }\n            scores[output] = total;\n        }\n        return scores;\n    }\n\n    /**\n     * Walk every tree and average what the leaves hold.\n     *\n     * A leaf is marked by a negative `feature` entry, which also carries\n     * its slot in the value array — so the walk needs no second lookup and\n     * the file stores values only for leaves.\n     *\n     * **The comparison runs in float32** (`Math.fround`), because that is\n     * what scikit-learn does: it casts its input to float32 before\n     * traversing, so a threshold like 5.099999904632568 — a float32 value\n     * widened for storage — and an input of 5.1 compare *equal* and go\n     * left. Comparing in float64 sends that row right instead, which on an\n     * iris forest changed one tree's vote and moved a probability by 0.05.\n     *\n     * @param row The prepared feature values.\n     * @returns One averaged score per output.\n     */\n    private treeScores(row: readonly number[]): number[] {\n        const feature = this.arrays.node_feature as Int32Array;\n        const threshold = this.arrays.node_threshold as Float32Array;\n        const left = this.arrays.node_left as Int32Array;\n        const right = this.arrays.node_right as Int32Array;\n        const leaf = this.arrays.leaf_value as Float32Array;\n        const offsets = this.arrays.tree_offset as Int32Array;\n        const outputs = this.header.n_outputs;\n        const trees = offsets.length - 1;\n\n        const totals = new Array<number>(outputs).fill(0);\n        for (let tree = 0; tree < trees; tree += 1) {\n            let node = offsets[tree] as number;\n            let column = feature[node] as number;\n            while (column >= 0) {\n                node =\n                    Math.fround(row[column] as number) <= (threshold[node] as number)\n                        ? (left[node] as number)\n                        : (right[node] as number);\n                column = feature[node] as number;\n            }\n            const slot = (-1 - column) * outputs;\n            for (let output = 0; output < outputs; output += 1) {\n                totals[output] += leaf[slot + output] as number;\n            }\n        }\n        for (let output = 0; output < outputs; output += 1) {\n            totals[output] = (totals[output] as number) / trees;\n        }\n        return totals;\n    }\n\n    /**\n     * Turn raw scores into labels and probabilities.\n     *\n     * @param scores One score array per row.\n     * @returns Labels and probabilities in the shape the ONNX route uses,\n     *   so an app can swap runtimes without touching its own code. That\n     *   includes the label's **type**: an integer class comes back as a\n     *   number here exactly as ONNX returns it, because two routes over one\n     *   model that disagree on `0` versus `\"0\"` break the day someone\n     *   switches.\n     */\n    private finish(scores: readonly number[][]): {\n        labels: PredictedLabel[];\n        probabilities: number[][];\n    } {\n        const link = this.header.link;\n        const classes = this.header.classes;\n\n        if (link === \"identity\") {\n            return { labels: scores.map((row) => row[0] as number), probabilities: [] };\n        }\n\n        const probabilities = scores.map((row) => {\n            if (link === \"sigmoid\") {\n                const positive = 1 / (1 + Math.exp(-(row[0] as number)));\n                return [1 - positive, positive];\n            }\n            if (link === \"softmax\") {\n                const highest = Math.max(...row);\n                const exponentiated = row.map((value) => Math.exp(value - highest));\n                const total = exponentiated.reduce((sum, value) => sum + value, 0);\n                return exponentiated.map((value) => value / total);\n            }\n            const total = row.reduce((sum, value) => sum + value, 0);\n            return total === 0 ? [...row] : row.map((value) => value / total);\n        });\n\n        const numeric = this.header.class_type !== \"str\";\n        const labels = probabilities.map((row) => {\n            let best = 0;\n            for (let index = 1; index < row.length; index += 1) {\n                if ((row[index] as number) > (row[best] as number)) best = index;\n            }\n            const label = classes[best];\n            if (label === undefined) return best;\n            return numeric ? Number(label) : label;\n        });\n\n        return { labels, probabilities };\n    }\n}\n\n/**\n * Read a source into an `ArrayBuffer`.\n *\n * @param source A URL or the bytes.\n * @param requestInit `fetch` options.\n * @returns The bytes.\n * @throws {@link ModelFetchError} when the URL cannot be read.\n */\nasync function toBuffer(\n    source: string | ArrayBuffer | Uint8Array,\n    requestInit?: RequestInit,\n): Promise<ArrayBuffer> {\n    if (typeof source === \"string\") {\n        let response: Response;\n        try {\n            response = await fetch(source, requestInit);\n        } catch (error) {\n            throw new ModelFetchError(`Could not download the model: ${source}`, {\n                cause: error,\n            });\n        }\n        if (!response.ok) {\n            throw new ModelFetchError(\n                `Could not download the model: ${response.status} ${response.statusText}`,\n            );\n        }\n        return await response.arrayBuffer();\n    }\n    if (source instanceof Uint8Array) {\n        return source.buffer.slice(\n            source.byteOffset,\n            source.byteOffset + source.byteLength,\n        ) as ArrayBuffer;\n    }\n    return source;\n}\n\n/**\n * Check a batch before predicting on it.\n *\n * @param rows The batch.\n * @param expected Features the model wants per row.\n * @returns The batch width.\n * @throws {@link FeatureShapeError} when the batch cannot be predicted on.\n */\nfunction validateRows(rows: readonly FeatureRow[], expected: number): number {\n    if (!Array.isArray(rows) || rows.length === 0) {\n        throw new FeatureShapeError(\"predict() needs at least one row, shaped [[f1, f2, ...]].\");\n    }\n    const width = rows[0]?.length ?? 0;\n    const ragged = rows.findIndex((row) => row.length !== width);\n    if (ragged !== -1) {\n        throw new FeatureShapeError(\n            `All rows must have the same width; row ${ragged} has ` +\n                `${rows[ragged]?.length} values, expected ${width}.`,\n        );\n    }\n    if (width !== expected) {\n        throw new FeatureShapeError(\n            `The model expects ${expected} features per row, got ${width}.`,\n        );\n    }\n    return width;\n}\n"],"mappings":";;AAkCA,IAAMA,IAAQ,QA+DDC,IAAb,MAAaA,EAAiB;CAEL;CACA;CAED;CAJpB,YACI,GACA,GAEA,GACF;EADkB,AAHC,KAAA,SAAAC,GACA,KAAA,SAAAC,GAED,KAAA,OAAAC;CACjB;CAYH,aAAa,OACT,GACA,GACyB;EACzB,IAAMC,IAAS,MAAMC,EAASC,GAAQC,CAAW,GAC3CC,IAAQ,IAAI,WAAWJ,CAAM;EAEnC,IAAI,OAAO,aAAa,GAAGI,EAAM,SAAS,GAAG,CAAC,CAAC,MAAMT,GACjD,MAAM,IAAIU,EACN,6DACQV,EAAM,uDAClB;EAIJ,IAAMW,IAAe,IADJ,SAASN,CACL,CAAA,CAAK,UAAU,GAAG,EAAI,GACrCH,IAAS,KAAK,MAChB,IAAI,YAAY,CAAC,CAAC,OAAOO,EAAM,SAAS,GAAG,IAAIE,CAAY,CAAC,CAChE;EAEA,IAAIT,EAAO,iBAAA,GACP,MAAM,IAAIQ,EACN,iCAAiCR,EAAO,eAAe,2EAG3D;EAGJ,IAAMC,IAAoD,CAAC,GACvDS,IAAS,IAAID;EACjB,KAAK,IAAME,KAAWX,EAAO,UAKzB,AAJA,EAAOW,EAAQ,QACXA,EAAQ,UAAU,YACZ,IAAI,aAAaR,GAAQO,GAAQC,EAAQ,MAAM,IAC/C,IAAI,WAAWR,GAAQO,GAAQC,EAAQ,MAAM,GACvD,KAAUA,EAAQ,SAAS;EAG/B,OAAO,IAAIZ,EAAiBC,GAAQC,GAAQ;GACxC,MAAMD,EAAO;GACb,SAASA,EAAO;GAChB,aAAaA,EAAO;GACpB,cAAcA,EAAO;GACrB,UAAUA,EAAO,WAAW;GAC5B,cAAcA,EAAO,SAAS;GAC9B,WAAWA,EAAO;EACtB,CAAC;CACL;CAYA,MAAM,QAAQ,GAAyD;EACnE,IAAMY,IAAQC,EAAaC,GAAM,KAAK,OAAO,UAAU,GACjDC,IAAU,YAAY,IAAI,GAE1BC,IAAqB,CAAC;EAC5B,KAAK,IAAMC,KAAOH,GAAM;GACpB,IAAMI,IAAW,KAAK,WAAWD,GAAKL,CAAK;GAC3C,EAAO,KACH,KAAK,OAAO,SAAS,WACf,KAAK,aAAaM,CAAQ,IAC1B,KAAK,WAAWA,CAAQ,CAClC;EACJ;EAEA,IAAM,EAAE,WAAQ,qBAAkB,KAAK,OAAOF,CAAM;EACpD,OAAO;GACH;GACA;GACA,SAASF,EAAK;GACd,IAAI,YAAY,IAAI,IAAIC;EAC5B;CACJ;CAGA,MAAM,UAAyB,CAE/B;CASA,WAAmB,GAAiB,GAAyB;EACzD,IAAMI,IAAa,KAAK,OAAO;EAC/B,IAAIA,MAAe,MAAM,OAAOF;EAChC,IAAMG,IAAa,MAAcR,CAAK;EACtC,KAAK,IAAIS,IAAQ,GAAGA,IAAQT,GAAO,KAAS,GACxC,EAAOS,MACDJ,EAAII,KAAqBF,EAAW,OAAOE,MAC5CF,EAAW,MAAME;EAE1B,OAAOD;CACX;CAQA,aAAqB,GAAkC;EACnD,IAAME,IAAO,KAAK,OAAO,MACnBC,IAAY,KAAK,OAAO,WACxBC,IAAW,KAAK,OAAO,YACvBC,IAAU,KAAK,OAAO,WAEtBT,IAAa,MAAcS,CAAO;EACxC,KAAK,IAAIC,IAAS,GAAGA,IAASD,GAAS,KAAU,GAAG;GAChD,IAAIE,IAAQJ,EAAUG,IAChBE,IAAOF,IAASF;GACtB,KAAK,IAAIH,IAAQ,GAAGA,IAAQG,GAAU,KAAS,GAC3C,KAAUF,EAAKM,IAAOP,KAAqBJ,EAAII;GAEnD,EAAOK,KAAUC;EACrB;EACA,OAAOX;CACX;CAmBA,WAAmB,GAAkC;EACjD,IAAMa,IAAU,KAAK,OAAO,cACtBC,IAAY,KAAK,OAAO,gBACxBC,IAAO,KAAK,OAAO,WACnBC,IAAQ,KAAK,OAAO,YACpBC,IAAO,KAAK,OAAO,YACnBC,IAAU,KAAK,OAAO,aACtBT,IAAU,KAAK,OAAO,WACtBU,IAAQD,EAAQ,SAAS,GAEzBE,IAAa,MAAcX,CAAO,CAAC,CAAC,KAAK,CAAC;EAChD,KAAK,IAAIY,IAAO,GAAGA,IAAOF,GAAO,KAAQ,GAAG;GACxC,IAAIG,IAAOJ,EAAQG,IACfE,IAASV,EAAQS;GACrB,OAAOC,KAAU,IAKb,AAJA,IACI,KAAK,OAAOtB,EAAIsB,EAAiB,KAAMT,EAAUQ,KAC1CP,EAAKO,KACLN,EAAMM,IACjB,IAAST,EAAQS;GAErB,IAAME,KAAQ,KAAKD,KAAUd;GAC7B,KAAK,IAAIC,IAAS,GAAGA,IAASD,GAAS,KAAU,GAC7C,EAAOC,MAAWO,EAAKO,IAAOd;EAEtC;EACA,KAAK,IAAIA,IAAS,GAAGA,IAASD,GAAS,KAAU,GAC7C,EAAOC,KAAWU,EAAOV,KAAqBS;EAElD,OAAOC;CACX;CAaA,OAAe,GAGb;EACE,IAAMK,IAAO,KAAK,OAAO,MACnBC,IAAU,KAAK,OAAO;EAE5B,IAAID,MAAS,YACT,OAAO;GAAE,QAAQzB,EAAO,KAAK,MAAQC,EAAI,EAAY;GAAG,eAAe,CAAC;EAAE;EAG9E,IAAM0B,IAAgB3B,EAAO,KAAK,MAAQ;GACtC,IAAIyB,MAAS,WAAW;IACpB,IAAMG,IAAW,KAAK,IAAI,KAAK,IAAI,CAAE3B,EAAI,EAAa;IACtD,OAAO,CAAC,IAAI2B,GAAUA,CAAQ;GAClC;GACA,IAAIH,MAAS,WAAW;IACpB,IAAMI,IAAU,KAAK,IAAI,GAAG5B,CAAG,GACzB6B,IAAgB7B,EAAI,KAAK,MAAU,KAAK,IAAI8B,IAAQF,CAAO,CAAC,GAC5DlB,IAAQmB,EAAc,QAAQ,GAAK,MAAUE,IAAMD,GAAO,CAAC;IACjE,OAAOD,EAAc,KAAK,MAAUC,IAAQpB,CAAK;GACrD;GACA,IAAMA,IAAQV,EAAI,QAAQ,GAAK,MAAU+B,IAAMD,GAAO,CAAC;GACvD,OAAOpB,MAAU,IAAI,CAAC,GAAGV,CAAG,IAAIA,EAAI,KAAK,MAAU8B,IAAQpB,CAAK;EACpE,CAAC,GAEKsB,IAAU,KAAK,OAAO,eAAe;EAW3C,OAAO;GAAE,QAVMN,EAAc,KAAK,MAAQ;IACtC,IAAIO,IAAO;IACX,KAAK,IAAI7B,IAAQ,GAAGA,IAAQJ,EAAI,QAAQ,KAAS,GAC7C,AAAKA,EAAII,KAAqBJ,EAAIiC,OAAkB,IAAO7B;IAE/D,IAAM8B,IAAQT,EAAQQ;IAEtB,OADIC,MAAU,KAAA,IAAkBD,IACzBD,IAAU,OAAOE,CAAK,IAAIA;GACrC,CAES;GAAQ;EAAc;CACnC;AACJ;AAUA,eAAe/C,EACX,GACA,GACoB;CACpB,IAAI,OAAOC,KAAW,UAAU;EAC5B,IAAI+C;EACJ,IAAI;GACA,IAAW,MAAM,MAAM/C,GAAQC,CAAW;EAC9C,SAAS+C,GAAO;GACZ,MAAM,IAAIC,EAAgB,iCAAiCjD,KAAU,EACjE,OAAOgD,EACX,CAAC;EACL;EACA,IAAI,CAACD,EAAS,IACV,MAAM,IAAIE,EACN,iCAAiCF,EAAS,OAAO,GAAGA,EAAS,YACjE;EAEJ,OAAO,MAAMA,EAAS,YAAY;CACtC;CAOA,OANI/C,aAAkB,aACXA,EAAO,OAAO,MACjBA,EAAO,YACPA,EAAO,aAAaA,EAAO,UAC/B,IAEGA;AACX;AAUA,SAASQ,EAAa,GAA6B,GAA0B;CACzE,IAAI,CAAC,MAAM,QAAQC,CAAI,KAAKA,EAAK,WAAW,GACxC,MAAM,IAAIyC,EAAkB,2DAA2D;CAE3F,IAAM3C,IAAQE,EAAK,EAAE,EAAE,UAAU,GAC3B0C,IAAS1C,EAAK,WAAW,MAAQG,EAAI,WAAWL,CAAK;CAC3D,IAAI4C,MAAW,IACX,MAAM,IAAID,EACN,0CAA0CC,EAAO,OAC1C1C,EAAK0C,EAAO,EAAE,OAAO,oBAAoB5C,EAAM,EAC1D;CAEJ,IAAIA,MAAU6C,GACV,MAAM,IAAIF,EACN,qBAAqBE,EAAS,yBAAyB7C,EAAM,EACjE;CAEJ,OAAOA;AACX"}
+{"version":3,"file":"compact.js","names":["MAGIC","CompactPredictor","header","arrays","info","buffer","toBuffer","source","requestInit","bytes","CompactFormatError","headerLength","cursor","section","assertTreesTerminate","width","validateRows","rows","started","scores","row","prepared","preprocess","scaled","index","coef","intercept","features","outputs","output","total","base","feature","threshold","left","right","leaf","offsets","trees","totals","tree","node","column","slot","link","classes","probabilities","positive","highest","exponentiated","value","sum","numeric","best","label","response","error","ModelFetchError","nodes","broken","detail","start","end","lower","upper","FeatureShapeError","ragged","expected"],"sources":["../../src/tabular/compact.ts"],"sourcesContent":["/**\n * @tempest-limits file-lines — the compact model format: the header, the column\n * dictionary, the quantised weights and the decoder that reads them back. Writer and\n * reader in one file because a format whose two halves live apart drifts.\n */\n/**\n * Running a model with no inference runtime at all.\n *\n * ONNX in the browser costs a **25.6 MB WebAssembly runtime** (6.0 MB\n * gzipped) before the first prediction, while the model itself is a few\n * hundred kilobytes. For an app whose only model is tabular, that runtime\n * *is* the download — so this reader replaces it with 1.49 KB of\n * arithmetic (measured, brotli).\n *\n * A linear model is a dot product. A tree is a chain of comparisons. That\n * is the whole implementation; there is nothing here that a WebAssembly\n * kernel would do better at this size.\n *\n * The file it reads (`.tmc`, magic `TMC1`) is written by\n * `tempest-fastapi-sdk`'s `export_sklearn_to_compact`, which verifies the\n * bytes against scikit-learn's own predictions and refuses to write a file\n * that disagrees. **It is data, not code**: no `eval`, no generated\n * JavaScript, nothing a strict CSP forbids.\n *\n * The trade against {@link TabularPredictor}: ONNX covers every estimator,\n * this covers linear models and tree ensembles. Pick by what your app\n * already ships — a page that loads ONNX for vision pays nothing extra for\n * tabular ONNX.\n */\n\nimport { CompactFormatError, FeatureShapeError, ModelFetchError } from \"./exceptions\";\nimport type { FeatureRow, PredictedLabel, TabularPrediction } from \"./types\";\n\n/** Magic bytes opening every compact file. */\nconst MAGIC = \"TMC1\";\n\n/** Layout version this reader understands. */\nexport const SUPPORTED_COMPACT_SCHEMA = 1;\n\n/** What the file holds. */\nexport type CompactKind = \"linear\" | \"tree_ensemble\";\n\n/** How raw scores become probabilities. */\ntype CompactLink = \"softmax\" | \"sigmoid\" | \"normalize\" | \"identity\";\n\n/** One array stored after the header. */\ninterface CompactSection {\n    readonly name: string;\n    readonly dtype: \"float32\" | \"int32\";\n    readonly length: number;\n}\n\n/** The parsed header. */\ninterface CompactHeader {\n    readonly schema_version: number;\n    readonly kind: CompactKind;\n    readonly task: \"classification\" | \"regression\";\n    readonly link: CompactLink;\n    readonly classes: readonly string[];\n    /** How scikit-learn typed those labels: ``int``, ``float`` or ``str``. */\n    readonly class_type?: \"int\" | \"float\" | \"str\";\n    readonly n_features: number;\n    readonly n_outputs: number;\n    readonly n_trees?: number;\n    readonly estimator: string;\n    readonly feature_names: readonly string[];\n    readonly preprocess: { readonly offset: number[]; readonly scale: number[] } | null;\n    readonly sections: readonly CompactSection[];\n}\n\n/** What a loaded compact model is. */\nexport interface CompactPredictorInfo {\n    /** Which reader path the file uses. */\n    readonly kind: CompactKind;\n    /** Class labels in score-column order; empty for a regressor. */\n    readonly classes: readonly string[];\n    /** Values expected per row. */\n    readonly numFeatures: number;\n    /** Column order recorded at training time, when the export had one. */\n    readonly featureNames: readonly string[];\n    /** Trees in the ensemble; `0` for a linear model. */\n    readonly numTrees: number;\n    /** Whether the model produces class scores. */\n    readonly isClassifier: boolean;\n    /** Class name of the exported estimator. */\n    readonly estimator: string;\n}\n\n/**\n * A compact model, loaded and ready to answer.\n *\n * @example\n * ```ts\n * const predictor = await CompactPredictor.create(\"/models/risk.tmc\");\n * const { labels, probabilities } = await predictor.predict([[5.1, 3.5, 1.4, 0.2]]);\n * ```\n */\nexport class CompactPredictor {\n    private constructor(\n        private readonly header: CompactHeader,\n        private readonly arrays: Record<string, Float32Array | Int32Array>,\n        /** What is loaded. */\n        public readonly info: CompactPredictorInfo,\n    ) {}\n\n    /**\n     * Load a `.tmc` file.\n     *\n     * @param source A URL, or the bytes when the app already has them.\n     * @param requestInit `fetch` options, when `source` is a URL.\n     * @returns The loaded predictor.\n     * @throws {@link ModelFetchError} when a URL cannot be read.\n     * @throws {@link CompactFormatError} when the bytes are not a compact\n     *   model, or use a layout newer than this reader.\n     */\n    static async create(\n        source: string | ArrayBuffer | Uint8Array,\n        requestInit?: RequestInit,\n    ): Promise<CompactPredictor> {\n        const buffer = await toBuffer(source, requestInit);\n        const bytes = new Uint8Array(buffer);\n\n        if (String.fromCharCode(...bytes.subarray(0, 4)) !== MAGIC) {\n            throw new CompactFormatError(\n                \"This is not a compact model file: it does not start with \" +\n                    `\"${MAGIC}\". A .onnx file goes through TabularPredictor instead.`,\n            );\n        }\n\n        const view = new DataView(buffer);\n        const headerLength = view.getUint32(4, true);\n        const header = JSON.parse(\n            new TextDecoder().decode(bytes.subarray(8, 8 + headerLength)),\n        ) as CompactHeader;\n\n        if (header.schema_version > SUPPORTED_COMPACT_SCHEMA) {\n            throw new CompactFormatError(\n                `This file uses compact layout ${header.schema_version}, and this ` +\n                    `SDK understands ${SUPPORTED_COMPACT_SCHEMA}. Upgrade ` +\n                    \"tempest-react-sdk before serving it.\",\n            );\n        }\n\n        const arrays: Record<string, Float32Array | Int32Array> = {};\n        let cursor = 8 + headerLength;\n        for (const section of header.sections) {\n            arrays[section.name] =\n                section.dtype === \"float32\"\n                    ? new Float32Array(buffer, cursor, section.length)\n                    : new Int32Array(buffer, cursor, section.length);\n            cursor += section.length * 4;\n        }\n\n        if (header.kind === \"tree_ensemble\") assertTreesTerminate(header, arrays);\n\n        return new CompactPredictor(header, arrays, {\n            kind: header.kind,\n            classes: header.classes,\n            numFeatures: header.n_features,\n            featureNames: header.feature_names,\n            numTrees: header.n_trees ?? 0,\n            isClassifier: header.task === \"classification\",\n            estimator: header.estimator,\n        });\n    }\n\n    /**\n     * Predict for a batch of rows.\n     *\n     * @param rows One array of feature values per row, in training column\n     *   order. A single row is still wrapped: `[[...]]`.\n     * @returns Labels, class scores when the model is a classifier, and the\n     *   call's duration.\n     * @throws {@link FeatureShapeError} when the batch is empty, ragged, or\n     *   the wrong width.\n     */\n    async predict(rows: readonly FeatureRow[]): Promise<TabularPrediction> {\n        const width = validateRows(rows, this.header.n_features);\n        const started = performance.now();\n\n        const scores: number[][] = [];\n        for (const row of rows) {\n            const prepared = this.preprocess(row, width);\n            scores.push(\n                this.header.kind === \"linear\"\n                    ? this.linearScores(prepared)\n                    : this.treeScores(prepared),\n            );\n        }\n\n        const { labels, probabilities } = this.finish(scores);\n        return {\n            labels,\n            probabilities,\n            numRows: rows.length,\n            ms: performance.now() - started,\n        };\n    }\n\n    /** Releasing nothing, so callers can swap predictors without branching. */\n    async dispose(): Promise<void> {\n        /* no runtime, nothing to release */\n    }\n\n    /**\n     * Apply the folded scaler, when the export had one.\n     *\n     * @param row The raw feature values.\n     * @param width How many there are.\n     * @returns The values the model was trained on.\n     */\n    private preprocess(row: FeatureRow, width: number): number[] {\n        const preprocess = this.header.preprocess;\n        if (preprocess === null) return row as number[];\n        const scaled = new Array<number>(width);\n        for (let index = 0; index < width; index += 1) {\n            scaled[index] =\n                ((row[index] as number) - (preprocess.offset[index] as number)) /\n                (preprocess.scale[index] as number);\n        }\n        return scaled;\n    }\n\n    /**\n     * Score one row against the coefficient matrix.\n     *\n     * @param row The prepared feature values.\n     * @returns One raw score per output.\n     */\n    private linearScores(row: readonly number[]): number[] {\n        const coef = this.arrays.coef as Float32Array;\n        const intercept = this.arrays.intercept as Float32Array;\n        const features = this.header.n_features;\n        const outputs = this.header.n_outputs;\n\n        const scores = new Array<number>(outputs);\n        for (let output = 0; output < outputs; output += 1) {\n            let total = intercept[output] as number;\n            const base = output * features;\n            for (let index = 0; index < features; index += 1) {\n                total += (coef[base + index] as number) * (row[index] as number);\n            }\n            scores[output] = total;\n        }\n        return scores;\n    }\n\n    /**\n     * Walk every tree and average what the leaves hold.\n     *\n     * A leaf is marked by a negative `feature` entry, which also carries\n     * its slot in the value array — so the walk needs no second lookup and\n     * the file stores values only for leaves.\n     *\n     * **The comparison runs in float32** (`Math.fround`), because that is\n     * what scikit-learn does: it casts its input to float32 before\n     * traversing, so a threshold like 5.099999904632568 — a float32 value\n     * widened for storage — and an input of 5.1 compare *equal* and go\n     * left. Comparing in float64 sends that row right instead, which on an\n     * iris forest changed one tree's vote and moved a probability by 0.05.\n     *\n     * @param row The prepared feature values.\n     * @returns One averaged score per output.\n     */\n    private treeScores(row: readonly number[]): number[] {\n        const feature = this.arrays.node_feature as Int32Array;\n        const threshold = this.arrays.node_threshold as Float32Array;\n        const left = this.arrays.node_left as Int32Array;\n        const right = this.arrays.node_right as Int32Array;\n        const leaf = this.arrays.leaf_value as Float32Array;\n        const offsets = this.arrays.tree_offset as Int32Array;\n        const outputs = this.header.n_outputs;\n        const trees = offsets.length - 1;\n\n        const totals = new Array<number>(outputs).fill(0);\n        for (let tree = 0; tree < trees; tree += 1) {\n            let node = offsets[tree] as number;\n            let column = feature[node] as number;\n            while (column >= 0) {\n                node =\n                    Math.fround(row[column] as number) <= (threshold[node] as number)\n                        ? (left[node] as number)\n                        : (right[node] as number);\n                column = feature[node] as number;\n            }\n            const slot = (-1 - column) * outputs;\n            for (let output = 0; output < outputs; output += 1) {\n                totals[output] += leaf[slot + output] as number;\n            }\n        }\n        for (let output = 0; output < outputs; output += 1) {\n            totals[output] = (totals[output] as number) / trees;\n        }\n        return totals;\n    }\n\n    /**\n     * Turn raw scores into labels and probabilities.\n     *\n     * @param scores One score array per row.\n     * @returns Labels and probabilities in the shape the ONNX route uses,\n     *   so an app can swap runtimes without touching its own code. That\n     *   includes the label's **type**: an integer class comes back as a\n     *   number here exactly as ONNX returns it, because two routes over one\n     *   model that disagree on `0` versus `\"0\"` break the day someone\n     *   switches.\n     */\n    private finish(scores: readonly number[][]): {\n        labels: PredictedLabel[];\n        probabilities: number[][];\n    } {\n        const link = this.header.link;\n        const classes = this.header.classes;\n\n        if (link === \"identity\") {\n            return { labels: scores.map((row) => row[0] as number), probabilities: [] };\n        }\n\n        const probabilities = scores.map((row) => {\n            if (link === \"sigmoid\") {\n                const positive = 1 / (1 + Math.exp(-(row[0] as number)));\n                return [1 - positive, positive];\n            }\n            if (link === \"softmax\") {\n                const highest = Math.max(...row);\n                const exponentiated = row.map((value) => Math.exp(value - highest));\n                const total = exponentiated.reduce((sum, value) => sum + value, 0);\n                return exponentiated.map((value) => value / total);\n            }\n            const total = row.reduce((sum, value) => sum + value, 0);\n            return total === 0 ? [...row] : row.map((value) => value / total);\n        });\n\n        const numeric = this.header.class_type !== \"str\";\n        const labels = probabilities.map((row) => {\n            let best = 0;\n            for (let index = 1; index < row.length; index += 1) {\n                if ((row[index] as number) > (row[best] as number)) best = index;\n            }\n            const label = classes[best];\n            if (label === undefined) return best;\n            return numeric ? Number(label) : label;\n        });\n\n        return { labels, probabilities };\n    }\n}\n\n/**\n * Read a source into an `ArrayBuffer`.\n *\n * @param source A URL or the bytes.\n * @param requestInit `fetch` options.\n * @returns The bytes.\n * @throws {@link ModelFetchError} when the URL cannot be read.\n */\nasync function toBuffer(\n    source: string | ArrayBuffer | Uint8Array,\n    requestInit?: RequestInit,\n): Promise<ArrayBuffer> {\n    if (typeof source === \"string\") {\n        let response: Response;\n        try {\n            response = await fetch(source, requestInit);\n        } catch (error) {\n            throw new ModelFetchError(`Could not download the model: ${source}`, {\n                cause: error,\n            });\n        }\n        if (!response.ok) {\n            throw new ModelFetchError(\n                `Could not download the model: ${response.status} ${response.statusText}`,\n            );\n        }\n        return await response.arrayBuffer();\n    }\n    if (source instanceof Uint8Array) {\n        return source.buffer.slice(\n            source.byteOffset,\n            source.byteOffset + source.byteLength,\n        ) as ArrayBuffer;\n    }\n    return source;\n}\n\n/**\n * Prove once, at load, that every tree walk ends inside its arrays.\n *\n * The walk in `treeScores` follows `node_left`/`node_right` until it meets a\n * leaf, with no step limit — a limit there would be paid on every node of every\n * row. So a file whose child index points back up the tree spun the tab forever\n * on the first `predict`, and one pointing past the arrays read `undefined` and\n * answered `NaN`. scikit-learn numbers a node before its children, and\n * `export_sklearn_to_compact` keeps that order with each tree's nodes contiguous,\n * so a well-formed file satisfies `node < child < next tree's offset` everywhere\n * — measured on every tree fixture in `__fixtures__` (1 761 nodes, 0 violations).\n * Checking that invariant is linear in the node count and rules out a cycle by\n * construction, since every step strictly increases the index.\n *\n * @param header The parsed header.\n * @param arrays The sections, by name.\n * @throws {@link CompactFormatError} when a tree could loop, leave its arrays,\n *   read a feature the model does not have, or a leaf slot past the values.\n */\nfunction assertTreesTerminate(\n    header: CompactHeader,\n    arrays: Record<string, Float32Array | Int32Array>,\n): void {\n    const feature = arrays.node_feature;\n    const left = arrays.node_left;\n    const right = arrays.node_right;\n    const leaf = arrays.leaf_value;\n    const offsets = arrays.tree_offset;\n    if (!feature || !left || !right || !leaf || !offsets || offsets.length < 2) {\n        throw new CompactFormatError(\"This tree model is missing its node or offset sections.\");\n    }\n    const nodes = feature.length;\n    const outputs = header.n_outputs;\n    const broken = (detail: string): CompactFormatError =>\n        new CompactFormatError(`This tree model is malformed: ${detail}. Re-export it.`);\n\n    if (left.length !== nodes || right.length !== nodes || offsets[0] !== 0) {\n        throw broken(\"the node sections disagree on the node count\");\n    }\n    if (offsets[offsets.length - 1] !== nodes) {\n        throw broken(\"the last tree does not end at the last node\");\n    }\n    for (let tree = 0; tree < offsets.length - 1; tree += 1) {\n        const start = offsets[tree] as number;\n        const end = offsets[tree + 1] as number;\n        if (end <= start) throw broken(`tree ${tree} has no nodes`);\n        for (let node = start; node < end; node += 1) {\n            const column = feature[node] as number;\n            if (column >= 0) {\n                const lower = left[node] as number;\n                const upper = right[node] as number;\n                if (column >= header.n_features) {\n                    throw broken(`node ${node} splits on feature ${column}`);\n                }\n                if (lower <= node || upper <= node || lower >= end || upper >= end) {\n                    throw broken(`node ${node} points outside its tree or back up it`);\n                }\n            } else if ((-1 - column) * outputs + outputs > leaf.length) {\n                throw broken(`leaf ${node} reads past the leaf values`);\n            }\n        }\n    }\n}\n\n/**\n * Check a batch before predicting on it.\n *\n * @param rows The batch.\n * @param expected Features the model wants per row.\n * @returns The batch width.\n * @throws {@link FeatureShapeError} when the batch cannot be predicted on.\n */\nfunction validateRows(rows: readonly FeatureRow[], expected: number): number {\n    if (!Array.isArray(rows) || rows.length === 0) {\n        throw new FeatureShapeError(\"predict() needs at least one row, shaped [[f1, f2, ...]].\");\n    }\n    const width = rows[0]?.length ?? 0;\n    const ragged = rows.findIndex((row) => row.length !== width);\n    if (ragged !== -1) {\n        throw new FeatureShapeError(\n            `All rows must have the same width; row ${ragged} has ` +\n                `${rows[ragged]?.length} values, expected ${width}.`,\n        );\n    }\n    if (width !== expected) {\n        throw new FeatureShapeError(\n            `The model expects ${expected} features per row, got ${width}.`,\n        );\n    }\n    return width;\n}\n"],"mappings":";;AAkCA,IAAMA,IAAQ,QA+DDC,IAAb,MAAaA,EAAiB;CAEL;CACA;CAED;CAJpB,YACI,GACA,GAEA,GACF;EADkB,AAHC,KAAA,SAAAC,GACA,KAAA,SAAAC,GAED,KAAA,OAAAC;CACjB;CAYH,aAAa,OACT,GACA,GACyB;EACzB,IAAMC,IAAS,MAAMC,EAASC,GAAQC,CAAW,GAC3CC,IAAQ,IAAI,WAAWJ,CAAM;EAEnC,IAAI,OAAO,aAAa,GAAGI,EAAM,SAAS,GAAG,CAAC,CAAC,MAAMT,GACjD,MAAM,IAAIU,EACN,6DACQV,EAAM,uDAClB;EAIJ,IAAMW,IAAe,IADJ,SAASN,CACL,CAAA,CAAK,UAAU,GAAG,EAAI,GACrCH,IAAS,KAAK,MAChB,IAAI,YAAY,CAAC,CAAC,OAAOO,EAAM,SAAS,GAAG,IAAIE,CAAY,CAAC,CAChE;EAEA,IAAIT,EAAO,iBAAA,GACP,MAAM,IAAIQ,EACN,iCAAiCR,EAAO,eAAe,2EAG3D;EAGJ,IAAMC,IAAoD,CAAC,GACvDS,IAAS,IAAID;EACjB,KAAK,IAAME,KAAWX,EAAO,UAKzB,AAJA,EAAOW,EAAQ,QACXA,EAAQ,UAAU,YACZ,IAAI,aAAaR,GAAQO,GAAQC,EAAQ,MAAM,IAC/C,IAAI,WAAWR,GAAQO,GAAQC,EAAQ,MAAM,GACvD,KAAUA,EAAQ,SAAS;EAK/B,OAFIX,EAAO,SAAS,mBAAiBY,EAAqBZ,GAAQC,CAAM,GAEjE,IAAIF,EAAiBC,GAAQC,GAAQ;GACxC,MAAMD,EAAO;GACb,SAASA,EAAO;GAChB,aAAaA,EAAO;GACpB,cAAcA,EAAO;GACrB,UAAUA,EAAO,WAAW;GAC5B,cAAcA,EAAO,SAAS;GAC9B,WAAWA,EAAO;EACtB,CAAC;CACL;CAYA,MAAM,QAAQ,GAAyD;EACnE,IAAMa,IAAQC,EAAaC,GAAM,KAAK,OAAO,UAAU,GACjDC,IAAU,YAAY,IAAI,GAE1BC,IAAqB,CAAC;EAC5B,KAAK,IAAMC,KAAOH,GAAM;GACpB,IAAMI,IAAW,KAAK,WAAWD,GAAKL,CAAK;GAC3C,EAAO,KACH,KAAK,OAAO,SAAS,WACf,KAAK,aAAaM,CAAQ,IAC1B,KAAK,WAAWA,CAAQ,CAClC;EACJ;EAEA,IAAM,EAAE,WAAQ,qBAAkB,KAAK,OAAOF,CAAM;EACpD,OAAO;GACH;GACA;GACA,SAASF,EAAK;GACd,IAAI,YAAY,IAAI,IAAIC;EAC5B;CACJ;CAGA,MAAM,UAAyB,CAE/B;CASA,WAAmB,GAAiB,GAAyB;EACzD,IAAMI,IAAa,KAAK,OAAO;EAC/B,IAAIA,MAAe,MAAM,OAAOF;EAChC,IAAMG,IAAa,MAAcR,CAAK;EACtC,KAAK,IAAIS,IAAQ,GAAGA,IAAQT,GAAO,KAAS,GACxC,EAAOS,MACDJ,EAAII,KAAqBF,EAAW,OAAOE,MAC5CF,EAAW,MAAME;EAE1B,OAAOD;CACX;CAQA,aAAqB,GAAkC;EACnD,IAAME,IAAO,KAAK,OAAO,MACnBC,IAAY,KAAK,OAAO,WACxBC,IAAW,KAAK,OAAO,YACvBC,IAAU,KAAK,OAAO,WAEtBT,IAAa,MAAcS,CAAO;EACxC,KAAK,IAAIC,IAAS,GAAGA,IAASD,GAAS,KAAU,GAAG;GAChD,IAAIE,IAAQJ,EAAUG,IAChBE,IAAOF,IAASF;GACtB,KAAK,IAAIH,IAAQ,GAAGA,IAAQG,GAAU,KAAS,GAC3C,KAAUF,EAAKM,IAAOP,KAAqBJ,EAAII;GAEnD,EAAOK,KAAUC;EACrB;EACA,OAAOX;CACX;CAmBA,WAAmB,GAAkC;EACjD,IAAMa,IAAU,KAAK,OAAO,cACtBC,IAAY,KAAK,OAAO,gBACxBC,IAAO,KAAK,OAAO,WACnBC,IAAQ,KAAK,OAAO,YACpBC,IAAO,KAAK,OAAO,YACnBC,IAAU,KAAK,OAAO,aACtBT,IAAU,KAAK,OAAO,WACtBU,IAAQD,EAAQ,SAAS,GAEzBE,IAAa,MAAcX,CAAO,CAAC,CAAC,KAAK,CAAC;EAChD,KAAK,IAAIY,IAAO,GAAGA,IAAOF,GAAO,KAAQ,GAAG;GACxC,IAAIG,IAAOJ,EAAQG,IACfE,IAASV,EAAQS;GACrB,OAAOC,KAAU,IAKb,AAJA,IACI,KAAK,OAAOtB,EAAIsB,EAAiB,KAAMT,EAAUQ,KAC1CP,EAAKO,KACLN,EAAMM,IACjB,IAAST,EAAQS;GAErB,IAAME,KAAQ,KAAKD,KAAUd;GAC7B,KAAK,IAAIC,IAAS,GAAGA,IAASD,GAAS,KAAU,GAC7C,EAAOC,MAAWO,EAAKO,IAAOd;EAEtC;EACA,KAAK,IAAIA,IAAS,GAAGA,IAASD,GAAS,KAAU,GAC7C,EAAOC,KAAWU,EAAOV,KAAqBS;EAElD,OAAOC;CACX;CAaA,OAAe,GAGb;EACE,IAAMK,IAAO,KAAK,OAAO,MACnBC,IAAU,KAAK,OAAO;EAE5B,IAAID,MAAS,YACT,OAAO;GAAE,QAAQzB,EAAO,KAAK,MAAQC,EAAI,EAAY;GAAG,eAAe,CAAC;EAAE;EAG9E,IAAM0B,IAAgB3B,EAAO,KAAK,MAAQ;GACtC,IAAIyB,MAAS,WAAW;IACpB,IAAMG,IAAW,KAAK,IAAI,KAAK,IAAI,CAAE3B,EAAI,EAAa;IACtD,OAAO,CAAC,IAAI2B,GAAUA,CAAQ;GAClC;GACA,IAAIH,MAAS,WAAW;IACpB,IAAMI,IAAU,KAAK,IAAI,GAAG5B,CAAG,GACzB6B,IAAgB7B,EAAI,KAAK,MAAU,KAAK,IAAI8B,IAAQF,CAAO,CAAC,GAC5DlB,IAAQmB,EAAc,QAAQ,GAAK,MAAUE,IAAMD,GAAO,CAAC;IACjE,OAAOD,EAAc,KAAK,MAAUC,IAAQpB,CAAK;GACrD;GACA,IAAMA,IAAQV,EAAI,QAAQ,GAAK,MAAU+B,IAAMD,GAAO,CAAC;GACvD,OAAOpB,MAAU,IAAI,CAAC,GAAGV,CAAG,IAAIA,EAAI,KAAK,MAAU8B,IAAQpB,CAAK;EACpE,CAAC,GAEKsB,IAAU,KAAK,OAAO,eAAe;EAW3C,OAAO;GAAE,QAVMN,EAAc,KAAK,MAAQ;IACtC,IAAIO,IAAO;IACX,KAAK,IAAI7B,IAAQ,GAAGA,IAAQJ,EAAI,QAAQ,KAAS,GAC7C,AAAKA,EAAII,KAAqBJ,EAAIiC,OAAkB,IAAO7B;IAE/D,IAAM8B,IAAQT,EAAQQ;IAEtB,OADIC,MAAU,KAAA,IAAkBD,IACzBD,IAAU,OAAOE,CAAK,IAAIA;GACrC,CAES;GAAQ;EAAc;CACnC;AACJ;AAUA,eAAehD,EACX,GACA,GACoB;CACpB,IAAI,OAAOC,KAAW,UAAU;EAC5B,IAAIgD;EACJ,IAAI;GACA,IAAW,MAAM,MAAMhD,GAAQC,CAAW;EAC9C,SAASgD,GAAO;GACZ,MAAM,IAAIC,EAAgB,iCAAiClD,KAAU,EACjE,OAAOiD,EACX,CAAC;EACL;EACA,IAAI,CAACD,EAAS,IACV,MAAM,IAAIE,EACN,iCAAiCF,EAAS,OAAO,GAAGA,EAAS,YACjE;EAEJ,OAAO,MAAMA,EAAS,YAAY;CACtC;CAOA,OANIhD,aAAkB,aACXA,EAAO,OAAO,MACjBA,EAAO,YACPA,EAAO,aAAaA,EAAO,UAC/B,IAEGA;AACX;AAqBA,SAASO,EACL,GACA,GACI;CACJ,IAAMkB,IAAU7B,EAAO,cACjB+B,IAAO/B,EAAO,WACdgC,IAAQhC,EAAO,YACfiC,IAAOjC,EAAO,YACdkC,IAAUlC,EAAO;CACvB,IAAI,CAAC6B,KAAW,CAACE,KAAQ,CAACC,KAAS,CAACC,KAAQ,CAACC,KAAWA,EAAQ,SAAS,GACrE,MAAM,IAAI3B,EAAmB,yDAAyD;CAE1F,IAAMgD,IAAQ1B,EAAQ,QAChBJ,IAAU1B,EAAO,WACjByD,KAAU,MACZ,IAAIjD,EAAmB,iCAAiCkD,EAAO,gBAAgB;CAEnF,IAAI1B,EAAK,WAAWwB,KAASvB,EAAM,WAAWuB,KAASrB,EAAQ,OAAO,GAClE,MAAMsB,EAAO,8CAA8C;CAE/D,IAAItB,EAAQA,EAAQ,SAAS,OAAOqB,GAChC,MAAMC,EAAO,6CAA6C;CAE9D,KAAK,IAAInB,IAAO,GAAGA,IAAOH,EAAQ,SAAS,GAAG,KAAQ,GAAG;EACrD,IAAMwB,IAAQxB,EAAQG,IAChBsB,IAAMzB,EAAQG,IAAO;EAC3B,IAAIsB,KAAOD,GAAO,MAAMF,EAAO,QAAQnB,EAAK,cAAc;EAC1D,KAAK,IAAIC,IAAOoB,GAAOpB,IAAOqB,GAAK,KAAQ,GAAG;GAC1C,IAAMpB,IAASV,EAAQS;GACvB,IAAIC,KAAU,GAAG;IACb,IAAMqB,IAAQ7B,EAAKO,IACbuB,IAAQ7B,EAAMM;IACpB,IAAIC,KAAUxC,EAAO,YACjB,MAAMyD,EAAO,QAAQlB,EAAK,qBAAqBC,GAAQ;IAE3D,IAAIqB,KAAStB,KAAQuB,KAASvB,KAAQsB,KAASD,KAAOE,KAASF,GAC3D,MAAMH,EAAO,QAAQlB,EAAK,uCAAuC;GAEzE,OAAO,KAAK,KAAKC,KAAUd,IAAUA,IAAUQ,EAAK,QAChD,MAAMuB,EAAO,QAAQlB,EAAK,4BAA4B;EAE9D;CACJ;AACJ;AAUA,SAASzB,EAAa,GAA6B,GAA0B;CACzE,IAAI,CAAC,MAAM,QAAQC,CAAI,KAAKA,EAAK,WAAW,GACxC,MAAM,IAAIgD,EAAkB,2DAA2D;CAE3F,IAAMlD,IAAQE,EAAK,EAAE,EAAE,UAAU,GAC3BiD,IAASjD,EAAK,WAAW,MAAQG,EAAI,WAAWL,CAAK;CAC3D,IAAImD,MAAW,IACX,MAAM,IAAID,EACN,0CAA0CC,EAAO,OAC1CjD,EAAKiD,EAAO,EAAE,OAAO,oBAAoBnD,EAAM,EAC1D;CAEJ,IAAIA,MAAUoD,GACV,MAAM,IAAIF,EACN,qBAAqBE,EAAS,yBAAyBpD,EAAM,EACjE;CAEJ,OAAOA;AACX"}

package/dist/tabular/manifest.cjs

@@ -1,2 +1,2 @@
-const e=require("./exceptions.cjs"),t=require("./cache.cjs"),n=require("./compact.cjs"),r=require("./predictor.cjs");var i=`manifest.json`;async function a(t,n){let r=l(t),i;try{i=await fetch(r,n)}catch(t){throw new e.ModelFetchError(`Could not read the manifest at ${r}`,{cause:t})}if(!i.ok)throw new e.ModelFetchError(`Could not read the manifest at ${r}: ${i.status} ${i.statusText}`);let a=await i.json();if(typeof a?.schema_version!=`number`)throw new e.ModelFetchError(`${r} is not an edge package manifest (no schema_version).`);if(a.schema_version>1)throw new e.ModelFetchError(`The manifest at ${r} uses schema_version ${a.schema_version}, and this SDK understands 1. Upgrade tempest-react-sdk before serving this package.`);return a}async function o(e,i={}){let o=await a(e),l=e.endsWith(`/`)?e:`${e}/`,u=s(o,i.runtime??`auto`),d=`${l}${c(o,u)}`,f=i.cache??!0,p=f===!1?d:await t.fetchModelBytes(d,typeof f==`object`?f:{}),m=u===`compact`?await n.CompactPredictor.create(p):await r.TabularPredictor.create(p,{providers:i.providers,warmup:i.warmup,sessionOptions:i.sessionOptions}),h=o.output.classes;return{manifest:o,predictor:m,runtime:u,featureNames:o.input.feature_names,classes:h,explain:e=>e.map((e,t)=>({name:h[t]??String(t),score:e})).sort((e,t)=>t.score-e.score)}}function s(t,n){let r=t.runtimes??[],i=r.some(e=>e.kind===`compact`);if(n===`auto`)return i?`compact`:`onnx`;if(n===`compact`&&!i)throw new e.ModelFetchError(`This package carries no compact model (${t.name} lists ${r.map(e=>e.kind).join(`, `)||`onnx`}). Re-export it with edge_pipeline(compact=True), or load it as ONNX.`);return n}function c(e,t){return(e.runtimes??[]).find(e=>e.kind===t)?.file??e.model.file}function l(e){return e.endsWith(`.json`)?e:e.endsWith(`/`)?`${e}${i}`:`${e}/${i}`}exports.MANIFEST_FILENAME=i,exports.fetchEdgeManifest=a,exports.loadEdgePackage=o;
+const e=require("./exceptions.cjs"),t=require("./cache.cjs"),n=require("./compact.cjs"),r=require("./predictor.cjs");var i=`manifest.json`;async function a(t,n){let r=d(t),i;try{i=await fetch(r,n)}catch(t){throw new e.ModelFetchError(`Could not read the manifest at ${r}`,{cause:t})}if(!i.ok)throw new e.ModelFetchError(`Could not read the manifest at ${r}: ${i.status} ${i.statusText}`);let a=await i.json();if(typeof a?.schema_version!=`number`)throw new e.ModelFetchError(`${r} is not an edge package manifest (no schema_version).`);if(a.schema_version>1)throw new e.ModelFetchError(`The manifest at ${r} uses schema_version ${a.schema_version}, and this SDK understands 1. Upgrade tempest-react-sdk before serving this package.`);return a}async function o(e,t={}){let i=await a(e),o=e.endsWith(`/`)?e:`${e}/`,l=s(i,t.runtime??`auto`),{file:d,sha256:f}=c(i,l),p=`${o}${d}`,m=t.cache??!0,h=m===!1?p:await u(p,f,typeof m==`object`?m:{}),g=l===`compact`?await n.CompactPredictor.create(h):await r.TabularPredictor.create(h,{providers:t.providers,warmup:t.warmup,sessionOptions:t.sessionOptions}),_=i.output.classes;return{manifest:i,predictor:g,runtime:l,featureNames:i.input.feature_names,classes:_,explain:e=>e.map((e,t)=>({name:_[t]??String(t),score:e})).sort((e,t)=>t.score-e.score)}}function s(t,n){let r=t.runtimes??[],i=r.some(e=>e.kind===`compact`);if(n===`auto`)return i?`compact`:`onnx`;if(n===`compact`&&!i)throw new e.ModelFetchError(`This package carries no compact model (${t.name} lists ${r.map(e=>e.kind).join(`, `)||`onnx`}). Re-export it with edge_pipeline(compact=True), or load it as ONNX.`);return n}function c(e,t){return(e.runtimes??[]).find(e=>e.kind===t)??e.model}async function l(e,t){let n=globalThis.crypto?.subtle;if(!n||!t)return!0;let r=new Uint8Array(await n.digest(`SHA-256`,e));return Array.from(r,e=>e.toString(16).padStart(2,`0`)).join(``)===t.toLowerCase()}async function u(n,r,i){let a=await t.fetchModelBytes(n,i);if(await l(a,r))return a;if(i.revalidate!==!0){let e=await t.fetchModelBytes(n,{...i,revalidate:!0});if(await l(e,r))return e}throw new e.ModelFetchError(`The model at ${n} does not match the sha256 in its manifest. The file and the manifest were published separately, or the download was altered — republish the package so both describe the same model.`)}function d(e){return e.endsWith(`.json`)?e:e.endsWith(`/`)?`${e}${i}`:`${e}/${i}`}exports.MANIFEST_FILENAME=i,exports.fetchEdgeManifest=a,exports.loadEdgePackage=o;
 //# sourceMappingURL=manifest.cjs.map

package/dist/tabular/manifest.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"manifest.cjs","names":["ModelFetchError","fetchModelBytes","CompactPredictor","TabularPredictor","require_exceptions","require_cache","require_compact","require_predictor","MANIFEST_FILENAME","fetchEdgeManifest","url","manifestUrl","directoryUrl","response","requestInit","error","manifest","loadEdgePackage","base","runtime","chooseRuntime","options","modelUrl","fileFor","cache","source","predictor","classes","probabilities","index","b","a","available","hasCompact","entry","requested","item"],"sources":["../../src/tabular/manifest.ts"],"sourcesContent":["/**\n * Reading an edge package published by `tempest-fastapi-sdk`.\n *\n * `edge_pipeline` writes a directory — the graph, a gzipped copy, a drift\n * baseline and a `manifest.json` — and the browser is one of its two\n * intended consumers. The manifest is what turns a `.onnx` URL into\n * something a UI can use: the **column order** the model was trained on,\n * the class names behind `probabilities[2]`, and a version to compare\n * against what is already cached.\n *\n * That column order is the field worth loading a manifest for. A model fed\n * the right features in the wrong order answers confidently and wrongly,\n * and no runtime check catches it.\n *\n * The contract is pinned by `schema_version`. Unknown fields are ignored\n * rather than rejected, so a package written by a newer SDK still loads.\n */\n\nimport { fetchModelBytes, type ModelCacheOptions } from \"./cache\";\nimport { CompactPredictor } from \"./compact\";\nimport { ModelFetchError } from \"./exceptions\";\nimport { TabularPredictor } from \"./predictor\";\nimport type { FeatureRow, TabularPrediction, TabularPredictorOptions } from \"./types\";\n\n/** Manifest schema version this reader was written against. */\nexport const SUPPORTED_MANIFEST_SCHEMA = 1;\n\n/** Fixed filename inside a package directory. */\nexport const MANIFEST_FILENAME = \"manifest.json\";\n\n/** The graph file and how to check you got it whole. */\nexport interface ManifestModelFile {\n    readonly file: string;\n    readonly sha256: string;\n    readonly bytes: number;\n    readonly gzip_file: string | null;\n    readonly gzip_bytes: number | null;\n    readonly opset: number;\n    readonly dtype: string;\n}\n\n/** What the graph expects per row. */\nexport interface ManifestInput {\n    readonly name: string;\n    readonly features: number;\n    /** Column order used at training time. */\n    readonly feature_names: readonly string[];\n}\n\n/** What the graph answers. */\nexport interface ManifestOutput {\n    readonly is_classifier: boolean;\n    readonly label_output: string;\n    readonly probability_output: string | null;\n    /** Class labels in score-column order. */\n    readonly classes: readonly string[];\n}\n\n/**\n * Where the packaged model came from, when it came from an existing\n * artifact.\n *\n * Present when the package was built with `edge_pipeline_from_pickle`: the\n * `.pkl` never reaches the browser (a pickle is a Python program, not\n * data), but its name and digest travel in the manifest, so a model\n * answering in a tab can be traced back to the file that produced it.\n */\nexport interface ManifestSource {\n    readonly file: string;\n    readonly kind: string;\n    readonly sha256: string;\n    readonly bytes: number;\n    readonly sklearn_version: string;\n    readonly warnings: readonly string[];\n}\n\n/**\n * One file in the package a runtime can load.\n *\n * A package may carry the same model twice — as ONNX, which any runtime\n * reads at the cost of a 25.6 MB WebAssembly download, and as the compact\n * format, which needs no runtime. The list is what lets the browser pick\n * by what it already ships.\n */\nexport interface ManifestRuntime {\n    readonly kind: \"onnx\" | \"compact\" | string;\n    readonly file: string;\n    readonly bytes: number;\n    readonly gzip_file: string | null;\n    readonly gzip_bytes: number | null;\n    readonly sha256: string;\n}\n\n/** The package manifest, as written by `edge_pipeline`. */\nexport interface EdgeManifest {\n    readonly schema_version: number;\n    readonly name: string;\n    readonly version: string;\n    readonly created_at: string;\n    readonly sdk_version: string;\n    readonly estimator: string;\n    readonly model: ManifestModelFile;\n    readonly input: ManifestInput;\n    readonly output: ManifestOutput;\n    readonly verified: boolean | null;\n    /** Every file a runtime can load. Absent on packages written before v0.194. */\n    readonly runtimes?: readonly ManifestRuntime[];\n    /** Absent on packages built straight from a fitted estimator. */\n    readonly source?: ManifestSource;\n    readonly baseline_file: string | null;\n    readonly baseline_samples: number;\n}\n\n/** Which reader served the package. */\nexport type TabularRuntime = \"onnx\" | \"compact\";\n\n/** A package loaded and ready to answer. */\nexport interface LoadedEdgePackage {\n    /** What was published. */\n    readonly manifest: EdgeManifest;\n    /** The running model, whichever runtime read it. */\n    readonly predictor: PredictorLike;\n    /** Which reader was used. */\n    readonly runtime: TabularRuntime;\n    /** Column order the rows must follow. */\n    readonly featureNames: readonly string[];\n    /** Class names behind each probability column. */\n    readonly classes: readonly string[];\n    /**\n     * Map a prediction's scores onto class names.\n     *\n     * @param probabilities One row of scores.\n     * @returns Name/score pairs, highest first.\n     */\n    readonly explain: (probabilities: readonly number[]) => { name: string; score: number }[];\n}\n\n/**\n * The shape both readers share.\n *\n * `TabularPredictor` (ONNX) and `CompactPredictor` (runtime-free) answer\n * with the same object, so an app can switch routes without touching a\n * line of its own code.\n */\nexport interface PredictorLike {\n    predict(rows: readonly FeatureRow[]): Promise<TabularPrediction>;\n    dispose(): Promise<void>;\n}\n\n/** Options for {@link loadEdgePackage}. */\nexport interface LoadEdgePackageOptions extends TabularPredictorOptions {\n    /** Cache the model bytes for offline use. `true` by default. */\n    readonly cache?: boolean | ModelCacheOptions;\n    /**\n     * Which reader to use.\n     *\n     * `\"auto\"` (the default) takes the compact form when the package has\n     * one, because it answers without downloading a WebAssembly runtime.\n     * Force `\"onnx\"` when the app already ships ONNX for something else —\n     * then the runtime is already paid for and ONNX covers more estimators.\n     */\n    readonly runtime?: TabularRuntime | \"auto\";\n}\n\n/**\n * Read a package's manifest.\n *\n * Cheap: it is a few hundred bytes, so an app can check for a new version\n * without downloading a model it may already have.\n *\n * @example\n * ```ts\n * const manifest = await fetchEdgeManifest(\"/models/risk/\");\n * if (manifest.version !== localStorage.getItem(\"risk-version\")) {\n *     // a new model was published\n * }\n * ```\n *\n * @param directoryUrl URL of the package directory, with or without a\n *   trailing slash. A full URL to the manifest file also works.\n * @param requestInit `fetch` options.\n * @returns The parsed manifest.\n * @throws {@link ModelFetchError} when the manifest cannot be read, or when\n *   its `schema_version` is newer than this reader understands — loading it\n *   anyway would risk misreading the field that defines column order.\n */\nexport async function fetchEdgeManifest(\n    directoryUrl: string,\n    requestInit?: RequestInit,\n): Promise<EdgeManifest> {\n    const url = manifestUrl(directoryUrl);\n    let response: Response;\n    try {\n        response = await fetch(url, requestInit);\n    } catch (error) {\n        throw new ModelFetchError(`Could not read the manifest at ${url}`, {\n            cause: error,\n        });\n    }\n    if (!response.ok) {\n        throw new ModelFetchError(\n            `Could not read the manifest at ${url}: ${response.status} ${response.statusText}`,\n        );\n    }\n\n    const manifest = (await response.json()) as EdgeManifest;\n    if (typeof manifest?.schema_version !== \"number\") {\n        throw new ModelFetchError(`${url} is not an edge package manifest (no schema_version).`);\n    }\n    if (manifest.schema_version > SUPPORTED_MANIFEST_SCHEMA) {\n        throw new ModelFetchError(\n            `The manifest at ${url} uses schema_version ${manifest.schema_version}, ` +\n                `and this SDK understands ${SUPPORTED_MANIFEST_SCHEMA}. Upgrade ` +\n                \"tempest-react-sdk before serving this package.\",\n        );\n    }\n    return manifest;\n}\n\n/**\n * Load a whole edge package: manifest, model, and the names to read it by.\n *\n * @example\n * ```ts\n * const pkg = await loadEdgePackage(\"/models/risk/\");\n *\n * console.log(pkg.featureNames); // [\"age\", \"income\", \"tenure\", \"score\", \"visits\"]\n *\n * const { probabilities } = await pkg.predictor.predict([[41, 5200, 3, 0.82, 12]]);\n * console.log(pkg.explain(probabilities[0]!)); // [{ name: \"approved\", score: 0.91 }, ...]\n * ```\n *\n * @param directoryUrl URL of the package directory.\n * @param options Predictor options plus caching.\n * @returns The loaded package.\n * @throws {@link ModelFetchError} when the manifest or model cannot be read.\n */\nexport async function loadEdgePackage(\n    directoryUrl: string,\n    options: LoadEdgePackageOptions = {},\n): Promise<LoadedEdgePackage> {\n    const manifest = await fetchEdgeManifest(directoryUrl);\n    const base = directoryUrl.endsWith(\"/\") ? directoryUrl : `${directoryUrl}/`;\n    const runtime = chooseRuntime(manifest, options.runtime ?? \"auto\");\n    const file = fileFor(manifest, runtime);\n    const modelUrl = `${base}${file}`;\n\n    const cache = options.cache ?? true;\n    const source =\n        cache === false\n            ? modelUrl\n            : await fetchModelBytes(modelUrl, typeof cache === \"object\" ? cache : {});\n\n    const predictor: PredictorLike =\n        runtime === \"compact\"\n            ? await CompactPredictor.create(source)\n            : await TabularPredictor.create(source, {\n                  providers: options.providers,\n                  warmup: options.warmup,\n                  sessionOptions: options.sessionOptions,\n              });\n\n    const classes = manifest.output.classes;\n    return {\n        manifest,\n        predictor,\n        runtime,\n        featureNames: manifest.input.feature_names,\n        classes,\n        explain: (probabilities: readonly number[]) =>\n            probabilities\n                .map((score, index) => ({\n                    name: classes[index] ?? String(index),\n                    score,\n                }))\n                .sort((a, b) => b.score - a.score),\n    };\n}\n\n/**\n * Decide which reader serves this package.\n *\n * @param manifest The package manifest.\n * @param requested What the caller asked for.\n * @returns The runtime to use.\n * @throws {@link ModelFetchError} when the package does not carry the\n *   requested form — asking for a compact model that was never written\n *   should say so, not silently download 25 MB of WebAssembly instead.\n */\nfunction chooseRuntime(manifest: EdgeManifest, requested: TabularRuntime | \"auto\"): TabularRuntime {\n    const available = manifest.runtimes ?? [];\n    const hasCompact = available.some((entry) => entry.kind === \"compact\");\n\n    if (requested === \"auto\") return hasCompact ? \"compact\" : \"onnx\";\n    if (requested === \"compact\" && !hasCompact) {\n        throw new ModelFetchError(\n            `This package carries no compact model (${manifest.name} lists ` +\n                `${available.map((entry) => entry.kind).join(\", \") || \"onnx\"}). ` +\n                \"Re-export it with edge_pipeline(compact=True), or load it as ONNX.\",\n        );\n    }\n    return requested;\n}\n\n/**\n * Find the file a runtime reads.\n *\n * @param manifest The package manifest.\n * @param runtime The chosen runtime.\n * @returns The filename inside the package directory.\n */\nfunction fileFor(manifest: EdgeManifest, runtime: TabularRuntime): string {\n    const entry = (manifest.runtimes ?? []).find((item) => item.kind === runtime);\n    return entry?.file ?? manifest.model.file;\n}\n\n/**\n * Resolve a directory URL to its manifest file.\n *\n * @param directoryUrl The package directory, or the manifest itself.\n * @returns The manifest URL.\n */\nfunction manifestUrl(directoryUrl: string): string {\n    if (directoryUrl.endsWith(\".json\")) return directoryUrl;\n    return directoryUrl.endsWith(\"/\")\n        ? `${directoryUrl}${MANIFEST_FILENAME}`\n        : `${directoryUrl}/${MANIFEST_FILENAME}`;\n}\n"],"mappings":"A,qHA4BA,IAAaQ,EAAoB,gBA8JjC,eAAsBC,EAClB,EACA,EACqB,CACrB,IAAMC,EAAMC,EAAYC,CAAY,EAChCC,EACJ,GAAI,CACA,EAAW,MAAM,MAAMH,EAAKI,CAAW,CAC3C,OAASC,EAAO,CACZ,MAAM,IAAIf,EAAAA,gBAAgB,kCAAkCU,IAAO,CAC/D,MAAOK,CACX,CAAC,CACL,CACA,GAAI,CAACF,EAAS,GACV,MAAM,IAAIb,EAAAA,gBACN,kCAAkCU,EAAI,IAAIG,EAAS,OAAO,GAAGA,EAAS,YAC1E,EAGJ,IAAMG,EAAY,MAAMH,EAAS,KAAK,EACtC,GAAI,OAAOG,GAAU,gBAAmB,SACpC,MAAM,IAAIhB,EAAAA,gBAAgB,GAAGU,EAAI,sDAAsD,EAE3F,GAAIM,EAAS,eAAA,EACT,MAAM,IAAIhB,EAAAA,gBACN,mBAAmBU,EAAI,uBAAuBM,EAAS,eAAe,qFAG1E,EAEJ,OAAOA,CACX,CAoBA,eAAsBC,EAClB,EACA,EAAkC,CAAC,EACT,CAC1B,IAAMD,EAAW,MAAMP,EAAkBG,CAAY,EAC/CM,EAAON,EAAa,SAAS,GAAG,EAAIA,EAAe,GAAGA,EAAa,GACnEO,EAAUC,EAAcJ,EAAUK,EAAQ,SAAW,MAAM,EAE3DC,EAAW,GAAGJ,IADPK,EAAQP,EAAUG,CACJ,IAErBK,EAAQH,EAAQ,OAAS,GACzBI,EACFD,IAAU,GACJF,EACA,MAAMrB,EAAAA,gBAAgBqB,EAAU,OAAOE,GAAU,SAAWA,EAAQ,CAAC,CAAC,EAE1EE,EACFP,IAAY,UACN,MAAMjB,EAAAA,iBAAiB,OAAOuB,CAAM,EACpC,MAAMtB,EAAAA,iBAAiB,OAAOsB,EAAQ,CAClC,UAAWJ,EAAQ,UACnB,OAAQA,EAAQ,OAChB,eAAgBA,EAAQ,cAC5B,CAAC,EAELM,EAAUX,EAAS,OAAO,QAChC,MAAO,CACH,WACA,YACA,UACA,aAAcA,EAAS,MAAM,cAC7B,UACA,QAAU,GACNY,EACK,KAAK,EAAO,KAAW,CACpB,KAAMD,EAAQE,IAAU,OAAOA,CAAK,EACpC,OACJ,EAAE,CAAC,CACF,MAAM,EAAG,IAAMC,EAAE,MAAQC,EAAE,KAAK,CAC7C,CACJ,CAYA,SAASX,EAAc,EAAwB,EAAoD,CAC/F,IAAMY,EAAYhB,EAAS,UAAY,CAAC,EAClCiB,EAAaD,EAAU,KAAM,GAAUE,EAAM,OAAS,SAAS,EAErE,GAAIC,IAAc,OAAQ,OAAOF,EAAa,UAAY,OAC1D,GAAIE,IAAc,WAAa,CAACF,EAC5B,MAAM,IAAIjC,EAAAA,gBACN,0CAA0CgB,EAAS,KAAK,SACjDgB,EAAU,IAAK,GAAUE,EAAM,IAAI,CAAC,CAAC,KAAK,IAAI,GAAK,OAAO,sEAErE,EAEJ,OAAOC,CACX,CASA,SAASZ,EAAQ,EAAwB,EAAiC,CAEtE,OADeP,EAAS,UAAY,CAAC,EAAA,CAAG,KAAM,GAASoB,EAAK,OAASjB,CAC9D,CAAA,EAAO,MAAQH,EAAS,MAAM,IACzC,CAQA,SAASL,EAAY,EAA8B,CAE/C,OADIC,EAAa,SAAS,OAAO,EAAUA,EACpCA,EAAa,SAAS,GAAG,EAC1B,GAAGA,IAAeJ,IAClB,GAAGI,EAAa,GAAGJ,GAC7B,C"}
+{"version":3,"file":"manifest.cjs","names":["ModelFetchError","CompactPredictor","TabularPredictor","fetchModelBytes","require_exceptions","require_cache","require_compact","require_predictor","MANIFEST_FILENAME","fetchEdgeManifest","url","manifestUrl","directoryUrl","response","requestInit","error","manifest","loadEdgePackage","base","runtime","chooseRuntime","options","fileFor","modelUrl","file","cache","source","verifiedModelBytes","sha256","predictor","classes","probabilities","index","b","a","available","hasCompact","entry","requested","item","digestMatches","subtle","expected","digest","bytes","byte","fresh"],"sources":["../../src/tabular/manifest.ts"],"sourcesContent":["/**\n * Reading an edge package published by `tempest-fastapi-sdk`.\n *\n * `edge_pipeline` writes a directory — the graph, a gzipped copy, a drift\n * baseline and a `manifest.json` — and the browser is one of its two\n * intended consumers. The manifest is what turns a `.onnx` URL into\n * something a UI can use: the **column order** the model was trained on,\n * the class names behind `probabilities[2]`, and a version to compare\n * against what is already cached.\n *\n * That column order is the field worth loading a manifest for. A model fed\n * the right features in the wrong order answers confidently and wrongly,\n * and no runtime check catches it.\n *\n * The contract is pinned by `schema_version`. Unknown fields are ignored\n * rather than rejected, so a package written by a newer SDK still loads.\n */\n\nimport { fetchModelBytes, type ModelCacheOptions } from \"./cache\";\nimport { CompactPredictor } from \"./compact\";\nimport { ModelFetchError } from \"./exceptions\";\nimport { TabularPredictor } from \"./predictor\";\nimport type { FeatureRow, TabularPrediction, TabularPredictorOptions } from \"./types\";\n\n/** Manifest schema version this reader was written against. */\nexport const SUPPORTED_MANIFEST_SCHEMA = 1;\n\n/** Fixed filename inside a package directory. */\nexport const MANIFEST_FILENAME = \"manifest.json\";\n\n/** The graph file and how to check you got it whole. */\nexport interface ManifestModelFile {\n    readonly file: string;\n    readonly sha256: string;\n    readonly bytes: number;\n    readonly gzip_file: string | null;\n    readonly gzip_bytes: number | null;\n    readonly opset: number;\n    readonly dtype: string;\n}\n\n/** What the graph expects per row. */\nexport interface ManifestInput {\n    readonly name: string;\n    readonly features: number;\n    /** Column order used at training time. */\n    readonly feature_names: readonly string[];\n}\n\n/** What the graph answers. */\nexport interface ManifestOutput {\n    readonly is_classifier: boolean;\n    readonly label_output: string;\n    readonly probability_output: string | null;\n    /** Class labels in score-column order. */\n    readonly classes: readonly string[];\n}\n\n/**\n * Where the packaged model came from, when it came from an existing\n * artifact.\n *\n * Present when the package was built with `edge_pipeline_from_pickle`: the\n * `.pkl` never reaches the browser (a pickle is a Python program, not\n * data), but its name and digest travel in the manifest, so a model\n * answering in a tab can be traced back to the file that produced it.\n */\nexport interface ManifestSource {\n    readonly file: string;\n    readonly kind: string;\n    readonly sha256: string;\n    readonly bytes: number;\n    readonly sklearn_version: string;\n    readonly warnings: readonly string[];\n}\n\n/**\n * One file in the package a runtime can load.\n *\n * A package may carry the same model twice — as ONNX, which any runtime\n * reads at the cost of a 25.6 MB WebAssembly download, and as the compact\n * format, which needs no runtime. The list is what lets the browser pick\n * by what it already ships.\n */\nexport interface ManifestRuntime {\n    readonly kind: \"onnx\" | \"compact\" | string;\n    readonly file: string;\n    readonly bytes: number;\n    readonly gzip_file: string | null;\n    readonly gzip_bytes: number | null;\n    readonly sha256: string;\n}\n\n/** The package manifest, as written by `edge_pipeline`. */\nexport interface EdgeManifest {\n    readonly schema_version: number;\n    readonly name: string;\n    readonly version: string;\n    readonly created_at: string;\n    readonly sdk_version: string;\n    readonly estimator: string;\n    readonly model: ManifestModelFile;\n    readonly input: ManifestInput;\n    readonly output: ManifestOutput;\n    readonly verified: boolean | null;\n    /** Every file a runtime can load. Absent on packages written before v0.194. */\n    readonly runtimes?: readonly ManifestRuntime[];\n    /** Absent on packages built straight from a fitted estimator. */\n    readonly source?: ManifestSource;\n    readonly baseline_file: string | null;\n    readonly baseline_samples: number;\n}\n\n/** Which reader served the package. */\nexport type TabularRuntime = \"onnx\" | \"compact\";\n\n/** A package loaded and ready to answer. */\nexport interface LoadedEdgePackage {\n    /** What was published. */\n    readonly manifest: EdgeManifest;\n    /** The running model, whichever runtime read it. */\n    readonly predictor: PredictorLike;\n    /** Which reader was used. */\n    readonly runtime: TabularRuntime;\n    /** Column order the rows must follow. */\n    readonly featureNames: readonly string[];\n    /** Class names behind each probability column. */\n    readonly classes: readonly string[];\n    /**\n     * Map a prediction's scores onto class names.\n     *\n     * @param probabilities One row of scores.\n     * @returns Name/score pairs, highest first.\n     */\n    readonly explain: (probabilities: readonly number[]) => { name: string; score: number }[];\n}\n\n/**\n * The shape both readers share.\n *\n * `TabularPredictor` (ONNX) and `CompactPredictor` (runtime-free) answer\n * with the same object, so an app can switch routes without touching a\n * line of its own code.\n */\nexport interface PredictorLike {\n    predict(rows: readonly FeatureRow[]): Promise<TabularPrediction>;\n    dispose(): Promise<void>;\n}\n\n/** Options for {@link loadEdgePackage}. */\nexport interface LoadEdgePackageOptions extends TabularPredictorOptions {\n    /** Cache the model bytes for offline use. `true` by default. */\n    readonly cache?: boolean | ModelCacheOptions;\n    /**\n     * Which reader to use.\n     *\n     * `\"auto\"` (the default) takes the compact form when the package has\n     * one, because it answers without downloading a WebAssembly runtime.\n     * Force `\"onnx\"` when the app already ships ONNX for something else —\n     * then the runtime is already paid for and ONNX covers more estimators.\n     */\n    readonly runtime?: TabularRuntime | \"auto\";\n}\n\n/**\n * Read a package's manifest.\n *\n * Cheap: it is a few hundred bytes, so an app can check for a new version\n * without downloading a model it may already have.\n *\n * @example\n * ```ts\n * const manifest = await fetchEdgeManifest(\"/models/risk/\");\n * if (manifest.version !== localStorage.getItem(\"risk-version\")) {\n *     // a new model was published\n * }\n * ```\n *\n * @param directoryUrl URL of the package directory, with or without a\n *   trailing slash. A full URL to the manifest file also works.\n * @param requestInit `fetch` options.\n * @returns The parsed manifest.\n * @throws {@link ModelFetchError} when the manifest cannot be read, or when\n *   its `schema_version` is newer than this reader understands — loading it\n *   anyway would risk misreading the field that defines column order.\n */\nexport async function fetchEdgeManifest(\n    directoryUrl: string,\n    requestInit?: RequestInit,\n): Promise<EdgeManifest> {\n    const url = manifestUrl(directoryUrl);\n    let response: Response;\n    try {\n        response = await fetch(url, requestInit);\n    } catch (error) {\n        throw new ModelFetchError(`Could not read the manifest at ${url}`, {\n            cause: error,\n        });\n    }\n    if (!response.ok) {\n        throw new ModelFetchError(\n            `Could not read the manifest at ${url}: ${response.status} ${response.statusText}`,\n        );\n    }\n\n    const manifest = (await response.json()) as EdgeManifest;\n    if (typeof manifest?.schema_version !== \"number\") {\n        throw new ModelFetchError(`${url} is not an edge package manifest (no schema_version).`);\n    }\n    if (manifest.schema_version > SUPPORTED_MANIFEST_SCHEMA) {\n        throw new ModelFetchError(\n            `The manifest at ${url} uses schema_version ${manifest.schema_version}, ` +\n                `and this SDK understands ${SUPPORTED_MANIFEST_SCHEMA}. Upgrade ` +\n                \"tempest-react-sdk before serving this package.\",\n        );\n    }\n    return manifest;\n}\n\n/**\n * Load a whole edge package: manifest, model, and the names to read it by.\n *\n * @example\n * ```ts\n * const pkg = await loadEdgePackage(\"/models/risk/\");\n *\n * console.log(pkg.featureNames); // [\"age\", \"income\", \"tenure\", \"score\", \"visits\"]\n *\n * const { probabilities } = await pkg.predictor.predict([[41, 5200, 3, 0.82, 12]]);\n * console.log(pkg.explain(probabilities[0]!)); // [{ name: \"approved\", score: 0.91 }, ...]\n * ```\n *\n * The model bytes are checked against the `sha256` the manifest lists for\n * them before any runtime sees them. The cache is keyed by URL and read\n * cache-first, so a model republished in place under the same file name kept\n * answering from the old bytes next to the new manifest — new column order,\n * old weights, no error. A mismatch now refetches once from the network, and a\n * second mismatch throws. Two cases are not checked: `cache: false`, where the\n * runtime fetches the URL itself and the bytes never pass through here, and a\n * non-secure context, which has no `crypto.subtle` (nor Cache Storage).\n *\n * @param directoryUrl URL of the package directory.\n * @param options Predictor options plus caching.\n * @returns The loaded package.\n * @throws {@link ModelFetchError} when the manifest or model cannot be read, or\n *   when the model does not match the manifest's `sha256` even after a refetch.\n */\nexport async function loadEdgePackage(\n    directoryUrl: string,\n    options: LoadEdgePackageOptions = {},\n): Promise<LoadedEdgePackage> {\n    const manifest = await fetchEdgeManifest(directoryUrl);\n    const base = directoryUrl.endsWith(\"/\") ? directoryUrl : `${directoryUrl}/`;\n    const runtime = chooseRuntime(manifest, options.runtime ?? \"auto\");\n    const { file, sha256 } = fileFor(manifest, runtime);\n    const modelUrl = `${base}${file}`;\n\n    const cache = options.cache ?? true;\n    const source =\n        cache === false\n            ? modelUrl\n            : await verifiedModelBytes(modelUrl, sha256, typeof cache === \"object\" ? cache : {});\n\n    const predictor: PredictorLike =\n        runtime === \"compact\"\n            ? await CompactPredictor.create(source)\n            : await TabularPredictor.create(source, {\n                  providers: options.providers,\n                  warmup: options.warmup,\n                  sessionOptions: options.sessionOptions,\n              });\n\n    const classes = manifest.output.classes;\n    return {\n        manifest,\n        predictor,\n        runtime,\n        featureNames: manifest.input.feature_names,\n        classes,\n        explain: (probabilities: readonly number[]) =>\n            probabilities\n                .map((score, index) => ({\n                    name: classes[index] ?? String(index),\n                    score,\n                }))\n                .sort((a, b) => b.score - a.score),\n    };\n}\n\n/**\n * Decide which reader serves this package.\n *\n * @param manifest The package manifest.\n * @param requested What the caller asked for.\n * @returns The runtime to use.\n * @throws {@link ModelFetchError} when the package does not carry the\n *   requested form — asking for a compact model that was never written\n *   should say so, not silently download 25 MB of WebAssembly instead.\n */\nfunction chooseRuntime(manifest: EdgeManifest, requested: TabularRuntime | \"auto\"): TabularRuntime {\n    const available = manifest.runtimes ?? [];\n    const hasCompact = available.some((entry) => entry.kind === \"compact\");\n\n    if (requested === \"auto\") return hasCompact ? \"compact\" : \"onnx\";\n    if (requested === \"compact\" && !hasCompact) {\n        throw new ModelFetchError(\n            `This package carries no compact model (${manifest.name} lists ` +\n                `${available.map((entry) => entry.kind).join(\", \") || \"onnx\"}). ` +\n                \"Re-export it with edge_pipeline(compact=True), or load it as ONNX.\",\n        );\n    }\n    return requested;\n}\n\n/**\n * Find the file a runtime reads, with the digest the manifest records for it.\n *\n * @param manifest The package manifest.\n * @param runtime The chosen runtime.\n * @returns The filename inside the package directory and its `sha256`.\n */\nfunction fileFor(\n    manifest: EdgeManifest,\n    runtime: TabularRuntime,\n): { file: string; sha256: string } {\n    const entry = (manifest.runtimes ?? []).find((item) => item.kind === runtime);\n    return entry ?? manifest.model;\n}\n\n/**\n * Whether `bytes` hash to `expected`.\n *\n * Answers `true` when nothing can be compared — no digest in the manifest, or\n * no `crypto.subtle` outside a secure context — because refusing to load there\n * would take inference away from an app over a check it never had.\n *\n * @param bytes The model bytes.\n * @param expected Hex SHA-256 from the manifest.\n * @returns `false` only on a real mismatch.\n */\nasync function digestMatches(bytes: Uint8Array, expected: string): Promise<boolean> {\n    const subtle = globalThis.crypto?.subtle;\n    if (!subtle || !expected) return true;\n    const digest = new Uint8Array(await subtle.digest(\"SHA-256\", bytes as Uint8Array<ArrayBuffer>));\n    const hex = Array.from(digest, (byte) => byte.toString(16).padStart(2, \"0\")).join(\"\");\n    return hex === expected.toLowerCase();\n}\n\n/**\n * Read the model bytes and prove they are the ones the manifest describes.\n *\n * A cached copy that does not match is refetched from the network once (the\n * refetch also replaces the cache entry); only a second mismatch throws.\n *\n * @param url The model URL.\n * @param expected Hex SHA-256 from the manifest.\n * @param options Cache options as the caller gave them.\n * @returns The verified bytes.\n * @throws {@link ModelFetchError} when the bytes cannot be read or never match.\n */\nasync function verifiedModelBytes(\n    url: string,\n    expected: string,\n    options: ModelCacheOptions,\n): Promise<Uint8Array> {\n    const bytes = await fetchModelBytes(url, options);\n    if (await digestMatches(bytes, expected)) return bytes;\n    if (options.revalidate !== true) {\n        const fresh = await fetchModelBytes(url, { ...options, revalidate: true });\n        if (await digestMatches(fresh, expected)) return fresh;\n    }\n    throw new ModelFetchError(\n        `The model at ${url} does not match the sha256 in its manifest. The file and ` +\n            \"the manifest were published separately, or the download was altered — \" +\n            \"republish the package so both describe the same model.\",\n    );\n}\n\n/**\n * Resolve a directory URL to its manifest file.\n *\n * @param directoryUrl The package directory, or the manifest itself.\n * @returns The manifest URL.\n */\nfunction manifestUrl(directoryUrl: string): string {\n    if (directoryUrl.endsWith(\".json\")) return directoryUrl;\n    return directoryUrl.endsWith(\"/\")\n        ? `${directoryUrl}${MANIFEST_FILENAME}`\n        : `${directoryUrl}/${MANIFEST_FILENAME}`;\n}\n"],"mappings":"A,qHA4BA,IAAaQ,EAAoB,gBA8JjC,eAAsBC,EAClB,EACA,EACqB,CACrB,IAAMC,EAAMC,EAAYC,CAAY,EAChCC,EACJ,GAAI,CACA,EAAW,MAAM,MAAMH,EAAKI,CAAW,CAC3C,OAASC,EAAO,CACZ,MAAM,IAAIf,EAAAA,gBAAgB,kCAAkCU,IAAO,CAC/D,MAAOK,CACX,CAAC,CACL,CACA,GAAI,CAACF,EAAS,GACV,MAAM,IAAIb,EAAAA,gBACN,kCAAkCU,EAAI,IAAIG,EAAS,OAAO,GAAGA,EAAS,YAC1E,EAGJ,IAAMG,EAAY,MAAMH,EAAS,KAAK,EACtC,GAAI,OAAOG,GAAU,gBAAmB,SACpC,MAAM,IAAIhB,EAAAA,gBAAgB,GAAGU,EAAI,sDAAsD,EAE3F,GAAIM,EAAS,eAAA,EACT,MAAM,IAAIhB,EAAAA,gBACN,mBAAmBU,EAAI,uBAAuBM,EAAS,eAAe,qFAG1E,EAEJ,OAAOA,CACX,CA8BA,eAAsBC,EAClB,EACA,EAAkC,CAAC,EACT,CAC1B,IAAMD,EAAW,MAAMP,EAAkBG,CAAY,EAC/CM,EAAON,EAAa,SAAS,GAAG,EAAIA,EAAe,GAAGA,EAAa,GACnEO,EAAUC,EAAcJ,EAAUK,EAAQ,SAAW,MAAM,EAC3D,CAAE,OAAM,UAAWC,EAAQN,EAAUG,CAAO,EAC5CI,EAAW,GAAGL,IAAOM,IAErBC,EAAQJ,EAAQ,OAAS,GACzBK,EACFD,IAAU,GACJF,EACA,MAAMI,EAAmBJ,EAAUK,EAAQ,OAAOH,GAAU,SAAWA,EAAQ,CAAC,CAAC,EAErFI,EACFV,IAAY,UACN,MAAMlB,EAAAA,iBAAiB,OAAOyB,CAAM,EACpC,MAAMxB,EAAAA,iBAAiB,OAAOwB,EAAQ,CAClC,UAAWL,EAAQ,UACnB,OAAQA,EAAQ,OAChB,eAAgBA,EAAQ,cAC5B,CAAC,EAELS,EAAUd,EAAS,OAAO,QAChC,MAAO,CACH,WACA,YACA,UACA,aAAcA,EAAS,MAAM,cAC7B,UACA,QAAU,GACNe,EACK,KAAK,EAAO,KAAW,CACpB,KAAMD,EAAQE,IAAU,OAAOA,CAAK,EACpC,OACJ,EAAE,CAAC,CACF,MAAM,EAAG,IAAMC,EAAE,MAAQC,EAAE,KAAK,CAC7C,CACJ,CAYA,SAASd,EAAc,EAAwB,EAAoD,CAC/F,IAAMe,EAAYnB,EAAS,UAAY,CAAC,EAClCoB,EAAaD,EAAU,KAAM,GAAUE,EAAM,OAAS,SAAS,EAErE,GAAIC,IAAc,OAAQ,OAAOF,EAAa,UAAY,OAC1D,GAAIE,IAAc,WAAa,CAACF,EAC5B,MAAM,IAAIpC,EAAAA,gBACN,0CAA0CgB,EAAS,KAAK,SACjDmB,EAAU,IAAK,GAAUE,EAAM,IAAI,CAAC,CAAC,KAAK,IAAI,GAAK,OAAO,sEAErE,EAEJ,OAAOC,CACX,CASA,SAAShB,EACL,EACA,EACgC,CAEhC,OADeN,EAAS,UAAY,CAAC,EAAA,CAAG,KAAM,GAASuB,EAAK,OAASpB,CAC9D,GAASH,EAAS,KAC7B,CAaA,eAAewB,EAAc,EAAmB,EAAoC,CAChF,IAAMC,EAAS,WAAW,QAAQ,OAClC,GAAI,CAACA,GAAU,CAACC,EAAU,MAAO,GACjC,IAAMC,EAAS,IAAI,WAAW,MAAMF,EAAO,OAAO,UAAWG,CAAgC,CAAC,EAE9F,OADY,MAAM,KAAKD,EAAS,GAASE,EAAK,SAAS,EAAE,CAAC,CAAC,SAAS,EAAG,GAAG,CAAC,CAAC,CAAC,KAAK,EAC3E,IAAQH,EAAS,YAAY,CACxC,CAcA,eAAef,EACX,EACA,EACA,EACmB,CACnB,IAAMiB,EAAQ,MAAMzC,EAAAA,gBAAgBO,EAAKW,CAAO,EAChD,GAAI,MAAMmB,EAAcI,EAAOF,CAAQ,EAAG,OAAOE,EACjD,GAAIvB,EAAQ,aAAe,GAAM,CAC7B,IAAMyB,EAAQ,MAAM3C,EAAAA,gBAAgBO,EAAK,CAAE,GAAGW,EAAS,WAAY,EAAK,CAAC,EACzE,GAAI,MAAMmB,EAAcM,EAAOJ,CAAQ,EAAG,OAAOI,CACrD,CACA,MAAM,IAAI9C,EAAAA,gBACN,gBAAgBU,EAAI,sLAGxB,CACJ,CAQA,SAASC,EAAY,EAA8B,CAE/C,OADIC,EAAa,SAAS,OAAO,EAAUA,EACpCA,EAAa,SAAS,GAAG,EAC1B,GAAGA,IAAeJ,IAClB,GAAGI,EAAa,GAAGJ,GAC7B,C"}