agentdb 3.0.0-alpha.10 → 3.0.0-alpha.12

package/dist/src/backends/rvf/ContrastiveTrainer.d.ts

@@ -0,0 +1,148 @@
+/**
+ * ContrastiveTrainer - Contrastive Embedding Improvement for AgentDB
+ *
+ * Provides:
+ * - InfoNCE contrastive loss for embedding refinement
+ * - Hard negative mining for effective training samples
+ * - Curriculum scheduling for progressive difficulty
+ * - AdamW optimizer for weight updates (via @ruvector/attention when available)
+ *
+ * Security:
+ * - Operates on embeddings only (no user text)
+ * - Temperature bounded (0.01-1.0)
+ * - Batch sizes bounded to prevent OOM
+ * - All inputs dimension-validated
+ */
+/** Training sample for contrastive learning */
+export interface ContrastiveSample {
+    anchor: Float32Array;
+    positive: Float32Array;
+    negatives: Float32Array[];
+}
+/** Training batch result */
+export interface TrainBatchResult {
+    loss: number;
+    batchSize: number;
+    avgGradNorm: number;
+}
+/** Training session statistics */
+export interface TrainingStats {
+    totalBatches: number;
+    totalSamples: number;
+    avgLoss: number;
+    bestLoss: number;
+    currentDifficulty: number;
+    epoch: number;
+}
+/** Curriculum stage configuration */
+export interface CurriculumStage {
+    /** Number of negatives per sample in this stage */
+    negativeCount: number;
+    /** Minimum similarity for hard negatives */
+    hardNegativeThreshold: number;
+    /** Number of batches in this stage */
+    batches: number;
+}
+/** ContrastiveTrainer configuration */
+export interface ContrastiveConfig {
+    /** Embedding dimension */
+    dimension: number;
+    /** InfoNCE temperature (default: 0.07) */
+    temperature?: number;
+    /** Learning rate (default: 0.001) */
+    learningRate?: number;
+    /** Weight decay for AdamW (default: 0.01) */
+    weightDecay?: number;
+    /** Maximum batch size (default: 64) */
+    maxBatchSize?: number;
+    /** Curriculum stages (default: 3 stages) */
+    stages?: CurriculumStage[];
+}
+import type { NativeAccelerator } from './NativeAccelerator.js';
+/**
+ * ContrastiveTrainer - Self-improving embedding quality via contrastive learning
+ *
+ * Uses InfoNCE loss to train a lightweight projection that improves
+ * embedding quality over time. Hard negative mining ensures the model
+ * focuses on the most informative training signals.
+ *
+ * Automatically uses NativeAccelerator (ADR-007) for SIMD-accelerated
+ * cosine similarity and native AdamW when available.
+ */
+export declare class ContrastiveTrainer {
+    private dim;
+    private temperature;
+    private learningRate;
+    private weightDecay;
+    private maxBatchSize;
+    private stages;
+    private currentStage;
+    private batchesInStage;
+    private weights;
+    private bias;
+    private mW;
+    private vW;
+    private mB;
+    private vB;
+    private step;
+    private accel;
+    private _totalBatches;
+    private _totalSamples;
+    private _lossSum;
+    private _bestLoss;
+    private _destroyed;
+    private constructor();
+    /**
+     * Create a new contrastive trainer.
+     */
+    static create(config: ContrastiveConfig): Promise<ContrastiveTrainer>;
+    /**
+     * Project an embedding through the learned transformation.
+     * Use this to enhance query/document embeddings after training.
+     */
+    project(embedding: Float32Array): Float32Array;
+    /**
+     * Compute InfoNCE loss for a batch of contrastive samples.
+     *
+     * L = -log(exp(sim(a, p) / t) / sum(exp(sim(a, n_i) / t)))
+     */
+    computeLoss(samples: ContrastiveSample[]): number;
+    /**
+     * Train on a batch of contrastive samples.
+     * Uses analytical backpropagation through the InfoNCE loss.
+     */
+    trainBatch(samples: ContrastiveSample[]): TrainBatchResult;
+    /**
+     * Mine hard negatives from a pool of embeddings.
+     * Returns negatives that are similar to anchor but from different classes.
+     *
+     * SOTA: Positive-aware filtering (NV-Retriever, 2024). Rejects candidates
+     * with high similarity to known positives to eliminate false negatives,
+     * which constitute ~70% of naively mined hard negatives.
+     */
+    mineHardNegatives(anchor: Float32Array, pool: Float32Array[], excludeIndices: Set<number>, count?: number, positives?: Float32Array[]): Float32Array[];
+    /**
+     * Get the NativeAccelerator instance (for testing/introspection).
+     */
+    getAccelerator(): NativeAccelerator | null;
+    /**
+     * Get the current curriculum stage info.
+     */
+    get currentCurriculumStage(): number;
+    /**
+     * Get training statistics.
+     */
+    getStats(): TrainingStats;
+    /** Check if destroyed */
+    get isDestroyed(): boolean;
+    /** Get configured dimension */
+    get dimension(): number;
+    /** Destroy and free resources */
+    destroy(): void;
+    private matVecMul;
+    private cosineSimilarity;
+    private vecNorm;
+    private adamWUpdate;
+    private ensureAlive;
+}
+//# sourceMappingURL=ContrastiveTrainer.d.ts.map

package/dist/src/backends/rvf/ContrastiveTrainer.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"ContrastiveTrainer.d.ts","sourceRoot":"","sources":["../../../../src/backends/rvf/ContrastiveTrainer.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAEH,+CAA+C;AAC/C,MAAM,WAAW,iBAAiB;IAChC,MAAM,EAAE,YAAY,CAAC;IACrB,QAAQ,EAAE,YAAY,CAAC;IACvB,SAAS,EAAE,YAAY,EAAE,CAAC;CAC3B;AAED,4BAA4B;AAC5B,MAAM,WAAW,gBAAgB;IAC/B,IAAI,EAAE,MAAM,CAAC;IACb,SAAS,EAAE,MAAM,CAAC;IAClB,WAAW,EAAE,MAAM,CAAC;CACrB;AAED,kCAAkC;AAClC,MAAM,WAAW,aAAa;IAC5B,YAAY,EAAE,MAAM,CAAC;IACrB,YAAY,EAAE,MAAM,CAAC;IACrB,OAAO,EAAE,MAAM,CAAC;IAChB,QAAQ,EAAE,MAAM,CAAC;IACjB,iBAAiB,EAAE,MAAM,CAAC;IAC1B,KAAK,EAAE,MAAM,CAAC;CACf;AAED,qCAAqC;AACrC,MAAM,WAAW,eAAe;IAC9B,mDAAmD;IACnD,aAAa,EAAE,MAAM,CAAC;IACtB,4CAA4C;IAC5C,qBAAqB,EAAE,MAAM,CAAC;IAC9B,sCAAsC;IACtC,OAAO,EAAE,MAAM,CAAC;CACjB;AAED,uCAAuC;AACvC,MAAM,WAAW,iBAAiB;IAChC,0BAA0B;IAC1B,SAAS,EAAE,MAAM,CAAC;IAClB,0CAA0C;IAC1C,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,qCAAqC;IACrC,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,6CAA6C;IAC7C,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,uCAAuC;IACvC,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,4CAA4C;IAC5C,MAAM,CAAC,EAAE,eAAe,EAAE,CAAC;CAC5B;AAED,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,wBAAwB,CAAC;AAShE;;;;;;;;;GASG;AACH,qBAAa,kBAAkB;IAC7B,OAAO,CAAC,GAAG,CAAS;IACpB,OAAO,CAAC,WAAW,CAAS;IAC5B,OAAO,CAAC,YAAY,CAAS;IAC7B,OAAO,CAAC,WAAW,CAAS;IAC5B,OAAO,CAAC,YAAY,CAAS;IAC7B,OAAO,CAAC,MAAM,CAAoB;IAClC,OAAO,CAAC,YAAY,CAAK;IACzB,OAAO,CAAC,cAAc,CAAK;IAG3B,OAAO,CAAC,OAAO,CAAe;IAC9B,OAAO,CAAC,IAAI,CAAe;IAG3B,OAAO,CAAC,EAAE,CAAe;IACzB,OAAO,CAAC,EAAE,CAAe;IACzB,OAAO,CAAC,EAAE,CAAe;IACzB,OAAO,CAAC,EAAE,CAAe;IACzB,OAAO,CAAC,IAAI,CAAK;IAGjB,OAAO,CAAC,KAAK,CAAkC;IAG/C,OAAO,CAAC,aAAa,CAAK;IAC1B,OAAO,CAAC,aAAa,CAAK;IAC1B,OAAO,CAAC,QAAQ,CAAK;IACrB,OAAO,CAAC,SAAS,CAAY;IAC7B,OAAO,CAAC,UAAU,CAAS;IAE3B,OAAO;IAqCP;;OAEG;WACU,MAAM,CAAC,MAAM,EAAE,iBAAiB,GAAG,OAAO,CAAC,kBAAkB,CAAC;IAc3E;;;OAGG;IACH,OAAO,CAAC,SAAS,EAAE,YAAY,GAAG,YAAY;IAQ9C;;;;OAIG;IACH,WAAW,CAAC,OAAO,EAAE,iBAAiB,EAAE,GAAG,MAAM;IA+BjD;;;OAGG;IACH,UAAU,CAAC,OAAO,EAAE,iBAAiB,EAAE,GAAG,gBAAgB;IAsK1D;;;;;;;OAOG;IACH,iBAAiB,CACf,MAAM,EAAE,YAAY,EACpB,IAAI,EAAE,YAAY,EAAE,EACpB,cAAc,EAAE,GAAG,CAAC,MAAM,CAAC,EAC3B,KAAK,CAAC,EAAE,MAAM,EACd,SAAS,CAAC,EAAE,YAAY,EAAE,GACzB,YAAY,EAAE;IAmCjB;;OAEG;IACH,cAAc,IAAI,iBAAiB,GAAG,IAAI;IAI1C;;OAEG;IACH,IAAI,sBAAsB,IAAI,MAAM,CAEnC;IAED;;OAEG;IACH,QAAQ,IAAI,aAAa;IAWzB,yBAAyB;IACzB,IAAI,WAAW,IAAI,OAAO,CAEzB;IAED,+BAA+B;IAC/B,IAAI,SAAS,IAAI,MAAM,CAEtB;IAED,iCAAiC;IACjC,OAAO,IAAI,IAAI;IAQf,OAAO,CAAC,SAAS;IAcjB,OAAO,CAAC,gBAAgB;IAaxB,OAAO,CAAC,OAAO;IAMf,OAAO,CAAC,WAAW;IAoCnB,OAAO,CAAC,WAAW;CAKpB"}

package/dist/src/backends/rvf/ContrastiveTrainer.js

@@ -0,0 +1,438 @@
+/**
+ * ContrastiveTrainer - Contrastive Embedding Improvement for AgentDB
+ *
+ * Provides:
+ * - InfoNCE contrastive loss for embedding refinement
+ * - Hard negative mining for effective training samples
+ * - Curriculum scheduling for progressive difficulty
+ * - AdamW optimizer for weight updates (via @ruvector/attention when available)
+ *
+ * Security:
+ * - Operates on embeddings only (no user text)
+ * - Temperature bounded (0.01-1.0)
+ * - Batch sizes bounded to prevent OOM
+ * - All inputs dimension-validated
+ */
+// Bounds
+const MIN_TEMPERATURE = 0.01;
+const MAX_TEMPERATURE = 1.0;
+const MAX_BATCH_SIZE = 256;
+const MAX_NEGATIVES = 128;
+const MAX_DIMENSION = 4096;
+/**
+ * ContrastiveTrainer - Self-improving embedding quality via contrastive learning
+ *
+ * Uses InfoNCE loss to train a lightweight projection that improves
+ * embedding quality over time. Hard negative mining ensures the model
+ * focuses on the most informative training signals.
+ *
+ * Automatically uses NativeAccelerator (ADR-007) for SIMD-accelerated
+ * cosine similarity and native AdamW when available.
+ */
+export class ContrastiveTrainer {
+    dim;
+    temperature;
+    learningRate;
+    weightDecay;
+    maxBatchSize;
+    stages;
+    currentStage = 0;
+    batchesInStage = 0;
+    // Lightweight linear projection: W * x + b
+    weights;
+    bias;
+    // AdamW state
+    mW;
+    vW;
+    mB;
+    vB;
+    step = 0;
+    // Native accelerator (ADR-007) — optional SIMD + native ops
+    accel = null;
+    // Stats
+    _totalBatches = 0;
+    _totalSamples = 0;
+    _lossSum = 0;
+    _bestLoss = Infinity;
+    _destroyed = false;
+    constructor(config) {
+        this.dim = config.dimension;
+        this.temperature = Math.min(Math.max(MIN_TEMPERATURE, config.temperature ?? 0.07), MAX_TEMPERATURE);
+        this.learningRate = Math.max(1e-6, config.learningRate ?? 0.001);
+        this.weightDecay = Math.max(0, config.weightDecay ?? 0.01);
+        this.maxBatchSize = Math.min(Math.max(1, config.maxBatchSize ?? 64), MAX_BATCH_SIZE);
+        this.stages = config.stages ?? [
+            { negativeCount: 4, hardNegativeThreshold: 0.5, batches: 100 },
+            { negativeCount: 8, hardNegativeThreshold: 0.3, batches: 100 },
+            { negativeCount: 16, hardNegativeThreshold: 0.1, batches: 100 },
+        ];
+        // Initialize projection as identity + small noise
+        const d = this.dim;
+        this.weights = new Float32Array(d * d);
+        this.bias = new Float32Array(d);
+        for (let i = 0; i < d; i++) {
+            this.weights[i * d + i] = 1.0; // Identity diagonal
+        }
+        // Small random perturbation
+        for (let i = 0; i < this.weights.length; i++) {
+            this.weights[i] += (Math.random() - 0.5) * 0.01;
+        }
+        // AdamW momentum/velocity
+        this.mW = new Float32Array(d * d);
+        this.vW = new Float32Array(d * d);
+        this.mB = new Float32Array(d);
+        this.vB = new Float32Array(d);
+    }
+    /**
+     * Create a new contrastive trainer.
+     */
+    static async create(config) {
+        if (!Number.isFinite(config.dimension) || config.dimension < 1 || config.dimension > MAX_DIMENSION) {
+            throw new Error(`dimension must be between 1 and ${MAX_DIMENSION}`);
+        }
+        const trainer = new ContrastiveTrainer(config);
+        // Lazy-load NativeAccelerator for SIMD cosine + native AdamW (ADR-007)
+        try {
+            const { NativeAccelerator: N } = await import('./NativeAccelerator.js');
+            trainer.accel = new N();
+            await trainer.accel.initialize();
+        }
+        catch { /* proceed without acceleration */ }
+        return trainer;
+    }
+    /**
+     * Project an embedding through the learned transformation.
+     * Use this to enhance query/document embeddings after training.
+     */
+    project(embedding) {
+        this.ensureAlive();
+        if (embedding.length !== this.dim) {
+            throw new Error(`Expected dimension ${this.dim}, got ${embedding.length}`);
+        }
+        return this.matVecMul(this.weights, embedding, this.bias);
+    }
+    /**
+     * Compute InfoNCE loss for a batch of contrastive samples.
+     *
+     * L = -log(exp(sim(a, p) / t) / sum(exp(sim(a, n_i) / t)))
+     */
+    computeLoss(samples) {
+        this.ensureAlive();
+        if (samples.length === 0)
+            return 0;
+        let totalLoss = 0;
+        for (const sample of samples) {
+            const projAnchor = this.matVecMul(this.weights, sample.anchor, this.bias);
+            const projPos = this.matVecMul(this.weights, sample.positive, this.bias);
+            // ADR-007 Phase 1: delegate to native InfoNCE when available
+            if (this.accel && this.accel.nativeInfoNceAvailable) {
+                const projNegs = sample.negatives.map(n => this.matVecMul(this.weights, n, this.bias));
+                totalLoss += this.accel.infoNceLoss(projAnchor, projPos, projNegs, this.temperature);
+                continue;
+            }
+            const posSim = this.cosineSimilarity(projAnchor, projPos) / this.temperature;
+            let logSumExp = Math.exp(posSim);
+            for (const neg of sample.negatives) {
+                const projNeg = this.matVecMul(this.weights, neg, this.bias);
+                const negSim = this.cosineSimilarity(projAnchor, projNeg) / this.temperature;
+                logSumExp += Math.exp(negSim);
+            }
+            totalLoss += -posSim + Math.log(logSumExp);
+        }
+        return totalLoss / samples.length;
+    }
+    /**
+     * Train on a batch of contrastive samples.
+     * Uses analytical backpropagation through the InfoNCE loss.
+     */
+    trainBatch(samples) {
+        this.ensureAlive();
+        const batchSize = Math.min(samples.length, this.maxBatchSize);
+        const batch = samples.slice(0, batchSize);
+        const d = this.dim;
+        // Forward pass: compute loss
+        const loss = this.computeLoss(batch);
+        // Backward pass: analytical gradients via chain rule
+        const wGrad = new Float32Array(d * d);
+        const bGrad = new Float32Array(d);
+        for (const sample of batch) {
+            // Forward: project all vectors
+            const aProj = this.matVecMul(this.weights, sample.anchor, this.bias);
+            const pProj = this.matVecMul(this.weights, sample.positive, this.bias);
+            const aNorm = this.vecNorm(aProj);
+            const pNorm = this.vecNorm(pProj);
+            if (aNorm < 1e-10 || pNorm < 1e-10)
+                continue;
+            // Normalized anchor and positive
+            const invAN = 1 / aNorm;
+            const invPN = 1 / pNorm;
+            const aHat = new Float32Array(d);
+            const pHat = new Float32Array(d);
+            for (let i = 0; i < d; i++) {
+                aHat[i] = aProj[i] * invAN;
+                pHat[i] = pProj[i] * invPN;
+            }
+            // Positive cosine similarity
+            let cosPos = 0;
+            for (let i = 0; i < d; i++)
+                cosPos += aHat[i] * pHat[i];
+            const sPos = cosPos / this.temperature;
+            // Project and normalize negatives, compute cosine similarities
+            const negHats = [];
+            const negNorms = [];
+            const negCos = [];
+            let Z = Math.exp(sPos);
+            for (const neg of sample.negatives) {
+                const nProj = this.matVecMul(this.weights, neg, this.bias);
+                const nNorm = this.vecNorm(nProj);
+                negNorms.push(nNorm);
+                const nHat = new Float32Array(d);
+                if (nNorm > 1e-10) {
+                    const invNN = 1 / nNorm;
+                    for (let i = 0; i < d; i++)
+                        nHat[i] = nProj[i] * invNN;
+                }
+                negHats.push(nHat);
+                let cosN = 0;
+                for (let i = 0; i < d; i++)
+                    cosN += aHat[i] * nHat[i];
+                negCos.push(cosN);
+                Z += Math.exp(cosN / this.temperature);
+            }
+            // Gradient of loss w.r.t. cosine similarities
+            // dL/d(cos_pos) = (-1 + exp(s_pos)/Z) / τ
+            const dLdCosPos = (-1 + Math.exp(sPos) / Z) / this.temperature;
+            // Accumulate dL/da' from positive + all negatives
+            // d(cos(â,b̂))/da' = (b̂ - cos*â) / ||a'||
+            const dLda = new Float32Array(d);
+            for (let i = 0; i < d; i++) {
+                dLda[i] = dLdCosPos * (pHat[i] - cosPos * aHat[i]) * invAN;
+            }
+            // dL/dp' = dLdCosPos * (â - cos*p̂) / ||p'||
+            const dLdp = new Float32Array(d);
+            for (let i = 0; i < d; i++) {
+                dLdp[i] = dLdCosPos * (aHat[i] - cosPos * pHat[i]) * invPN;
+            }
+            // Process negatives
+            const dLdn = [];
+            for (let n = 0; n < negHats.length; n++) {
+                const dLdCosNeg = Math.exp(negCos[n] / this.temperature) / (this.temperature * Z);
+                const nNorm = negNorms[n];
+                // Accumulate anchor gradient from this negative
+                if (nNorm > 1e-10) {
+                    const invNN = 1 / nNorm;
+                    for (let i = 0; i < d; i++) {
+                        dLda[i] += dLdCosNeg * (negHats[n][i] - negCos[n] * aHat[i]) * invAN;
+                    }
+                    // Gradient w.r.t. negative projection
+                    const g = new Float32Array(d);
+                    for (let i = 0; i < d; i++) {
+                        g[i] = dLdCosNeg * (aHat[i] - negCos[n] * negHats[n][i]) * invNN;
+                    }
+                    dLdn.push(g);
+                }
+                else {
+                    dLdn.push(new Float32Array(d));
+                }
+            }
+            // Accumulate weight and bias gradients via outer products
+            // dL/dW += outer(dL/da', anchor) + outer(dL/dp', positive) + sum(outer(dL/dn_i', neg_i))
+            for (let i = 0; i < d; i++) {
+                const rowOff = i * d;
+                const dLda_i = dLda[i];
+                const dLdp_i = dLdp[i];
+                for (let j = 0; j < d; j++) {
+                    wGrad[rowOff + j] += dLda_i * sample.anchor[j] + dLdp_i * sample.positive[j];
+                }
+                bGrad[i] += dLda_i + dLdp_i;
+            }
+            // Negative outer products
+            for (let n = 0; n < sample.negatives.length; n++) {
+                for (let i = 0; i < d; i++) {
+                    const rowOff = i * d;
+                    const g_i = dLdn[n][i];
+                    for (let j = 0; j < d; j++) {
+                        wGrad[rowOff + j] += g_i * sample.negatives[n][j];
+                    }
+                    bGrad[i] += g_i;
+                }
+            }
+        }
+        // Average over batch
+        const invN = 1 / batchSize;
+        let totalGradNorm = 0;
+        for (let i = 0; i < wGrad.length; i++) {
+            wGrad[i] *= invN;
+            totalGradNorm += wGrad[i] * wGrad[i];
+        }
+        for (let i = 0; i < bGrad.length; i++) {
+            bGrad[i] *= invN;
+            totalGradNorm += bGrad[i] * bGrad[i];
+        }
+        totalGradNorm = Math.sqrt(totalGradNorm);
+        // AdamW update
+        this.step++;
+        this.adamWUpdate(this.weights, wGrad, this.mW, this.vW);
+        this.adamWUpdate(this.bias, bGrad, this.mB, this.vB);
+        // Update stats
+        this._totalBatches++;
+        this._totalSamples += batchSize;
+        this._lossSum += loss;
+        if (loss < this._bestLoss)
+            this._bestLoss = loss;
+        // Advance curriculum
+        this.batchesInStage++;
+        if (this.currentStage < this.stages.length &&
+            this.batchesInStage >= this.stages[this.currentStage].batches) {
+            this.currentStage = Math.min(this.currentStage + 1, this.stages.length - 1);
+            this.batchesInStage = 0;
+        }
+        return {
+            loss,
+            batchSize,
+            avgGradNorm: totalGradNorm,
+        };
+    }
+    /**
+     * Mine hard negatives from a pool of embeddings.
+     * Returns negatives that are similar to anchor but from different classes.
+     *
+     * SOTA: Positive-aware filtering (NV-Retriever, 2024). Rejects candidates
+     * with high similarity to known positives to eliminate false negatives,
+     * which constitute ~70% of naively mined hard negatives.
+     */
+    mineHardNegatives(anchor, pool, excludeIndices, count, positives) {
+        this.ensureAlive();
+        const stage = this.stages[Math.min(this.currentStage, this.stages.length - 1)];
+        const n = Math.min(count ?? stage.negativeCount, MAX_NEGATIVES, pool.length);
+        const falseNegThreshold = 0.85; // NV-Retriever recommended threshold
+        // Score all candidates by similarity to anchor
+        const scored = [];
+        for (let i = 0; i < pool.length; i++) {
+            if (excludeIndices.has(i))
+                continue;
+            const sim = this.cosineSimilarity(anchor, pool[i]);
+            if (sim < stage.hardNegativeThreshold)
+                continue;
+            // SOTA: Positive-aware false negative filtering
+            // Skip candidates too similar to any known positive (likely false negatives)
+            if (positives && positives.length > 0) {
+                let isFalseNeg = false;
+                for (const pos of positives) {
+                    if (this.cosineSimilarity(pool[i], pos) > falseNegThreshold) {
+                        isFalseNeg = true;
+                        break;
+                    }
+                }
+                if (isFalseNeg)
+                    continue;
+            }
+            scored.push({ idx: i, sim });
+        }
+        // Sort by similarity descending (hardest negatives first)
+        scored.sort((a, b) => b.sim - a.sim);
+        return scored.slice(0, n).map((s) => pool[s.idx]);
+    }
+    /**
+     * Get the NativeAccelerator instance (for testing/introspection).
+     */
+    getAccelerator() {
+        return this.accel;
+    }
+    /**
+     * Get the current curriculum stage info.
+     */
+    get currentCurriculumStage() {
+        return this.currentStage;
+    }
+    /**
+     * Get training statistics.
+     */
+    getStats() {
+        return {
+            totalBatches: this._totalBatches,
+            totalSamples: this._totalSamples,
+            avgLoss: this._totalBatches > 0 ? this._lossSum / this._totalBatches : 0,
+            bestLoss: this._bestLoss === Infinity ? 0 : this._bestLoss,
+            currentDifficulty: this.currentStage / Math.max(1, this.stages.length - 1),
+            epoch: this.currentStage,
+        };
+    }
+    /** Check if destroyed */
+    get isDestroyed() {
+        return this._destroyed;
+    }
+    /** Get configured dimension */
+    get dimension() {
+        return this.dim;
+    }
+    /** Destroy and free resources */
+    destroy() {
+        if (!this._destroyed) {
+            this._destroyed = true;
+        }
+    }
+    // --- Private helpers ---
+    matVecMul(W, x, b) {
+        const d = this.dim;
+        const result = new Float32Array(d);
+        for (let i = 0; i < d; i++) {
+            let sum = b[i];
+            const rowOffset = i * d;
+            for (let j = 0; j < d; j++) {
+                sum += W[rowOffset + j] * x[j];
+            }
+            result[i] = sum;
+        }
+        return result;
+    }
+    cosineSimilarity(a, b) {
+        // Delegate to NativeAccelerator for SIMD-accelerated cosine when available
+        if (this.accel)
+            return this.accel.cosineSimilarity(a, b);
+        let dot = 0, normA = 0, normB = 0;
+        for (let i = 0; i < a.length; i++) {
+            dot += a[i] * b[i];
+            normA += a[i] * a[i];
+            normB += b[i] * b[i];
+        }
+        const denom = Math.sqrt(normA * normB);
+        return denom > 0 ? dot / denom : 0;
+    }
+    vecNorm(v) {
+        let sum = 0;
+        for (let i = 0; i < v.length; i++)
+            sum += v[i] * v[i];
+        return Math.sqrt(sum);
+    }
+    adamWUpdate(params, grads, m, v) {
+        // Delegate to NativeAccelerator when available (ADR-007)
+        if (this.accel) {
+            this.accel.adamWStep(params, grads, m, v, this.step, this.learningRate, this.weightDecay);
+            return;
+        }
+        const beta1 = 0.9;
+        const beta2 = 0.999;
+        const eps = 1e-8;
+        const lr = this.learningRate;
+        const wd = this.weightDecay;
+        const bc1 = 1 - Math.pow(beta1, this.step);
+        const bc2 = 1 - Math.pow(beta2, this.step);
+        for (let i = 0; i < params.length; i++) {
+            // AdamW: decouple weight decay
+            params[i] -= lr * wd * params[i];
+            // Adam update
+            m[i] = beta1 * m[i] + (1 - beta1) * grads[i];
+            v[i] = beta2 * v[i] + (1 - beta2) * grads[i] * grads[i];
+            const mHat = m[i] / bc1;
+            const vHat = v[i] / bc2;
+            params[i] -= lr * mHat / (Math.sqrt(vHat) + eps);
+        }
+    }
+    ensureAlive() {
+        if (this._destroyed) {
+            throw new Error('ContrastiveTrainer has been destroyed. Create a new instance.');
+        }
+    }
+}
+//# sourceMappingURL=ContrastiveTrainer.js.map

package/dist/src/backends/rvf/ContrastiveTrainer.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"ContrastiveTrainer.js","sourceRoot":"","sources":["../../../../src/backends/rvf/ContrastiveTrainer.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AAsDH,SAAS;AACT,MAAM,eAAe,GAAG,IAAI,CAAC;AAC7B,MAAM,eAAe,GAAG,GAAG,CAAC;AAC5B,MAAM,cAAc,GAAG,GAAG,CAAC;AAC3B,MAAM,aAAa,GAAG,GAAG,CAAC;AAC1B,MAAM,aAAa,GAAG,IAAI,CAAC;AAE3B;;;;;;;;;GASG;AACH,MAAM,OAAO,kBAAkB;IACrB,GAAG,CAAS;IACZ,WAAW,CAAS;IACpB,YAAY,CAAS;IACrB,WAAW,CAAS;IACpB,YAAY,CAAS;IACrB,MAAM,CAAoB;IAC1B,YAAY,GAAG,CAAC,CAAC;IACjB,cAAc,GAAG,CAAC,CAAC;IAE3B,2CAA2C;IACnC,OAAO,CAAe;IACtB,IAAI,CAAe;IAE3B,cAAc;IACN,EAAE,CAAe;IACjB,EAAE,CAAe;IACjB,EAAE,CAAe;IACjB,EAAE,CAAe;IACjB,IAAI,GAAG,CAAC,CAAC;IAEjB,4DAA4D;IACpD,KAAK,GAA6B,IAAI,CAAC;IAE/C,QAAQ;IACA,aAAa,GAAG,CAAC,CAAC;IAClB,aAAa,GAAG,CAAC,CAAC;IAClB,QAAQ,GAAG,CAAC,CAAC;IACb,SAAS,GAAG,QAAQ,CAAC;IACrB,UAAU,GAAG,KAAK,CAAC;IAE3B,YAAoB,MAAyB;QAC3C,IAAI,CAAC,GAAG,GAAG,MAAM,CAAC,SAAS,CAAC;QAC5B,IAAI,CAAC,WAAW,GAAG,IAAI,CAAC,GAAG,CACzB,IAAI,CAAC,GAAG,CAAC,eAAe,EAAE,MAAM,CAAC,WAAW,IAAI,IAAI,CAAC,EACrD,eAAe,CAChB,CAAC;QACF,IAAI,CAAC,YAAY,GAAG,IAAI,CAAC,GAAG,CAAC,IAAI,EAAE,MAAM,CAAC,YAAY,IAAI,KAAK,CAAC,CAAC;QACjE,IAAI,CAAC,WAAW,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,MAAM,CAAC,WAAW,IAAI,IAAI,CAAC,CAAC;QAC3D,IAAI,CAAC,YAAY,GAAG,IAAI,CAAC,GAAG,CAC1B,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,MAAM,CAAC,YAAY,IAAI,EAAE,CAAC,EACtC,cAAc,CACf,CAAC;QACF,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC,MAAM,IAAI;YAC7B,EAAE,aAAa,EAAE,CAAC,EAAE,qBAAqB,EAAE,GAAG,EAAE,OAAO,EAAE,GAAG,EAAE;YAC9D,EAAE,aAAa,EAAE,CAAC,EAAE,qBAAqB,EAAE,GAAG,EAAE,OAAO,EAAE,GAAG,EAAE;YAC9D,EAAE,aAAa,EAAE,EAAE,EAAE,qBAAqB,EAAE,GAAG,EAAE,OAAO,EAAE,GAAG,EAAE;SAChE,CAAC;QAEF,kDAAkD;QAClD,MAAM,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC;QACnB,IAAI,CAAC,OAAO,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;QACvC,IAAI,CAAC,IAAI,GAAG,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;QAChC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YAC3B,IAAI,CAAC,OAAO,CAAC,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,oBAAoB;QACrD,CAAC;QACD,4BAA4B;QAC5B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,OAAO,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YAC7C,IAAI,CAAC,OAAO,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,MAAM,EAAE,GAAG,GAAG,CAAC,GAAG,IAAI,CAAC;QAClD,CAAC;QAED,0BAA0B;QAC1B,IAAI,CAAC,EAAE,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;QAClC,IAAI,CAAC,EAAE,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;QAClC,IAAI,CAAC,EAAE,GAAG,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;QAC9B,IAAI,CAAC,EAAE,GAAG,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;IAChC,CAAC;IAED;;OAEG;IACH,MAAM,CAAC,KAAK,CAAC,MAAM,CAAC,MAAyB;QAC3C,IAAI,CAAC,MAAM,CAAC,QAAQ,CAAC,MAAM,CAAC,SAAS,CAAC,IAAI,MAAM,CAAC,SAAS,GAAG,CAAC,IAAI,MAAM,CAAC,SAAS,GAAG,aAAa,EAAE,CAAC;YACnG,MAAM,IAAI,KAAK,CAAC,mCAAmC,aAAa,EAAE,CAAC,CAAC;QACtE,CAAC;QACD,MAAM,OAAO,GAAG,IAAI,kBAAkB,CAAC,MAAM,CAAC,CAAC;QAC/C,uEAAuE;QACvE,IAAI,CAAC;YACH,MAAM,EAAE,iBAAiB,EAAE,CAAC,EAAE,GAAG,MAAM,MAAM,CAAC,wBAAwB,CAAC,CAAC;YACxE,OAAO,CAAC,KAAK,GAAG,IAAI,CAAC,EAAE,CAAC;YACxB,MAAM,OAAO,CAAC,KAAK,CAAC,UAAU,EAAE,CAAC;QACnC,CAAC;QAAC,MAAM,CAAC,CAAC,kCAAkC,CAAC,CAAC;QAC9C,OAAO,OAAO,CAAC;IACjB,CAAC;IAED;;;OAGG;IACH,OAAO,CAAC,SAAuB;QAC7B,IAAI,CAAC,WAAW,EAAE,CAAC;QACnB,IAAI,SAAS,CAAC,MAAM,KAAK,IAAI,CAAC,GAAG,EAAE,CAAC;YAClC,MAAM,IAAI,KAAK,CAAC,sBAAsB,IAAI,CAAC,GAAG,SAAS,SAAS,CAAC,MAAM,EAAE,CAAC,CAAC;QAC7E,CAAC;QACD,OAAO,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC,OAAO,EAAE,SAAS,EAAE,IAAI,CAAC,IAAI,CAAC,CAAC;IAC5D,CAAC;IAED;;;;OAIG;IACH,WAAW,CAAC,OAA4B;QACtC,IAAI,CAAC,WAAW,EAAE,CAAC;QACnB,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC;YAAE,OAAO,CAAC,CAAC;QAEnC,IAAI,SAAS,GAAG,CAAC,CAAC;QAClB,KAAK,MAAM,MAAM,IAAI,OAAO,EAAE,CAAC;YAC7B,MAAM,UAAU,GAAG,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC,OAAO,EAAE,MAAM,CAAC,MAAM,EAAE,IAAI,CAAC,IAAI,CAAC,CAAC;YAC1E,MAAM,OAAO,GAAG,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC,OAAO,EAAE,MAAM,CAAC,QAAQ,EAAE,IAAI,CAAC,IAAI,CAAC,CAAC;YAEzE,6DAA6D;YAC7D,IAAI,IAAI,CAAC,KAAK,IAAI,IAAI,CAAC,KAAK,CAAC,sBAAsB,EAAE,CAAC;gBACpD,MAAM,QAAQ,GAAG,MAAM,CAAC,SAAS,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC,OAAO,EAAE,CAAC,EAAE,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC;gBACvF,SAAS,IAAI,IAAI,CAAC,KAAK,CAAC,WAAW,CAAC,UAAU,EAAE,OAAO,EAAE,QAAQ,EAAE,IAAI,CAAC,WAAW,CAAC,CAAC;gBACrF,SAAS;YACX,CAAC;YAED,MAAM,MAAM,GAAG,IAAI,CAAC,gBAAgB,CAAC,UAAU,EAAE,OAAO,CAAC,GAAG,IAAI,CAAC,WAAW,CAAC;YAE7E,IAAI,SAAS,GAAG,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC;YACjC,KAAK,MAAM,GAAG,IAAI,MAAM,CAAC,SAAS,EAAE,CAAC;gBACnC,MAAM,OAAO,GAAG,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC,OAAO,EAAE,GAAG,EAAE,IAAI,CAAC,IAAI,CAAC,CAAC;gBAC7D,MAAM,MAAM,GAAG,IAAI,CAAC,gBAAgB,CAAC,UAAU,EAAE,OAAO,CAAC,GAAG,IAAI,CAAC,WAAW,CAAC;gBAC7E,SAAS,IAAI,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC;YAChC,CAAC;YAED,SAAS,IAAI,CAAC,MAAM,GAAG,IAAI,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC;QAC7C,CAAC;QAED,OAAO,SAAS,GAAG,OAAO,CAAC,MAAM,CAAC;IACpC,CAAC;IAED;;;OAGG;IACH,UAAU,CAAC,OAA4B;QACrC,IAAI,CAAC,WAAW,EAAE,CAAC;QACnB,MAAM,SAAS,GAAG,IAAI,CAAC,GAAG,CAAC,OAAO,CAAC,MAAM,EAAE,IAAI,CAAC,YAAY,CAAC,CAAC;QAC9D,MAAM,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,SAAS,CAAC,CAAC;QAC1C,MAAM,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC;QAEnB,6BAA6B;QAC7B,MAAM,IAAI,GAAG,IAAI,CAAC,WAAW,CAAC,KAAK,CAAC,CAAC;QAErC,qDAAqD;QACrD,MAAM,KAAK,GAAG,IAAI,YAAY,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;QACtC,MAAM,KAAK,GAAG,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;QAElC,KAAK,MAAM,MAAM,IAAI,KAAK,EAAE,CAAC;YAC3B,+BAA+B;YAC/B,MAAM,KAAK,GAAG,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC,OAAO,EAAE,MAAM,CAAC,MAAM,EAAE,IAAI,CAAC,IAAI,CAAC,CAAC;YACrE,MAAM,KAAK,GAAG,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC,OAAO,EAAE,MAAM,CAAC,QAAQ,EAAE,IAAI,CAAC,IAAI,CAAC,CAAC;YAEvE,MAAM,KAAK,GAAG,IAAI,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;YAClC,MAAM,KAAK,GAAG,IAAI,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;YAClC,IAAI,KAAK,GAAG,KAAK,IAAI,KAAK,GAAG,KAAK;gBAAE,SAAS;YAE7C,iCAAiC;YACjC,MAAM,KAAK,GAAG,CAAC,GAAG,KAAK,CAAC;YACxB,MAAM,KAAK,GAAG,CAAC,GAAG,KAAK,CAAC;YACxB,MAAM,IAAI,GAAG,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;YACjC,MAAM,IAAI,GAAG,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;YACjC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;gBAC3B,IAAI,CAAC,CAAC,CAAC,GAAG,KAAK,CAAC,CAAC,CAAC,GAAG,KAAK,CAAC;gBAC3B,IAAI,CAAC,CAAC,CAAC,GAAG,KAAK,CAAC,CAAC,CAAC,GAAG,KAAK,CAAC;YAC7B,CAAC;YAED,6BAA6B;YAC7B,IAAI,MAAM,GAAG,CAAC,CAAC;YACf,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE;gBAAE,MAAM,IAAI,IAAI,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC;YACxD,MAAM,IAAI,GAAG,MAAM,GAAG,IAAI,CAAC,WAAW,CAAC;YAEvC,+DAA+D;YAC/D,MAAM,OAAO,GAAmB,EAAE,CAAC;YACnC,MAAM,QAAQ,GAAa,EAAE,CAAC;YAC9B,MAAM,MAAM,GAAa,EAAE,CAAC;YAC5B,IAAI,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;YAEvB,KAAK,MAAM,GAAG,IAAI,MAAM,CAAC,SAAS,EAAE,CAAC;gBACnC,MAAM,KAAK,GAAG,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC,OAAO,EAAE,GAAG,EAAE,IAAI,CAAC,IAAI,CAAC,CAAC;gBAC3D,MAAM,KAAK,GAAG,IAAI,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;gBAClC,QAAQ,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;gBAErB,MAAM,IAAI,GAAG,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;gBACjC,IAAI,KAAK,GAAG,KAAK,EAAE,CAAC;oBAClB,MAAM,KAAK,GAAG,CAAC,GAAG,KAAK,CAAC;oBACxB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE;wBAAE,IAAI,CAAC,CAAC,CAAC,GAAG,KAAK,CAAC,CAAC,CAAC,GAAG,KAAK,CAAC;gBACzD,CAAC;gBACD,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;gBAEnB,IAAI,IAAI,GAAG,CAAC,CAAC;gBACb,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE;oBAAE,IAAI,IAAI,IAAI,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC;gBACtD,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;gBAClB,CAAC,IAAI,IAAI,CAAC,GAAG,CAAC,IAAI,GAAG,IAAI,CAAC,WAAW,CAAC,CAAC;YACzC,CAAC;YAED,8CAA8C;YAC9C,0CAA0C;YAC1C,MAAM,SAAS,GAAG,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,WAAW,CAAC;YAE/D,kDAAkD;YAClD,2CAA2C;YAC3C,MAAM,IAAI,GAAG,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;YACjC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;gBAC3B,IAAI,CAAC,CAAC,CAAC,GAAG,SAAS,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,GAAG,MAAM,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC,GAAG,KAAK,CAAC;YAC7D,CAAC;YAED,6CAA6C;YAC7C,MAAM,IAAI,GAAG,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;YACjC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;gBAC3B,IAAI,CAAC,CAAC,CAAC,GAAG,SAAS,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,GAAG,MAAM,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC,GAAG,KAAK,CAAC;YAC7D,CAAC;YAED,oBAAoB;YACpB,MAAM,IAAI,GAAmB,EAAE,CAAC;YAChC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,OAAO,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;gBACxC,MAAM,SAAS,GAAG,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,WAAW,CAAC,GAAG,CAAC,IAAI,CAAC,WAAW,GAAG,CAAC,CAAC,CAAC;gBAClF,MAAM,KAAK,GAAG,QAAQ,CAAC,CAAC,CAAC,CAAC;gBAE1B,gDAAgD;gBAChD,IAAI,KAAK,GAAG,KAAK,EAAE,CAAC;oBAClB,MAAM,KAAK,GAAG,CAAC,GAAG,KAAK,CAAC;oBACxB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;wBAC3B,IAAI,CAAC,CAAC,CAAC,IAAI,SAAS,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,MAAM,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC,GAAG,KAAK,CAAC;oBACvE,CAAC;oBAED,sCAAsC;oBACtC,MAAM,CAAC,GAAG,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;oBAC9B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;wBAC3B,CAAC,CAAC,CAAC,CAAC,GAAG,SAAS,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC,GAAG,MAAM,CAAC,CAAC,CAAC,GAAG,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,KAAK,CAAC;oBACnE,CAAC;oBACD,IAAI,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC;gBACf,CAAC;qBAAM,CAAC;oBACN,IAAI,CAAC,IAAI,CAAC,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC,CAAC;gBACjC,CAAC;YACH,CAAC;YAED,0DAA0D;YAC1D,yFAAyF;YACzF,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;gBAC3B,MAAM,MAAM,GAAG,CAAC,GAAG,CAAC,CAAC;gBACrB,MAAM,MAAM,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC;gBACvB,MAAM,MAAM,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC;gBACvB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;oBAC3B,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,IAAI,MAAM,GAAG,MAAM,CAAC,MAAM,CAAC,CAAC,CAAC,GAAG,MAAM,GAAG,MAAM,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC;gBAC/E,CAAC;gBACD,KAAK,CAAC,CAAC,CAAC,IAAI,MAAM,GAAG,MAAM,CAAC;YAC9B,CAAC;YAED,0BAA0B;YAC1B,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,CAAC,SAAS,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;gBACjD,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;oBAC3B,MAAM,MAAM,GAAG,CAAC,GAAG,CAAC,CAAC;oBACrB,MAAM,GAAG,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;oBACvB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;wBAC3B,KAAK,CAAC,MAAM,GAAG,CAAC,CAAC,IAAI,GAAG,GAAG,MAAM,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;oBACpD,CAAC;oBACD,KAAK,CAAC,CAAC,CAAC,IAAI,GAAG,CAAC;gBAClB,CAAC;YACH,CAAC;QACH,CAAC;QAED,qBAAqB;QACrB,MAAM,IAAI,GAAG,CAAC,GAAG,SAAS,CAAC;QAC3B,IAAI,aAAa,GAAG,CAAC,CAAC;QACtB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,KAAK,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACtC,KAAK,CAAC,CAAC,CAAC,IAAI,IAAI,CAAC;YACjB,aAAa,IAAI,KAAK,CAAC,CAAC,CAAC,GAAG,KAAK,CAAC,CAAC,CAAC,CAAC;QACvC,CAAC;QACD,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,KAAK,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACtC,KAAK,CAAC,CAAC,CAAC,IAAI,IAAI,CAAC;YACjB,aAAa,IAAI,KAAK,CAAC,CAAC,CAAC,GAAG,KAAK,CAAC,CAAC,CAAC,CAAC;QACvC,CAAC;QACD,aAAa,GAAG,IAAI,CAAC,IAAI,CAAC,aAAa,CAAC,CAAC;QAEzC,eAAe;QACf,IAAI,CAAC,IAAI,EAAE,CAAC;QACZ,IAAI,CAAC,WAAW,CAAC,IAAI,CAAC,OAAO,EAAE,KAAK,EAAE,IAAI,CAAC,EAAE,EAAE,IAAI,CAAC,EAAE,CAAC,CAAC;QACxD,IAAI,CAAC,WAAW,CAAC,IAAI,CAAC,IAAI,EAAE,KAAK,EAAE,IAAI,CAAC,EAAE,EAAE,IAAI,CAAC,EAAE,CAAC,CAAC;QAErD,eAAe;QACf,IAAI,CAAC,aAAa,EAAE,CAAC;QACrB,IAAI,CAAC,aAAa,IAAI,SAAS,CAAC;QAChC,IAAI,CAAC,QAAQ,IAAI,IAAI,CAAC;QACtB,IAAI,IAAI,GAAG,IAAI,CAAC,SAAS;YAAE,IAAI,CAAC,SAAS,GAAG,IAAI,CAAC;QAEjD,qBAAqB;QACrB,IAAI,CAAC,cAAc,EAAE,CAAC;QACtB,IAAI,IAAI,CAAC,YAAY,GAAG,IAAI,CAAC,MAAM,CAAC,MAAM;YACtC,IAAI,CAAC,cAAc,IAAI,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,YAAY,CAAC,CAAC,OAAO,EAAE,CAAC;YAClE,IAAI,CAAC,YAAY,GAAG,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,YAAY,GAAG,CAAC,EAAE,IAAI,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;YAC5E,IAAI,CAAC,cAAc,GAAG,CAAC,CAAC;QAC1B,CAAC;QAED,OAAO;YACL,IAAI;YACJ,SAAS;YACT,WAAW,EAAE,aAAa;SAC3B,CAAC;IACJ,CAAC;IAED;;;;;;;OAOG;IACH,iBAAiB,CACf,MAAoB,EACpB,IAAoB,EACpB,cAA2B,EAC3B,KAAc,EACd,SAA0B;QAE1B,IAAI,CAAC,WAAW,EAAE,CAAC;QACnB,MAAM,KAAK,GAAG,IAAI,CAAC,MAAM,CAAC,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,YAAY,EAAE,IAAI,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC;QAC/E,MAAM,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,IAAI,KAAK,CAAC,aAAa,EAAE,aAAa,EAAE,IAAI,CAAC,MAAM,CAAC,CAAC;QAC7E,MAAM,iBAAiB,GAAG,IAAI,CAAC,CAAC,qCAAqC;QAErE,+CAA+C;QAC/C,MAAM,MAAM,GAAwC,EAAE,CAAC;QACvD,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACrC,IAAI,cAAc,CAAC,GAAG,CAAC,CAAC,CAAC;gBAAE,SAAS;YACpC,MAAM,GAAG,GAAG,IAAI,CAAC,gBAAgB,CAAC,MAAM,EAAE,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC;YACnD,IAAI,GAAG,GAAG,KAAK,CAAC,qBAAqB;gBAAE,SAAS;YAEhD,gDAAgD;YAChD,6EAA6E;YAC7E,IAAI,SAAS,IAAI,SAAS,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;gBACtC,IAAI,UAAU,GAAG,KAAK,CAAC;gBACvB,KAAK,MAAM,GAAG,IAAI,SAAS,EAAE,CAAC;oBAC5B,IAAI,IAAI,CAAC,gBAAgB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,GAAG,CAAC,GAAG,iBAAiB,EAAE,CAAC;wBAC5D,UAAU,GAAG,IAAI,CAAC;wBAClB,MAAM;oBACR,CAAC;gBACH,CAAC;gBACD,IAAI,UAAU;oBAAE,SAAS;YAC3B,CAAC;YAED,MAAM,CAAC,IAAI,CAAC,EAAE,GAAG,EAAE,CAAC,EAAE,GAAG,EAAE,CAAC,CAAC;QAC/B,CAAC;QAED,0DAA0D;QAC1D,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC;QAErC,OAAO,MAAM,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,IAAI,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC;IACpD,CAAC;IAED;;OAEG;IACH,cAAc;QACZ,OAAO,IAAI,CAAC,KAAK,CAAC;IACpB,CAAC;IAED;;OAEG;IACH,IAAI,sBAAsB;QACxB,OAAO,IAAI,CAAC,YAAY,CAAC;IAC3B,CAAC;IAED;;OAEG;IACH,QAAQ;QACN,OAAO;YACL,YAAY,EAAE,IAAI,CAAC,aAAa;YAChC,YAAY,EAAE,IAAI,CAAC,aAAa;YAChC,OAAO,EAAE,IAAI,CAAC,aAAa,GAAG,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,QAAQ,GAAG,IAAI,CAAC,aAAa,CAAC,CAAC,CAAC,CAAC;YACxE,QAAQ,EAAE,IAAI,CAAC,SAAS,KAAK,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,SAAS;YAC1D,iBAAiB,EAAE,IAAI,CAAC,YAAY,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,IAAI,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,CAAC;YAC1E,KAAK,EAAE,IAAI,CAAC,YAAY;SACzB,CAAC;IACJ,CAAC;IAED,yBAAyB;IACzB,IAAI,WAAW;QACb,OAAO,IAAI,CAAC,UAAU,CAAC;IACzB,CAAC;IAED,+BAA+B;IAC/B,IAAI,SAAS;QACX,OAAO,IAAI,CAAC,GAAG,CAAC;IAClB,CAAC;IAED,iCAAiC;IACjC,OAAO;QACL,IAAI,CAAC,IAAI,CAAC,UAAU,EAAE,CAAC;YACrB,IAAI,CAAC,UAAU,GAAG,IAAI,CAAC;QACzB,CAAC;IACH,CAAC;IAED,0BAA0B;IAElB,SAAS,CAAC,CAAe,EAAE,CAAe,EAAE,CAAe;QACjE,MAAM,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC;QACnB,MAAM,MAAM,GAAG,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;QACnC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;YAC3B,IAAI,GAAG,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;YACf,MAAM,SAAS,GAAG,CAAC,GAAG,CAAC,CAAC;YACxB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;gBAC3B,GAAG,IAAI,CAAC,CAAC,SAAS,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;YACjC,CAAC;YACD,MAAM,CAAC,CAAC,CAAC,GAAG,GAAG,CAAC;QAClB,CAAC;QACD,OAAO,MAAM,CAAC;IAChB,CAAC;IAEO,gBAAgB,CAAC,CAAe,EAAE,CAAe;QACvD,2EAA2E;QAC3E,IAAI,IAAI,CAAC,KAAK;YAAE,OAAO,IAAI,CAAC,KAAK,CAAC,gBAAgB,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;QACzD,IAAI,GAAG,GAAG,CAAC,EAAE,KAAK,GAAG,CAAC,EAAE,KAAK,GAAG,CAAC,CAAC;QAClC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YAClC,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;YACnB,KAAK,IAAI,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;YACrB,KAAK,IAAI,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;QACvB,CAAC;QACD,MAAM,KAAK,GAAG,IAAI,CAAC,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC,CAAC;QACvC,OAAO,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,GAAG,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC;IACrC,CAAC;IAEO,OAAO,CAAC,CAAe;QAC7B,IAAI,GAAG,GAAG,CAAC,CAAC;QACZ,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,MAAM,EAAE,CAAC,EAAE;YAAE,GAAG,IAAI,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC;QACtD,OAAO,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;IACxB,CAAC;IAEO,WAAW,CACjB,MAAoB,EACpB,KAAmB,EACnB,CAAe,EACf,CAAe;QAEf,yDAAyD;QACzD,IAAI,IAAI,CAAC,KAAK,EAAE,CAAC;YACf,IAAI,CAAC,KAAK,CAAC,SAAS,CAAC,MAAM,EAAE,KAAK,EAAE,CAAC,EAAE,CAAC,EAAE,IAAI,CAAC,IAAI,EAAE,IAAI,CAAC,YAAY,EAAE,IAAI,CAAC,WAAW,CAAC,CAAC;YAC1F,OAAO;QACT,CAAC;QAED,MAAM,KAAK,GAAG,GAAG,CAAC;QAClB,MAAM,KAAK,GAAG,KAAK,CAAC;QACpB,MAAM,GAAG,GAAG,IAAI,CAAC;QACjB,MAAM,EAAE,GAAG,IAAI,CAAC,YAAY,CAAC;QAC7B,MAAM,EAAE,GAAG,IAAI,CAAC,WAAW,CAAC;QAE5B,MAAM,GAAG,GAAG,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,EAAE,IAAI,CAAC,IAAI,CAAC,CAAC;QAC3C,MAAM,GAAG,GAAG,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,EAAE,IAAI,CAAC,IAAI,CAAC,CAAC;QAE3C,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,MAAM,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;YACvC,+BAA+B;YAC/B,MAAM,CAAC,CAAC,CAAC,IAAI,EAAE,GAAG,EAAE,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC;YAEjC,cAAc;YACd,CAAC,CAAC,CAAC,CAAC,GAAG,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,KAAK,CAAC,GAAG,KAAK,CAAC,CAAC,CAAC,CAAC;YAC7C,CAAC,CAAC,CAAC,CAAC,GAAG,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,GAAG,KAAK,CAAC,GAAG,KAAK,CAAC,CAAC,CAAC,GAAG,KAAK,CAAC,CAAC,CAAC,CAAC;YAExD,MAAM,IAAI,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,GAAG,CAAC;YACxB,MAAM,IAAI,GAAG,CAAC,CAAC,CAAC,CAAC,GAAG,GAAG,CAAC;YAExB,MAAM,CAAC,CAAC,CAAC,IAAI,EAAE,GAAG,IAAI,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,GAAG,CAAC,CAAC;QACnD,CAAC;IACH,CAAC;IAEO,WAAW;QACjB,IAAI,IAAI,CAAC,UAAU,EAAE,CAAC;YACpB,MAAM,IAAI,KAAK,CAAC,+DAA+D,CAAC,CAAC;QACnF,CAAC;IACH,CAAC;CACF"}