@claude-flow/cli 3.32.3 → 3.32.4

package/dist/src/mcp-tools/neural-tools.js

@@ -0,0 +1,1168 @@
+/**
+ * Neural MCP Tools for CLI
+ *
+ * V2 Compatibility - Neural network and ML tools
+ *
+ * ✅ HYBRID Implementation:
+ * - Uses @claude-flow/embeddings for REAL ML embeddings when available
+ * - Falls back to deterministic hash-based embeddings when ML model not installed
+ * - Pattern storage and search with cosine similarity (real math in all tiers)
+ * - Training stores patterns as searchable embeddings (not simulated)
+ *
+ * Note: For production neural features, use @claude-flow/neural module
+ */
+import { getProjectCwd } from './types.js';
+import { validateIdentifier, validateText } from './validate-input.js';
+import { existsSync, readFileSync, writeFileSync, mkdirSync } from 'node:fs';
+import { join } from 'node:path';
+let _corpusStatsCache = null;
+// (query, store) → embedding + cosine cache. Small LRU (not size-1): the flywheel
+// scores many configs across several queries in interleaved order, so a size-1
+// cache would thrash. Cap keeps memory bounded; eviction is oldest-first.
+const _cosineCache = new Map();
+const _COSINE_CACHE_MAX = 128;
+function _cosineCacheGet(key) {
+    const v = _cosineCache.get(key);
+    if (v) {
+        _cosineCache.delete(key);
+        _cosineCache.set(key, v);
+    } // LRU touch
+    return v;
+}
+function _cosineCacheSet(key, v) {
+    _cosineCache.set(key, v);
+    if (_cosineCache.size > _COSINE_CACHE_MAX)
+        _cosineCache.delete(_cosineCache.keys().next().value);
+}
+function corpusFingerprint(patterns) {
+    let h = 2166136261 >>> 0;
+    for (const p of patterns) {
+        const s = `${p.id}|${p.name?.length ?? 0}|${p.content?.length ?? 0}`;
+        for (let i = 0; i < s.length; i++) {
+            h ^= s.charCodeAt(i);
+            h = Math.imul(h, 16777619) >>> 0;
+        }
+    }
+    return `${patterns.length}:${h.toString(16)}`;
+}
+let _champCache = null;
+function activeChampionParams() {
+    try {
+        if (_champCache && Date.now() - _champCache.at < 30_000)
+            return _champCache.params;
+        const p = join(getProjectCwd(), '.claude-flow', 'harness-active-policy.json');
+        let params = {};
+        if (existsSync(p)) {
+            const active = JSON.parse(readFileSync(p, 'utf-8'));
+            if (active && !active.rolledBack && active.params && typeof active.params === 'object')
+                params = active.params;
+        }
+        _champCache = { at: Date.now(), params };
+        return params;
+    }
+    catch {
+        return {};
+    }
+}
+// Real embeddings — resolved LAZILY on first use.
+// Perf (measured 2026-07): the previous top-level-await version of this block
+// ran `await import('ruvector')` + `initOnnxEmbedder()` + a probe embed at
+// module-import time, adding ~450ms (warm) to ~2800ms (cold) to EVERY CLI
+// command, because mcp-client statically imports this module. ensureEmbeddings()
+// keeps the exact same tier chain and degraded/fallback semantics, but only
+// pays the cost when a handler actually needs an embedding.
+// Tier 0 (ADR-089): ruvector@0.2.27 bundled ONNX (no sharp dep, fixes ADR-086's
+//   silent-fallback bug at source; closes the chain described in ruvnet/ruvector#523).
+// Tier 1: agentic-flow v3 ReasoningBank (was Tier 1 — broken on darwin-arm64 without sharp)
+// Tier 2-3: @claude-flow/embeddings
+let realEmbeddings = null;
+let embeddingServiceName = 'none';
+let embeddingsPromise = null;
+/**
+ * Memoized lazy initialiser for the embedding provider chain. Safe to call
+ * concurrently (single shared promise). If every tier fails to import or
+ * probe, `realEmbeddings` stays null and callers use the explicit
+ * hash-fallback path — identical degraded semantics to the old eager block.
+ */
+function ensureEmbeddings() {
+    if (embeddingsPromise)
+        return embeddingsPromise;
+    embeddingsPromise = (async () => {
+        try {
+            // Tier -1 (PRIMARY): optional WASM embedder — a pluggable real embedder
+            // ahead of everything. OPT-IN (RUFLO_EMBED_WASM_PKG) + FAIL-CLOSED: unset
+            // or absent/init-fail ⇒ fall straight through to the existing
+            // ruvector-ONNX → hash tiers with zero regression.
+            try {
+                const we = await import('../ruvector/wasm-embedder.js').catch(() => null);
+                if (we && await we.wasmEmbedderAvailable()) {
+                    const model = we.DEFAULT_EMBED_MODEL;
+                    realEmbeddings = {
+                        embed: async (text) => {
+                            const v = await we.wasmEmbed(text, model);
+                            if (!v || !v.length)
+                                throw new Error('wasm embed failed'); // → generateEmbedding falls to next tier
+                            return v;
+                        },
+                    };
+                    embeddingServiceName = `wasm-embedder/${model} (${we.wasmEmbedderModels().length} model${we.wasmEmbedderModels().length === 1 ? '' : 's'})`;
+                }
+            }
+            catch { /* not configured — fall through */ }
+            // Tier 0: ruvector@0.2.27 — bundled all-MiniLM-L6-v2 + parallel worker pool.
+            // Probe with isOnnxAvailable() and verify an actual embed succeeds (avoids
+            // the type-load-success-but-runtime-fails trap from ADR-086). The probe now
+            // runs on first embed request instead of at import time.
+            // NOTE: ruvector's embed() returns `{embedding, dimension, timeMs}` — we
+            // unwrap to plain number[] for the shared interface.
+            const rv = !realEmbeddings ? await import('ruvector').catch(() => null) : null;
+            if (!realEmbeddings && rv?.embed && typeof rv.embed === 'function' && rv.isOnnxAvailable?.()) {
+                try {
+                    if (typeof rv.initOnnxEmbedder === 'function')
+                        await rv.initOnnxEmbedder();
+                    const probe = await rv.embed('probe');
+                    // Handle both shapes: ruvector wraps as {embedding, dimension, timeMs};
+                    // some versions returned raw Float32Array.
+                    const probeVec = probe?.embedding ?? probe;
+                    if (probeVec && (Array.isArray(probeVec) || probeVec.length > 0)) {
+                        realEmbeddings = {
+                            embed: async (text) => {
+                                const r = await rv.embed(text);
+                                const v = r?.embedding ?? r;
+                                return Array.isArray(v) ? v : Array.from(v);
+                            },
+                        };
+                        embeddingServiceName = 'ruvector@0.2.27 (bundled all-MiniLM-L6-v2)';
+                    }
+                }
+                catch {
+                    // ruvector embed failed at runtime; fall through to next tier
+                }
+            }
+            // Tier 1: agentic-flow v3 ReasoningBank (kept for backward-compat; may
+            // silently fall back on darwin-arm64 without sharp — that's the bug
+            // Tier 0 was added to bypass).
+            if (!realEmbeddings) {
+                const rb = await import('agentic-flow/reasoningbank').catch(() => null);
+                if (rb?.computeEmbedding) {
+                    realEmbeddings = { embed: (text) => rb.computeEmbedding(text) };
+                    embeddingServiceName = 'agentic-flow/reasoningbank';
+                }
+            }
+            // Tier 2: @claude-flow/embeddings with agentic-flow provider
+            if (!realEmbeddings) {
+                const embeddingsModule = await import('@claude-flow/embeddings').catch(() => null);
+                if (embeddingsModule?.createEmbeddingService) {
+                    try {
+                        const service = embeddingsModule.createEmbeddingService({ provider: 'agentic-flow' });
+                        realEmbeddings = {
+                            embed: async (text) => {
+                                const result = await service.embed(text);
+                                return Array.from(result.embedding);
+                            },
+                        };
+                        embeddingServiceName = 'agentic-flow';
+                    }
+                    catch {
+                        // agentic-flow provider not available, try ONNX
+                    }
+                }
+            }
+            // Tier 3: @claude-flow/embeddings with ONNX provider
+            if (!realEmbeddings) {
+                const embeddingsModule = await import('@claude-flow/embeddings').catch(() => null);
+                if (embeddingsModule?.createEmbeddingService) {
+                    try {
+                        const service = embeddingsModule.createEmbeddingService({ provider: 'onnx' });
+                        realEmbeddings = {
+                            embed: async (text) => {
+                                const result = await service.embed(text);
+                                return Array.from(result.embedding);
+                            },
+                        };
+                        embeddingServiceName = 'onnx';
+                    }
+                    catch {
+                        // ONNX provider not available, fall through to mock
+                    }
+                }
+            }
+            // No Tier 4 mock fallback. If all real-embedder tiers fail to import or
+            // probe, leave realEmbeddings null and let downstream code use the
+            // explicit hash-fallback path with a clear _embeddingNote in stats.
+            // Silently substituting mock embeddings would hide a missing production
+            // dependency from callers — that's the bug ADR-086 was about.
+        }
+        catch {
+            // No embedding provider available, will use fallback
+        }
+    })();
+    return embeddingsPromise;
+}
+// Storage paths
+const STORAGE_DIR = '.claude-flow';
+const NEURAL_DIR = 'neural';
+const MODELS_FILE = 'models.json';
+const PATTERNS_FILE = 'patterns.json';
+function getNeuralDir() {
+    return join(getProjectCwd(), STORAGE_DIR, NEURAL_DIR);
+}
+function getNeuralPath() {
+    return join(getNeuralDir(), MODELS_FILE);
+}
+function ensureNeuralDir() {
+    const dir = getNeuralDir();
+    if (!existsSync(dir)) {
+        mkdirSync(dir, { recursive: true });
+    }
+}
+function loadNeuralStore() {
+    try {
+        const path = getNeuralPath();
+        if (existsSync(path)) {
+            return JSON.parse(readFileSync(path, 'utf-8'));
+        }
+    }
+    catch {
+        // Return empty store
+    }
+    return { models: {}, patterns: {}, version: '3.0.0' };
+}
+/**
+ * ADR-176 flywheel: expose the stored patterns (id/name/content) so the
+ * self-optimizing loop can harvest a benchmark corpus from real usage. Additive,
+ * read-only, never throws.
+ */
+export function getStorePatterns() {
+    try {
+        return Object.values(loadNeuralStore().patterns ?? {}).map((p) => ({ id: p.id, name: p.name, content: p.content }));
+    }
+    catch {
+        return [];
+    }
+}
+function saveNeuralStore(store) {
+    ensureNeuralDir();
+    writeFileSync(getNeuralPath(), JSON.stringify(store, null, 2), 'utf-8');
+}
+/**
+ * Public helper: read-only stats about the neural store, for the unified
+ * learning-stats aggregator. Returns total pattern count + per-type breakdown
+ * without exposing the embeddings.
+ */
+export function getNeuralStoreStats() {
+    const store = loadNeuralStore();
+    const patterns = Object.values(store.patterns ?? {});
+    const byType = {};
+    for (const p of patterns) {
+        const t = p.type || 'unknown';
+        byType[t] = (byType[t] ?? 0) + 1;
+    }
+    return {
+        patternCount: patterns.length,
+        byType,
+        modelCount: Object.values(store.models ?? {}).length,
+        source: '.claude-flow/neural/patterns.json (loadNeuralStore)',
+    };
+}
+/**
+ * Public helper: store an array of patterns into the neural store so they
+ * surface via `neural_patterns list`. Used by hooks_pretrain so its extracted
+ * patterns are actually queryable, not just bundled in the `pretrain` namespace.
+ * #2245.
+ *
+ * Returns the number of patterns written.
+ */
+export async function storeNeuralPatterns(items) {
+    if (!items || items.length === 0)
+        return { stored: 0, total: 0 };
+    // realEmbeddings is initialised lazily by ensureEmbeddings() (invoked from
+    // generateEmbedding()); it falls back to a hash-based embedding if no
+    // provider is available.
+    const store = loadNeuralStore();
+    let stored = 0;
+    for (const item of items) {
+        if (!item.name || !item.type)
+            continue;
+        const sourceText = item.content ?? item.name;
+        const embedding = await generateEmbedding(sourceText);
+        const id = `pattern-${Date.now()}-${stored.toString(36)}-${Math.random().toString(36).slice(2, 6)}`;
+        store.patterns[id] = {
+            id,
+            name: String(item.name).slice(0, 200),
+            type: String(item.type).slice(0, 64),
+            embedding,
+            content: typeof sourceText === 'string' ? sourceText.slice(0, 4096) : undefined,
+            metadata: item.metadata ?? {},
+            createdAt: new Date().toISOString(),
+            usageCount: 0,
+        };
+        stored++;
+    }
+    saveNeuralStore(store);
+    return { stored, total: items.length };
+}
+// Generate embedding - uses real ML embeddings if available, falls back to deterministic hash
+async function generateEmbedding(text, dims = 384) {
+    // Lazily resolve the embedding provider on first real use (perf: keeps
+    // ONNX init off the CLI startup path — see ensureEmbeddings()).
+    if (text)
+        await ensureEmbeddings();
+    // If real embeddings available and text provided, use them
+    if (realEmbeddings && text) {
+        try {
+            return await realEmbeddings.embed(text);
+        }
+        catch {
+            // Fall back to hash-based
+        }
+    }
+    // Hash-based deterministic embedding (better than pure random for consistency)
+    // NOTE: No semantic meaning — only useful for consistent deduplication, not similarity search
+    if (text) {
+        (await import('../memory/embedding-policy.js')).enforceNoStub('neural-tools.generateEmbedding'); // "no stubs" strict mode
+        if (embeddingServiceName === 'none') {
+            embeddingServiceName = 'hash-fallback';
+        }
+        const hash = text.split('').reduce((acc, char, i) => {
+            return acc + char.charCodeAt(0) * (i + 1);
+        }, 0);
+        // Use hash to seed a deterministic embedding
+        const embedding = [];
+        let seed = hash;
+        for (let i = 0; i < dims; i++) {
+            // Simple LCG random with seed
+            seed = (seed * 1103515245 + 12345) & 0x7fffffff;
+            embedding.push((seed / 0x7fffffff) * 2 - 1);
+        }
+        return embedding;
+    }
+    // No text provided — return zero vector (callers should always provide text)
+    return new Array(dims).fill(0);
+}
+// Cosine similarity for pattern search
+function cosineSimilarity(a, b) {
+    if (a.length !== b.length)
+        return 0;
+    let dotProduct = 0;
+    let normA = 0;
+    let normB = 0;
+    for (let i = 0; i < a.length; i++) {
+        dotProduct += a[i] * b[i];
+        normA += a[i] * a[i];
+        normB += b[i] * b[i];
+    }
+    return dotProduct / (Math.sqrt(normA) * Math.sqrt(normB) || 1);
+}
+export const neuralTools = [
+    {
+        name: 'neural_train',
+        description: 'Train a neural model Use when nothing native trains on your workflow — Claude Code has no learning loop. Use to train SONA/MoE/EWC patterns from successful task outcomes; query via neural_predict before spawning agents. Off-path for one-shot work.',
+        category: 'neural',
+        inputSchema: {
+            type: 'object',
+            properties: {
+                modelId: { type: 'string', description: 'Model ID to train' },
+                modelType: { type: 'string', enum: ['moe', 'transformer', 'classifier', 'embedding'], description: 'Model type' },
+                epochs: { type: 'number', description: 'Number of training epochs' },
+                learningRate: { type: 'number', description: 'Learning rate' },
+                data: { type: 'object', description: 'Training data' },
+            },
+            required: ['modelType'],
+        },
+        handler: async (input) => {
+            if (input.modelId) {
+                const v = validateIdentifier(input.modelId, 'modelId');
+                if (!v.valid)
+                    return { success: false, error: v.error };
+            }
+            const store = loadNeuralStore();
+            const modelId = input.modelId || `model-${Date.now()}-${Math.random().toString(36).slice(2, 6)}`;
+            const modelType = input.modelType;
+            const epochs = input.epochs || 10;
+            const model = {
+                id: modelId,
+                name: `${modelType}-model`,
+                type: modelType,
+                status: 'training',
+                accuracy: 0,
+                epochs,
+                config: {
+                    learningRate: input.learningRate || 0.001,
+                    batchSize: 32,
+                },
+            };
+            store.models[modelId] = model;
+            saveNeuralStore(store);
+            // Real training: embed training data and store as searchable patterns
+            const trainingData = input.data;
+            let patternsStored = 0;
+            if (trainingData) {
+                const entries = Array.isArray(trainingData) ? trainingData : [trainingData];
+                for (let i = 0; i < entries.length; i++) {
+                    const entry = entries[i];
+                    const text = typeof entry === 'string' ? entry
+                        : entry?.text
+                            || entry?.content
+                            || entry?.label
+                            || JSON.stringify(entry);
+                    if (!text)
+                        continue;
+                    const embedding = await generateEmbedding(text, 384);
+                    const patternId = `${modelId}-train-${i}`;
+                    // ADR-093 F11: extract a meaningful label instead of dumping raw
+                    // training JSON as the pattern name. Audit reported neural_predict
+                    // returned `label: <raw training data JSON>` because the previous
+                    // fallback was `text.slice(0, 100)` where text was `JSON.stringify(entry)`.
+                    let label;
+                    if (typeof entry === 'string') {
+                        label = entry.slice(0, 80);
+                    }
+                    else if (entry && typeof entry === 'object') {
+                        const e = entry;
+                        // Prefer common semantic fields over a JSON dump
+                        const labelField = e.label ?? e.category ?? e.class ?? e.tag ?? e.intent ?? e.name ?? e.title;
+                        if (typeof labelField === 'string' && labelField.length > 0) {
+                            label = labelField.slice(0, 80);
+                        }
+                        else {
+                            const summaryField = e.text ?? e.input ?? e.task ?? e.description ?? e.content;
+                            if (typeof summaryField === 'string' && summaryField.length > 0) {
+                                label = `${summaryField.slice(0, 60)}${summaryField.length > 60 ? '…' : ''}`;
+                            }
+                            else {
+                                // Last resort: reduce to a stable short hash-like id
+                                label = `${modelType}:entry-${i}`;
+                            }
+                        }
+                    }
+                    else {
+                        label = `${modelType}:entry-${i}`;
+                    }
+                    store.patterns[patternId] = {
+                        id: patternId,
+                        name: label,
+                        type: modelType,
+                        embedding,
+                        metadata: { modelId, epoch: epochs, index: i, raw: entry },
+                        createdAt: new Date().toISOString(),
+                        usageCount: 0,
+                    };
+                    patternsStored++;
+                }
+            }
+            model.status = 'ready';
+            model.accuracy = patternsStored > 0 ? 1.0 : 0; // accuracy = data stored, not simulated
+            model.trainedAt = new Date().toISOString();
+            saveNeuralStore(store);
+            return {
+                success: true,
+                _realEmbedding: !!realEmbeddings,
+                _embeddingSource: embeddingServiceName,
+                embeddingProvider: embeddingServiceName,
+                modelId,
+                type: modelType,
+                status: model.status,
+                patternsStored,
+                totalPatterns: Object.keys(store.patterns).length,
+                epochs,
+                trainedAt: model.trainedAt,
+                ...(embeddingServiceName === 'hash-fallback' || embeddingServiceName === 'none' ? {
+                    platformNote: 'ONNX embeddings not available — using hash-based fallback. Install @claude-flow/embeddings and run "embeddings init --download" for semantic search.',
+                } : {}),
+            };
+        },
+    },
+    {
+        name: 'neural_predict',
+        description: 'Make predictions using a neural model Use when nothing native trains on your workflow — Claude Code has no learning loop. Use to train SONA/MoE/EWC patterns from successful task outcomes; query via neural_predict before spawning agents. Off-path for one-shot work.',
+        category: 'neural',
+        inputSchema: {
+            type: 'object',
+            properties: {
+                modelId: { type: 'string', description: 'Model ID to use' },
+                input: { type: 'string', description: 'Input text or data' },
+                topK: { type: 'number', description: 'Number of top predictions' },
+            },
+            required: ['input'],
+        },
+        handler: async (input) => {
+            {
+                const v = validateText(input.input, 'input');
+                if (!v.valid)
+                    return { success: false, error: v.error };
+            }
+            if (input.modelId) {
+                const v = validateIdentifier(input.modelId, 'modelId');
+                if (!v.valid)
+                    return { success: false, error: v.error };
+            }
+            const store = loadNeuralStore();
+            const modelId = input.modelId;
+            const inputText = input.input;
+            const topK = input.topK || 3;
+            // Find model or use default
+            const model = modelId ? store.models[modelId] : Object.values(store.models).find(m => m.status === 'ready');
+            if (model && model.status !== 'ready') {
+                return { success: false, error: 'Model not ready' };
+            }
+            // Generate real embedding for the input
+            const startTime = performance.now();
+            const embedding = await generateEmbedding(inputText, 384);
+            const latency = Math.round(performance.now() - startTime);
+            // ADR-093 F11: real classifier head over stored patterns. Previously
+            // confidence was the raw cosine similarity (often clamped to 0 when
+            // stored embeddings were stale or zero-vectored). Now we run k-NN
+            // with cosine distance and apply a temperature-controlled softmax
+            // over the top-K so confidence is a proper distribution that sums
+            // to 1, and we surface enough metadata to trust the result.
+            const storedPatterns = Object.values(store.patterns);
+            let predictions;
+            if (storedPatterns.length > 0) {
+                // Step 1: k-NN with cosine
+                const scored = storedPatterns
+                    .map(p => {
+                    const sim = cosineSimilarity(embedding, p.embedding);
+                    return {
+                        patternId: p.id,
+                        label: p.name || p.type || p.id,
+                        cosineSimilarity: sim,
+                    };
+                })
+                    .sort((a, b) => b.cosineSimilarity - a.cosineSimilarity)
+                    .slice(0, topK);
+                // Step 2: temperature-softmax over the top-K so confidence sums to 1.
+                // Temperature 0.1 sharpens differences between similar candidates.
+                const tau = 0.1;
+                const exps = scored.map(s => Math.exp(s.cosineSimilarity / tau));
+                const z = exps.reduce((a, b) => a + b, 0) || 1;
+                predictions = scored.map((s, i) => ({
+                    label: s.label,
+                    patternId: s.patternId,
+                    cosineSimilarity: Number(s.cosineSimilarity.toFixed(4)),
+                    confidence: Number((exps[i] / z).toFixed(4)),
+                }));
+            }
+            else {
+                // No patterns stored — no predictions possible. Be honest about it
+                // instead of returning empty silently.
+                predictions = [];
+            }
+            const topConfidence = predictions[0]?.confidence ?? 0;
+            const topSimilarity = predictions[0]?.cosineSimilarity ?? 0;
+            return {
+                success: true,
+                _realEmbedding: !!realEmbeddings,
+                _embeddingSource: embeddingServiceName,
+                embeddingProvider: embeddingServiceName,
+                _hasStoredPatterns: storedPatterns.length > 0,
+                _classifierHead: storedPatterns.length > 0 ? 'knn-cosine+softmax(tau=0.1)' : 'none',
+                modelId: model?.id || 'default',
+                input: inputText,
+                predictions,
+                // Surface cosineSimilarity separately so callers know whether the
+                // softmax confidence reflects true match strength.
+                topPrediction: predictions[0]?.label ?? null,
+                topConfidence,
+                topSimilarity,
+                embedding: embedding.slice(0, 8), // Preview of embedding
+                embeddingDims: embedding.length,
+                latency,
+                ...(storedPatterns.length === 0 ? {
+                    _note: 'No patterns stored. Train with neural_train(modelType, trainingData) before predicting.',
+                } : {}),
+            };
+        },
+    },
+    {
+        name: 'neural_patterns',
+        description: 'Get or manage neural patterns Use when nothing native trains on your workflow — Claude Code has no learning loop. Use to train SONA/MoE/EWC patterns from successful task outcomes; query via neural_predict before spawning agents. Off-path for one-shot work.',
+        category: 'neural',
+        inputSchema: {
+            type: 'object',
+            properties: {
+                action: { type: 'string', enum: ['list', 'get', 'store', 'search', 'delete'], description: 'Action to perform' },
+                patternId: { type: 'string', description: 'Pattern ID' },
+                name: { type: 'string', description: 'Pattern name' },
+                type: { type: 'string', description: 'Pattern type' },
+                content: { type: 'string', description: 'Pattern source text (used for BM25 in hybrid search; falls back to name)' },
+                query: { type: 'string', description: 'Search query' },
+                limit: { type: 'number', description: 'Top-K results to return (default 10, max 100)' },
+                mode: { type: 'string', enum: ['hybrid', 'cosine'], description: 'Search mode — hybrid (cosine+BM25+MMR, default) or cosine (pre-3.10.18 behaviour, for A/B)' },
+                alpha: { type: 'number', description: 'Hybrid: cosine weight in [0,1]; (1-α) is BM25 weight (default 0.5, tuned ADR-082)' },
+                mmrLambda: { type: 'number', description: 'Hybrid: MMR balance — 1.0 = pure relevance, 0.0 = pure diversity (default 0.7, tuned ADR-082)' },
+                subjectWeight: { type: 'number', description: 'Hybrid: multi-field BM25 weight for subject/name (default 2.0 non-rerank, 3.0 with rerank — tuned ADR-082/083)' },
+                bodyWeight: { type: 'number', description: 'Hybrid: multi-field BM25 weight for body/content (default 1.0)' },
+                typePenaltyFactor: { type: 'number', description: 'Hybrid: meta-commit score multiplier — release/merge/bump commits × this factor (default 1.0 = disabled; set 0.5 for aggressive suppression)' },
+                rerank: { type: 'boolean', description: 'Hybrid: opt-in cross-encoder rerank pass over the top-K (ADR-080). Adds ~20-40 ms per (query, doc) pair; first call downloads ~30MB model. Gracefully degrades to hybrid+MMR order when unavailable.' },
+                hybridWeight: { type: 'number', description: 'Rerank: hybrid score weight in final combination (default 0.7, tuned ADR-083)' },
+                ceWeight: { type: 'number', description: 'Rerank: cross-encoder score weight in final combination (default 0.3, tuned ADR-083)' },
+                data: { type: 'object', description: 'Pattern data' },
+            },
+        },
+        handler: async (input) => {
+            if (input.patternId) {
+                const v = validateIdentifier(input.patternId, 'patternId');
+                if (!v.valid)
+                    return { success: false, error: v.error };
+            }
+            if (input.name) {
+                const v = validateText(input.name, 'name');
+                if (!v.valid)
+                    return { success: false, error: v.error };
+            }
+            if (input.type) {
+                const v = validateIdentifier(input.type, 'type');
+                if (!v.valid)
+                    return { success: false, error: v.error };
+            }
+            if (input.query) {
+                const v = validateText(input.query, 'query');
+                if (!v.valid)
+                    return { success: false, error: v.error };
+            }
+            const store = loadNeuralStore();
+            const action = input.action || 'list';
+            if (action === 'list') {
+                const patterns = Object.values(store.patterns);
+                const typeFilter = input.type;
+                const filtered = typeFilter ? patterns.filter(p => p.type === typeFilter) : patterns;
+                return {
+                    patterns: filtered.map(p => ({
+                        id: p.id,
+                        name: p.name,
+                        type: p.type,
+                        usageCount: p.usageCount,
+                        createdAt: p.createdAt,
+                    })),
+                    total: filtered.length,
+                };
+            }
+            if (action === 'get') {
+                const pattern = store.patterns[input.patternId];
+                if (!pattern) {
+                    return { success: false, error: 'Pattern not found' };
+                }
+                return { success: true, pattern };
+            }
+            if (action === 'store') {
+                const patternId = `pattern-${Date.now()}-${Math.random().toString(36).slice(2, 6)}`;
+                const patternName = input.name || 'Unnamed pattern';
+                const patternContent = input.content ?? patternName;
+                // Generate embedding from pattern content (falls back to name).
+                const embedding = await generateEmbedding(patternContent, 384);
+                const pattern = {
+                    id: patternId,
+                    name: patternName,
+                    type: input.type || 'general',
+                    embedding,
+                    content: typeof patternContent === 'string' ? patternContent.slice(0, 4096) : undefined,
+                    metadata: input.data || {},
+                    createdAt: new Date().toISOString(),
+                    usageCount: 0,
+                };
+                store.patterns[patternId] = pattern;
+                saveNeuralStore(store);
+                return {
+                    success: true,
+                    _realEmbedding: !!realEmbeddings,
+                    _embeddingSource: embeddingServiceName,
+                    embeddingProvider: embeddingServiceName,
+                    patternId,
+                    name: pattern.name,
+                    type: pattern.type,
+                    embeddingDims: embedding.length,
+                    createdAt: pattern.createdAt,
+                };
+            }
+            if (action === 'search') {
+                const query = input.query;
+                const k = Math.min(Math.max(Number(input.limit ?? input.topK ?? 10), 1), 100);
+                // ADR-078 hybrid retrieval controls. Cosine-only mode preserves the
+                // pre-3.10.18 behaviour for A/B tests via {mode:'cosine'}; default is
+                // hybrid (cosine + BM25 + MMR).
+                const mode = String(input.mode ?? 'hybrid');
+                const useRerank = input.rerank === true || String(input.rerank) === 'true';
+                // ADR-083 joint grid: rerank path benefits from DIFFERENT hybrid
+                // sub-params than non-rerank (the cross-encoder adds semantic depth,
+                // so the hybrid stage can be more keyword-focused). nDCG@3 0.900 →
+                // 0.963 on rerank just by switching sw 2.0 → 3.0 in the hybrid stage.
+                // Champion-provided defaults (ADR-176/177): explicit input wins, then the
+                // adopted proven-config champion, then the hardcoded ADR-082 defaults.
+                const champ = activeChampionParams();
+                const alpha = Number(input.alpha ?? champ.alpha ?? 0.5);
+                const mmrLambda = Number(input.mmrLambda ?? champ.mmrLambda ?? 0.7);
+                const { tokenize, buildCorpusStats, hybridScores, mmrRerank, multiFieldBM25, typePenalty } = await import('../memory/hybrid-retrieval.js');
+                const patterns = Object.values(store.patterns);
+                // Query embedding + cosine array depend only on (query, store) — NOT the
+                // retrieval config. Cache them so scoring many configs for one query
+                // (the flywheel's access pattern) embeds + cosines once, not per config.
+                const _cosKey = `${corpusFingerprint(patterns)}::${query}`;
+                let queryEmbedding, cosineArr;
+                const _hit = _cosineCacheGet(_cosKey);
+                if (_hit) {
+                    queryEmbedding = _hit.queryEmbedding;
+                    cosineArr = _hit.cosineArr;
+                }
+                else {
+                    queryEmbedding = await generateEmbedding(query, 384);
+                    cosineArr = patterns.map((p) => cosineSimilarity(queryEmbedding, p.embedding));
+                    _cosineCacheSet(_cosKey, { queryEmbedding, cosineArr });
+                }
+                if (mode === 'cosine') {
+                    const ranked = patterns
+                        .map((p, i) => ({ ...p, similarity: cosineArr[i] }))
+                        .sort((a, b) => b.similarity - a.similarity)
+                        .slice(0, k);
+                    return {
+                        _realSimilarity: true,
+                        _realEmbedding: !!realEmbeddings,
+                        _embeddingSource: embeddingServiceName,
+                        embeddingProvider: embeddingServiceName,
+                        mode: 'cosine',
+                        query,
+                        results: ranked.map((r) => ({ id: r.id, name: r.name, type: r.type, similarity: r.similarity })),
+                        total: ranked.length,
+                    };
+                }
+                // Hybrid path — multi-field BM25 (subject 3×, body 1×) + type penalty
+                // for meta-commits (release bumps / merges) per ADR-079. Falls back
+                // to single-field BM25 when no content is stored.
+                // Cache the (config-independent) tokenized docs + BM25 stats by store fingerprint.
+                const _fp = corpusFingerprint(patterns);
+                let subjectDocs, bodyDocs, subjectStats, bodyStats;
+                if (_corpusStatsCache && _corpusStatsCache.fp === _fp) {
+                    subjectDocs = _corpusStatsCache.subjectDocs;
+                    bodyDocs = _corpusStatsCache.bodyDocs;
+                    subjectStats = _corpusStatsCache.subjectStats;
+                    bodyStats = _corpusStatsCache.bodyStats;
+                }
+                else {
+                    subjectDocs = patterns.map((p) => tokenize(p.name ?? ''));
+                    bodyDocs = patterns.map((p) => {
+                        // Body is content minus the subject — if content starts with name,
+                        // strip it; otherwise use full content (with name removed if duplicated).
+                        const c = p.content ?? '';
+                        const n = p.name ?? '';
+                        return tokenize(c.startsWith(n) ? c.slice(n.length) : c);
+                    });
+                    subjectStats = buildCorpusStats(subjectDocs);
+                    bodyStats = buildCorpusStats(bodyDocs);
+                    _corpusStatsCache = { fp: _fp, subjectDocs, bodyDocs, subjectStats, bodyStats };
+                }
+                const queryTokens = tokenize(query);
+                // ADR-082: subjectWeight 3.0 → 2.0 from grid (sw=2 dominates at hybrid-only).
+                // ADR-083 joint grid: when rerank is on, the cross-encoder handles
+                // semantic understanding, so the hybrid stage can be MORE
+                // subject-focused (sw=3) — recovers nDCG@3 0.963.
+                const subjectWeight = Number(input.subjectWeight ?? champ.subjectWeight ?? (useRerank ? 3.0 : 2.0));
+                const bodyWeight = Number(input.bodyWeight ?? champ.bodyWeight ?? 1.0);
+                const bm25Arr = patterns.map((_, i) => multiFieldBM25(queryTokens, subjectDocs[i], bodyDocs[i], subjectStats, bodyStats, subjectWeight, bodyWeight));
+                const baseHybrid = hybridScores(cosineArr, bm25Arr, alpha);
+                // Type penalty — opt-in (default 1.0 = disabled). Ablation in ADR-079
+                // showed multi-field BM25 alone gives best top-1 (8/10 vs 7/10 with
+                // penalty enabled) because some relevant work commits also match the
+                // Merge/release regex. Callers wanting aggressive meta-commit
+                // suppression can set {typePenaltyFactor: 0.5}.
+                const typeFactor = Number(input.typePenaltyFactor ?? champ.typePenaltyFactor ?? 1.0);
+                const hybridArr = typeFactor === 1.0
+                    ? baseHybrid
+                    : baseHybrid.map((s, i) => s * typePenalty(patterns[i].name, typeFactor));
+                // Candidate pool sizing: k*3 for MMR, k*6 for cross-encoder (it needs
+                // more options to find the truly-best). ADR-080 ablation: rerank over
+                // a narrow post-MMR slice degrades top-1; reranking a wider hybrid
+                // top-K*6 pool restores and exceeds the no-rerank baseline.
+                // useRerank declared at top of search block for conditional defaults.
+                const poolSize = useRerank ? k * 6 : k * 3;
+                const prelim = patterns
+                    .map((p, i) => ({ p, hybrid: hybridArr[i], cosine: cosineArr[i], bm25: bm25Arr[i] }))
+                    .sort((a, b) => b.hybrid - a.hybrid)
+                    .slice(0, Math.min(poolSize, patterns.length));
+                const candidates = prelim.map(({ p, hybrid, cosine, bm25 }) => ({
+                    id: p.id, name: p.name, type: p.type,
+                    embedding: p.embedding,
+                    content: p.content,
+                    relevance: hybrid,
+                    _cosine: cosine,
+                    _bm25: bm25,
+                }));
+                let picked;
+                if (useRerank) {
+                    // ADR-080: cross-encoder reranks the wider candidate pool, then
+                    // final score = hybridWeight * hybrid + ceWeight * crossEncoder
+                    // on normalised scales. Ablation showed cross-encoder alone hits
+                    // 100% top-3 but loses top-1 (calibration on short commit subjects
+                    // is noisy); linear combination preserves hybrid's top-1 strength
+                    // while gaining the cross-encoder's recall.
+                    try {
+                        const { crossEncoderRerank } = await import('../memory/cross-encoder-rerank.js');
+                        const { normalise } = await import('../memory/hybrid-retrieval.js');
+                        const docs = candidates.map((c) => c.content || c.name);
+                        const reranked = await crossEncoderRerank(query, docs);
+                        const ceScores = new Array(candidates.length).fill(0);
+                        for (const { index, score } of reranked)
+                            ceScores[index] = score;
+                        const hybridNorm = normalise(candidates.map((c) => c.relevance));
+                        const ceNorm = normalise(ceScores);
+                        // ADR-083 joint grid: hw=0.7 cw=0.3 (with α=0.5 sw=3 from above) is
+                        // the joint optimum at nDCG@3=0.963 (vs 0.900 at 0.5/0.5). The
+                        // hybrid signal carries most of the relevance; the cross-encoder
+                        // contributes a 30% smoothing/disambiguation kick.
+                        const hybridWeight = Number(input.hybridWeight ?? 0.7);
+                        const ceWeight = Number(input.ceWeight ?? 0.3);
+                        const combined = candidates.map((c, i) => ({
+                            ...c,
+                            _crossEncoderScore: ceScores[i],
+                            _combinedScore: hybridWeight * hybridNorm[i] + ceWeight * ceNorm[i],
+                        }));
+                        picked = combined
+                            .sort((a, b) => b._combinedScore - a._combinedScore)
+                            .slice(0, k);
+                    }
+                    catch {
+                        picked = candidates.slice(0, k);
+                    }
+                }
+                else {
+                    // Default path: MMR diversity over top-K*3 hybrid candidates.
+                    picked = mmrRerank(candidates, k, mmrLambda);
+                }
+                return {
+                    _realSimilarity: true,
+                    _realEmbedding: !!realEmbeddings,
+                    _embeddingSource: embeddingServiceName,
+                    embeddingProvider: embeddingServiceName,
+                    mode: 'hybrid',
+                    alpha,
+                    mmrLambda,
+                    rerank: useRerank,
+                    query,
+                    results: picked.map((r) => ({
+                        id: r.id,
+                        name: r.name,
+                        type: r.type,
+                        similarity: r.relevance, // exposed as `similarity` for back-compat
+                        hybridScore: r.relevance,
+                        cosineScore: r._cosine,
+                        bm25Score: r._bm25,
+                        mmrScore: r.mmrScore,
+                        ...(r._crossEncoderScore !== undefined
+                            ? { crossEncoderScore: r._crossEncoderScore }
+                            : {}),
+                    })),
+                    total: picked.length,
+                };
+            }
+            if (action === 'delete') {
+                const patternId = input.patternId;
+                if (!store.patterns[patternId]) {
+                    return { success: false, error: 'Pattern not found' };
+                }
+                delete store.patterns[patternId];
+                saveNeuralStore(store);
+                return { success: true, deleted: patternId };
+            }
+            return { success: false, error: 'Unknown action' };
+        },
+    },
+    {
+        name: 'neural_compress',
+        description: 'Compress neural model or embeddings Use when nothing native trains on your workflow — Claude Code has no learning loop. Use to train SONA/MoE/EWC patterns from successful task outcomes; query via neural_predict before spawning agents. Off-path for one-shot work.',
+        category: 'neural',
+        inputSchema: {
+            type: 'object',
+            properties: {
+                modelId: { type: 'string', description: 'Model ID to compress' },
+                method: { type: 'string', enum: ['quantize', 'prune', 'distill'], description: 'Compression method' },
+                targetSize: { type: 'number', description: 'Target size reduction (0-1)' },
+            },
+        },
+        handler: async (input) => {
+            if (input.modelId) {
+                const v = validateIdentifier(input.modelId, 'modelId');
+                if (!v.valid)
+                    return { success: false, error: v.error };
+            }
+            // Resolve provider so embeddingProvider in the result matches the old
+            // eager-init reporting.
+            await ensureEmbeddings();
+            const store = loadNeuralStore();
+            const method = input.method || 'quantize';
+            const targetReduction = input.targetSize || 0.5;
+            const patterns = Object.values(store.patterns);
+            if (patterns.length === 0) {
+                return { success: false, error: 'No patterns to compress. Train patterns first with neural_train.' };
+            }
+            const beforeCount = patterns.length;
+            const beforeSize = patterns.reduce((s, p) => s + (p.embedding?.length || 0) * 4, 0); // Float32 = 4 bytes
+            if (method === 'quantize') {
+                try {
+                    const { quantizeInt8, getQuantizationStats } = await import('../memory/memory-initializer.js');
+                    let totalCompressed = 0;
+                    for (const pattern of patterns) {
+                        if (pattern.embedding && pattern.embedding.length > 0) {
+                            const stats = getQuantizationStats(pattern.embedding);
+                            const quantized = quantizeInt8(pattern.embedding);
+                            // Store quantized metadata (keep original embedding for search)
+                            pattern._quantized = {
+                                scale: quantized.scale,
+                                zeroPoint: quantized.zeroPoint,
+                                compressionRatio: stats.compressionRatio,
+                            };
+                            totalCompressed++;
+                        }
+                    }
+                    saveNeuralStore(store);
+                    return {
+                        success: true, _real: true, method,
+                        embeddingProvider: embeddingServiceName,
+                        patternsCompressed: totalCompressed,
+                        compressionRatio: '3.92x (Int8)',
+                        beforeBytes: beforeSize,
+                        afterBytes: Math.round(beforeSize / 3.92),
+                    };
+                }
+                catch {
+                    return { success: false, error: 'Quantization requires memory-initializer. Run `memory init` first.' };
+                }
+            }
+            if (method === 'prune') {
+                // Prune patterns with low usage count below threshold (targetReduction as min usage)
+                const threshold = targetReduction;
+                const toRemove = [];
+                for (const [id, pattern] of Object.entries(store.patterns)) {
+                    if ((pattern.usageCount || 0) < threshold)
+                        toRemove.push(id);
+                }
+                for (const id of toRemove)
+                    delete store.patterns[id];
+                saveNeuralStore(store);
+                return {
+                    success: true, _real: true, method,
+                    embeddingProvider: embeddingServiceName,
+                    threshold,
+                    patternsRemoved: toRemove.length,
+                    patternsBefore: beforeCount,
+                    patternsAfter: Object.keys(store.patterns).length,
+                };
+            }
+            if (method === 'distill') {
+                // Merge similar patterns by cosine similarity > 0.95
+                const patternList = Object.entries(store.patterns);
+                const merged = [];
+                for (let i = 0; i < patternList.length; i++) {
+                    const [idA, a] = patternList[i];
+                    if (merged.includes(idA))
+                        continue;
+                    for (let j = i + 1; j < patternList.length; j++) {
+                        const [idB, b] = patternList[j];
+                        if (!a.embedding || !b.embedding || merged.includes(idB))
+                            continue;
+                        const sim = cosineSimilarity(a.embedding, b.embedding);
+                        if (sim > 0.95) {
+                            // Merge: average embeddings, keep higher usage count
+                            for (let k = 0; k < a.embedding.length; k++) {
+                                a.embedding[k] = (a.embedding[k] + (b.embedding[k] || 0)) / 2;
+                            }
+                            a.usageCount = Math.max(a.usageCount || 0, b.usageCount || 0);
+                            delete store.patterns[idB];
+                            merged.push(idB);
+                        }
+                    }
+                }
+                saveNeuralStore(store);
+                return {
+                    success: true, _real: true, method,
+                    embeddingProvider: embeddingServiceName,
+                    patternsMerged: merged.length,
+                    patternsBefore: beforeCount,
+                    patternsAfter: Object.keys(store.patterns).length,
+                };
+            }
+            return { success: false, error: `Unknown method: ${method}. Use quantize, prune, or distill.` };
+        },
+    },
+    {
+        name: 'neural_status',
+        description: 'Get neural system status Use when nothing native trains on your workflow — Claude Code has no learning loop. Use to train SONA/MoE/EWC patterns from successful task outcomes; query via neural_predict before spawning agents. Off-path for one-shot work.',
+        category: 'neural',
+        inputSchema: {
+            type: 'object',
+            properties: {
+                modelId: { type: 'string', description: 'Specific model ID' },
+                detailed: { type: 'boolean', description: 'Include detailed info' },
+            },
+        },
+        handler: async (input) => {
+            if (input.modelId) {
+                const v = validateIdentifier(input.modelId, 'modelId');
+                if (!v.valid)
+                    return { success: false, error: v.error };
+            }
+            // Resolve provider so _realEmbeddings/embeddingProvider report the same
+            // values the old eager-init version produced.
+            await ensureEmbeddings();
+            const store = loadNeuralStore();
+            if (input.modelId) {
+                const model = store.models[input.modelId];
+                if (!model) {
+                    return { success: false, error: 'Model not found' };
+                }
+                return { success: true, model };
+            }
+            const models = Object.values(store.models);
+            const patterns = Object.values(store.patterns);
+            return {
+                _realEmbeddings: !!realEmbeddings,
+                embeddingProvider: realEmbeddings ? `@claude-flow/embeddings (${embeddingServiceName})` : 'hash-based (deterministic)',
+                models: {
+                    total: models.length,
+                    ready: models.filter(m => m.status === 'ready').length,
+                    training: models.filter(m => m.status === 'training').length,
+                    avgAccuracy: models.length > 0
+                        ? models.reduce((sum, m) => sum + m.accuracy, 0) / models.length
+                        : 0,
+                },
+                patterns: {
+                    total: patterns.length,
+                    byType: patterns.reduce((acc, p) => {
+                        acc[p.type] = (acc[p.type] || 0) + 1;
+                        return acc;
+                    }, {}),
+                    totalEmbeddingDims: patterns.length > 0 ? patterns[0].embedding.length : 384,
+                },
+                features: {
+                    hnsw: true,
+                    quantization: true,
+                    // #1770: probe the real loader instead of returning a literal false.
+                    // Was hardcoded false, which contradicted hooks_intelligence_stats's
+                    // simultaneous claim of `implementation: real-flash-attention`.
+                    // The two surfaces now agree on a single source of truth.
+                    flashAttention: await (async () => {
+                        try {
+                            // #1773 item 4 — flash-attention now lives in @claude-flow/neural
+                            const { getFlashAttention } = await import('@claude-flow/neural');
+                            return getFlashAttention() !== null;
+                        }
+                        catch {
+                            return false;
+                        }
+                    })(),
+                    reasoningBank: true,
+                },
+            };
+        },
+    },
+    {
+        name: 'neural_optimize',
+        description: 'Optimize neural model performance Use when nothing native trains on your workflow — Claude Code has no learning loop. Use to train SONA/MoE/EWC patterns from successful task outcomes; query via neural_predict before spawning agents. Off-path for one-shot work.',
+        category: 'neural',
+        inputSchema: {
+            type: 'object',
+            properties: {
+                modelId: { type: 'string', description: 'Model ID to optimize' },
+                target: { type: 'string', enum: ['speed', 'memory', 'accuracy', 'balanced'], description: 'Optimization target' },
+            },
+        },
+        handler: async (input) => {
+            if (input.modelId) {
+                const v = validateIdentifier(input.modelId, 'modelId');
+                if (!v.valid)
+                    return { success: false, error: v.error };
+            }
+            // Resolve provider so embeddingProvider in the result matches the old
+            // eager-init reporting.
+            await ensureEmbeddings();
+            const store = loadNeuralStore();
+            const target = input.target || 'balanced';
+            const patterns = Object.values(store.patterns);
+            if (patterns.length === 0) {
+                return { success: false, error: 'No patterns to optimize. Train patterns first with neural_train.' };
+            }
+            const startTime = performance.now();
+            const actions = [];
+            const beforeCount = patterns.length;
+            const dims = patterns[0]?.embedding?.length || 0;
+            let patternsRemoved = 0;
+            let patternsQuantized = 0;
+            let duplicatesRemoved = 0;
+            // speed / balanced: deduplicate identical or near-identical patterns
+            if (target === 'speed' || target === 'balanced') {
+                const seen = new Map(); // hash -> id
+                for (const [id, p] of Object.entries(store.patterns)) {
+                    if (!p.embedding || p.embedding.length === 0)
+                        continue;
+                    // Quick hash: first 8 dims rounded
+                    const hash = p.embedding.slice(0, 8).map(v => v.toFixed(4)).join(',');
+                    if (seen.has(hash)) {
+                        // Verify with full cosine similarity
+                        const existingId = seen.get(hash);
+                        const existing = store.patterns[existingId];
+                        if (existing && cosineSimilarity(p.embedding, existing.embedding) > 0.99) {
+                            existing.usageCount = Math.max(existing.usageCount || 0, p.usageCount || 0);
+                            delete store.patterns[id];
+                            duplicatesRemoved++;
+                        }
+                    }
+                    else {
+                        seen.set(hash, id);
+                    }
+                }
+                if (duplicatesRemoved > 0)
+                    actions.push(`Removed ${duplicatesRemoved} near-duplicate patterns`);
+            }
+            // memory / balanced: quantize large embeddings
+            if (target === 'memory' || target === 'balanced') {
+                try {
+                    const { quantizeInt8, getQuantizationStats } = await import('../memory/memory-initializer.js');
+                    for (const p of Object.values(store.patterns)) {
+                        if (p.embedding && p.embedding.length > 0 && !p._quantized) {
+                            const stats = getQuantizationStats(p.embedding);
+                            const q = quantizeInt8(p.embedding);
+                            p._quantized = { scale: q.scale, zeroPoint: q.zeroPoint, compressionRatio: stats.compressionRatio };
+                            patternsQuantized++;
+                        }
+                    }
+                    if (patternsQuantized > 0)
+                        actions.push(`Quantized ${patternsQuantized} pattern embeddings (Int8, ~3.92x)`);
+                }
+                catch {
+                    actions.push('Quantization skipped (memory-initializer not available)');
+                }
+            }
+            // accuracy / balanced: prune low-usage, zero-embedding patterns
+            if (target === 'accuracy' || target === 'balanced') {
+                for (const [id, p] of Object.entries(store.patterns)) {
+                    if (!p.embedding || p.embedding.length === 0) {
+                        delete store.patterns[id];
+                        patternsRemoved++;
+                        continue;
+                    }
+                    // Remove patterns with all-zero embeddings (no useful signal)
+                    const norm = p.embedding.reduce((s, v) => s + v * v, 0);
+                    if (norm < 1e-10) {
+                        delete store.patterns[id];
+                        patternsRemoved++;
+                    }
+                }
+                if (patternsRemoved > 0)
+                    actions.push(`Pruned ${patternsRemoved} empty/zero-signal patterns`);
+            }
+            saveNeuralStore(store);
+            const elapsed = Math.round(performance.now() - startTime);
+            return {
+                success: true, _real: true, target,
+                embeddingProvider: embeddingServiceName,
+                actions,
+                patternsBefore: beforeCount,
+                patternsAfter: Object.keys(store.patterns).length,
+                duplicatesRemoved,
+                patternsQuantized,
+                patternsRemoved,
+                embeddingDims: dims,
+                elapsedMs: elapsed,
+            };
+        },
+    },
+];
+//# sourceMappingURL=neural-tools.js.map