@jigyasudham/veto 3.0.0 → 3.1.1

package/dist/router/learning-updater.js

@@ -5,7 +5,7 @@ import { getConfig } from '../memory/config.js';
 // ─── Record ───────────────────────────────────────────────────────────────────
 export function recordOutcome(taskType, complexity, modelTier, agent, outputQuality, tokensUsed = 0, fileExt) {
     const db = getDb();
-    db.prepare(`INSERT INTO learning_data (id, task_type, complexity, model_tier, output_quality, tokens_used, agent)
+    db.prepare(`INSERT INTO learning_data (id, task_type, complexity, model_tier, output_quality, tokens_used, agent)
      VALUES (?, ?, ?, ?, ?, ?, ?)`).run(randomUUID(), taskType, complexity, modelTier, outputQuality, tokensUsed, agent);
     // Predictive routing: track agent quality per file extension and task type keyword
     if (agent && agent !== 'dynamic' && outputQuality > 0) {
@@ -50,8 +50,8 @@ export function getRecommendedAgent(taskType, fileExt) {
     const db = getDb();
     // Try file extension match first (more specific)
     if (fileExt) {
-        const rows = db.prepare(`SELECT pattern_key, pattern_val, confidence, seen_count FROM patterns
-       WHERE pattern_key LIKE ? AND seen_count >= 3
+        const rows = db.prepare(`SELECT pattern_key, pattern_val, confidence, seen_count FROM patterns
+       WHERE pattern_key LIKE ? AND seen_count >= 3
        ORDER BY confidence DESC LIMIT 1`).all(`file_agent:${fileExt.toLowerCase()}:%`);
         if (rows.length > 0 && rows[0].confidence >= 0.7)
             return rows[0].pattern_val;
@@ -59,8 +59,8 @@ export function getRecommendedAgent(taskType, fileExt) {
     // Fall back to task keyword match
     const keyword = taskType.toLowerCase().split(/[\s_-]/)[0];
     if (keyword) {
-        const rows = db.prepare(`SELECT pattern_key, pattern_val, confidence, seen_count FROM patterns
-       WHERE pattern_key LIKE ? AND seen_count >= 3
+        const rows = db.prepare(`SELECT pattern_key, pattern_val, confidence, seen_count FROM patterns
+       WHERE pattern_key LIKE ? AND seen_count >= 3
        ORDER BY confidence DESC LIMIT 1`).all(`task_agent:${keyword}:%`);
         if (rows.length > 0 && rows[0].confidence >= 0.7)
             return rows[0].pattern_val;
@@ -70,7 +70,7 @@ export function getRecommendedAgent(taskType, fileExt) {
 // ─── Stats ────────────────────────────────────────────────────────────────────
 export function getLearningStats() {
     const db = getDb();
-    const rows = db.prepare(`SELECT model_tier, COUNT(*) as count, AVG(output_quality) as avg_quality
+    const rows = db.prepare(`SELECT model_tier, COUNT(*) as count, AVG(output_quality) as avg_quality
      FROM learning_data GROUP BY model_tier`).all();
     const total_tasks = rows.reduce((sum, r) => sum + r.count, 0);
     const tier_breakdown = {};
@@ -85,16 +85,16 @@ export function getLearningStats() {
 }
 export function getAgentPerformanceStats() {
     const db = getDb();
-    const rows = db.prepare(`
-    SELECT agent,
-           COUNT(*) as task_count,
-           AVG(output_quality) as avg_quality,
-           AVG(tokens_used) as avg_tokens,
-           SUM(CASE WHEN output_quality < 60 THEN 1 ELSE 0 END) as low_quality_count
-    FROM learning_data
-    WHERE agent IS NOT NULL
-    GROUP BY agent
-    ORDER BY avg_quality ASC
+    const rows = db.prepare(`
+    SELECT agent,
+           COUNT(*) as task_count,
+           AVG(output_quality) as avg_quality,
+           AVG(tokens_used) as avg_tokens,
+           SUM(CASE WHEN output_quality < 60 THEN 1 ELSE 0 END) as low_quality_count
+    FROM learning_data
+    WHERE agent IS NOT NULL
+    GROUP BY agent
+    ORDER BY avg_quality ASC
   `).all();
     return rows.map(r => ({
         agent: r.agent,
@@ -107,23 +107,23 @@ export function getAgentPerformanceStats() {
 }
 export function getTaskTypeBreakdown() {
     const db = getDb();
-    const rows = db.prepare(`
-    SELECT task_type,
-           COUNT(*) as count,
-           AVG(complexity) as avg_complexity,
-           AVG(output_quality) as avg_quality,
-           model_tier as dominant_tier
-    FROM learning_data
-    GROUP BY task_type, model_tier
-    HAVING COUNT(*) = (
-      SELECT MAX(c) FROM (
-        SELECT COUNT(*) as c FROM learning_data l2
-        WHERE l2.task_type = learning_data.task_type
-        GROUP BY l2.model_tier
-      )
-    )
-    ORDER BY count DESC
-    LIMIT 50
+    const rows = db.prepare(`
+    SELECT task_type,
+           COUNT(*) as count,
+           AVG(complexity) as avg_complexity,
+           AVG(output_quality) as avg_quality,
+           model_tier as dominant_tier
+    FROM learning_data
+    GROUP BY task_type, model_tier
+    HAVING COUNT(*) = (
+      SELECT MAX(c) FROM (
+        SELECT COUNT(*) as c FROM learning_data l2
+        WHERE l2.task_type = learning_data.task_type
+        GROUP BY l2.model_tier
+      )
+    )
+    ORDER BY count DESC
+    LIMIT 50
   `).all();
     return rows.map(r => {
         // Under-tiered: task scored high quality but at a tier below what complexity suggests
@@ -142,15 +142,15 @@ export function getTaskTypeBreakdown() {
 export function getCouncilInsights() {
     const db = getDb();
     // Correlate council approvals (GREEN/YELLOW) with debugging sessions that came after
-    const rows = db.prepare(`
-    SELECT co.verdict, co.task, co.debated_at,
-           d.decision as related_decision, d.made_at as decision_at
-    FROM council_outcomes co
-    LEFT JOIN decisions d ON d.session_id = co.session_id
-      AND d.decision LIKE '%debug%' OR d.decision LIKE '%fix%' OR d.decision LIKE '%bug%'
-    WHERE co.verdict IN ('GREEN', 'YELLOW')
-    ORDER BY co.debated_at DESC
-    LIMIT 30
+    const rows = db.prepare(`
+    SELECT co.verdict, co.task, co.debated_at,
+           d.decision as related_decision, d.made_at as decision_at
+    FROM council_outcomes co
+    LEFT JOIN decisions d ON d.session_id = co.session_id
+      AND d.decision LIKE '%debug%' OR d.decision LIKE '%fix%' OR d.decision LIKE '%bug%'
+    WHERE co.verdict IN ('GREEN', 'YELLOW')
+    ORDER BY co.debated_at DESC
+    LIMIT 30
   `).all();
     return rows.map(r => {
         let days_to_debug = null;
@@ -273,9 +273,9 @@ export function recordRoutingFeedback(opts) {
     const now = new Date().toISOString();
     const expiresAt = new Date(Date.now() + FEEDBACK_TTL_DAYS * 86_400_000).toISOString();
     const task_hash = opts.task.toLowerCase().replace(/\s+/g, '_').slice(0, 40);
-    db.prepare(`
-    INSERT INTO routing_feedback (id, task_hash, task_snippet, complexity, model_tier, agent, outcome, session_id, recorded_at, expires_at)
-    VALUES (?, ?, ?, ?, ?, ?, 'pending', ?, ?, ?)
+    db.prepare(`
+    INSERT INTO routing_feedback (id, task_hash, task_snippet, complexity, model_tier, agent, outcome, session_id, recorded_at, expires_at)
+    VALUES (?, ?, ?, ?, ?, ?, 'pending', ?, ?, ?)
   `).run(id, task_hash, opts.task.slice(0, 100), opts.complexity, opts.model_tier, opts.agent ?? null, opts.session_id ?? null, now, expiresAt);
     return id;
 }

package/dist/server.js

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-// Veto MCP Server — 92 tools, LLM council + auto-learning router
+// Veto MCP Server — 93 tools, LLM council + auto-learning router
 // Suppress node:sqlite experimental warning — it would corrupt the MCP stdio protocol
 process.removeAllListeners('warning');
 import { Server } from '@modelcontextprotocol/sdk/server/index.js';

package/dist/tools/compact.js

@@ -1,5 +1,5 @@
-// Compact tool surface — progressive disclosure for clients where 92 tool
-// schemas would burn ~25K context tokens before the user types a word.
+// Compact tool surface — progressive disclosure for clients where 93 tool
+// schemas would burn ~20K context tokens before the user types a word.
 //
 // In compact mode (VETO_COMPACT=1 or config compact_tools=true) ListTools
 // exposes only a small first-class set plus two meta-tools: veto_find_tools
@@ -22,7 +22,7 @@ const FIRST_CLASS = [
 export const META_TOOL_DEFINITIONS = [
     {
         name: 'veto_find_tools',
-        description: 'Searches the full Veto catalog (92 tools: agents, reviews, advisors, generators, memory, learning, git, workflow) by keyword and returns matching tool schemas. Compact mode exposes only core tools up front — call this first to discover the right tool, then invoke it via veto_call.',
+        description: 'Searches the full Veto catalog (93 tools: agents, reviews, advisors, generators, memory, learning, git, workflow) by keyword and returns matching tool schemas. Compact mode exposes only core tools up front — call this first to discover the right tool, then invoke it via veto_call.',
         inputSchema: {
             type: 'object',
             properties: {

package/dist/transcripts/archive.js

@@ -79,14 +79,14 @@ export async function captureSession(opts = {}) {
         if (existing) {
             // Content changed (grew): overwrite the row, reset the index watermark AND
             // parser_version so Step 6 re-derives events from the new bytes.
-            db.prepare(`UPDATE archives SET archive_path=?, content_sha256=?, source_bytes=?, archive_bytes=?,
-           project_dir=COALESCE(?, project_dir), veto_session_id=COALESCE(?, veto_session_id),
+            db.prepare(`UPDATE archives SET archive_path=?, content_sha256=?, source_bytes=?, archive_bytes=?,
+           project_dir=COALESCE(?, project_dir), veto_session_id=COALESCE(?, veto_session_id),
            updated_at=?, indexed_through_seq=0, parser_version=0 WHERE id=?`).run(archivePath, sha256, bytes, archiveBytes, proj, opts.vetoSessionId ?? null, now, id);
         }
         else {
-            db.prepare(`INSERT INTO archives (id, source, source_session_id, project_dir, veto_session_id, archive_path,
-           content_sha256, source_bytes, archive_bytes, source_format_hint, parser_version, indexed_through_seq,
-           captured_at, updated_at)
+            db.prepare(`INSERT INTO archives (id, source, source_session_id, project_dir, veto_session_id, archive_path,
+           content_sha256, source_bytes, archive_bytes, source_format_hint, parser_version, indexed_through_seq,
+           captured_at, updated_at)
          VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, 0, 0, ?, ?)`).run(id, source, sessionId, proj, opts.vetoSessionId ?? null, archivePath, sha256, bytes, archiveBytes, formatHint, now, now);
         }
         return {

package/dist/transcripts/cache.d.ts

@@ -0,0 +1,66 @@
+export type DocCache = {
+    /** Highest doc id; arrays below are indexed by doc id, so length is maxId+1. */
+    maxId: number;
+    /** Token count per doc, for BM25 length normalization. */
+    len: Int32Array;
+    project: (string | null)[];
+    session: string[];
+    /** Doc count and average length, for BM25's corpus statistics. */
+    count: number;
+    avgdl: number;
+};
+/**
+ * How many leading dimensions the cheap first pass scores.
+ *
+ * potion applies PCA, so dimensions arrive ordered by variance and a prefix is
+ * unusually informative. Scoring a prefix costs a quarter of a full dot
+ * product and, combined with the tail norm, yields an exact upper bound via
+ * Cauchy-Schwarz: |tail contribution| <= ||q_tail|| * ||d_tail||. Chunks whose
+ * BEST possible score cannot reach the current cutoff are skipped without ever
+ * being fully scored — exact, never a recall loss, unlike an ANN index.
+ */
+export declare const PREFIX_DIMS = 64;
+export type VectorCache = {
+    dim: number;
+    count: number;
+    /** count * dim, row-major — one contiguous allocation, not one per chunk. */
+    data: Int8Array;
+    /** L2 norm of each int8 row. */
+    norms: Float32Array;
+    /** L2 norm of dims [PREFIX_DIMS, dim) per row — the tail a prefix scan has
+     *  not yet seen. Enables an exact Cauchy-Schwarz bound (see PREFIX_DIMS). */
+    tailNorms: Float32Array;
+    docIds: Int32Array;
+    chunkIdx: Int32Array;
+};
+/** Drop both caches. MUST be called after anything that mutates the index. */
+export declare function resetCaches(): void;
+export declare function getDocCache(): DocCache;
+export declare function getVectorCache(): VectorCache;
+/**
+ * A lookup table of in-scope doc ids, built from the cached metadata rather
+ * than a per-query SQL round trip. Returns null when no filter applies.
+ *
+ * Memoized per scope: building it walks every doc id doing string comparisons,
+ * which at 41k docs is a real per-query cost, and consecutive queries almost
+ * always share a scope. Cleared with the caches it is derived from.
+ */
+export declare function scopeMask(docs: DocCache, opts: {
+    projectDir?: string;
+    sourceSessionId?: string;
+}): Uint8Array | null;
+/**
+ * Reusable per-document accumulators for the vector scan.
+ *
+ * Allocating these per query costs hundreds of kilobytes of zero-fill on every
+ * search. They are handed back dirty by contract — the caller resets only the
+ * entries it touched, which is proportional to matches rather than to the id
+ * space.
+ */
+export declare function scoreScratch(maxId: number): {
+    score: Float32Array;
+    chunk: Int32Array;
+};
+/** Sentinel for "this document has not been scored in this query". */
+export declare const EMPTY_SCORE = -2;
+//# sourceMappingURL=cache.d.ts.map

package/dist/transcripts/cache.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"cache.d.ts","sourceRoot":"","sources":["../../src/transcripts/cache.ts"],"names":[],"mappings":"AAmBA,MAAM,MAAM,QAAQ,GAAG;IACrB,gFAAgF;IAChF,KAAK,EAAE,MAAM,CAAC;IACd,0DAA0D;IAC1D,GAAG,EAAE,UAAU,CAAC;IAChB,OAAO,EAAE,CAAC,MAAM,GAAG,IAAI,CAAC,EAAE,CAAC;IAC3B,OAAO,EAAE,MAAM,EAAE,CAAC;IAClB,kEAAkE;IAClE,KAAK,EAAE,MAAM,CAAC;IACd,KAAK,EAAE,MAAM,CAAC;CACf,CAAC;AAEF;;;;;;;;;GASG;AACH,eAAO,MAAM,WAAW,KAAK,CAAC;AAE9B,MAAM,MAAM,WAAW,GAAG;IACxB,GAAG,EAAE,MAAM,CAAC;IACZ,KAAK,EAAE,MAAM,CAAC;IACd,6EAA6E;IAC7E,IAAI,EAAE,SAAS,CAAC;IAChB,gCAAgC;IAChC,KAAK,EAAE,YAAY,CAAC;IACpB;iFAC6E;IAC7E,SAAS,EAAE,YAAY,CAAC;IACxB,MAAM,EAAE,UAAU,CAAC;IACnB,QAAQ,EAAE,UAAU,CAAC;CACtB,CAAC;AAKF,8EAA8E;AAC9E,wBAAgB,WAAW,IAAI,IAAI,CAMlC;AAED,wBAAgB,WAAW,IAAI,QAAQ,CA8BtC;AAED,wBAAgB,cAAc,IAAI,WAAW,CAkC5C;AAID;;;;;;;GAOG;AACH,wBAAgB,SAAS,CACvB,IAAI,EAAE,QAAQ,EACd,IAAI,EAAE;IAAE,UAAU,CAAC,EAAE,MAAM,CAAC;IAAC,eAAe,CAAC,EAAE,MAAM,CAAA;CAAE,GACtD,UAAU,GAAG,IAAI,CAcnB;AAKD;;;;;;;GAOG;AACH,wBAAgB,YAAY,CAAC,KAAK,EAAE,MAAM,GAAG;IAAE,KAAK,EAAE,YAAY,CAAC;IAAC,KAAK,EAAE,UAAU,CAAA;CAAE,CAMtF;AAED,sEAAsE;AACtE,eAAO,MAAM,WAAW,KAAK,CAAC"}

package/dist/transcripts/cache.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"cache.js","sourceRoot":"","sources":["../../src/transcripts/cache.ts"],"names":[],"mappings":"AAAA,8EAA8E;AAC9E,EAAE;AACF,4EAA4E;AAC5E,6EAA6E;AAC7E,6EAA6E;AAC7E,6EAA6E;AAC7E,yEAAyE;AACzE,6EAA6E;AAC7E,EAAE;AACF,mEAAmE;AACnE,yEAAyE;AACzE,yEAAyE;AACzE,uEAAuE;AACvE,EAAE;AACF,8EAA8E;AAC9E,iEAAiE;AAEjE,OAAO,EAAE,gBAAgB,EAAE,MAAM,YAAY,CAAC;AAc9C;;;;;;;;;GASG;AACH,MAAM,CAAC,MAAM,WAAW,GAAG,EAAE,CAAC;AAgB9B,IAAI,KAA2B,CAAC;AAChC,IAAI,QAAiC,CAAC;AAEtC,8EAA8E;AAC9E,MAAM,UAAU,WAAW;IACzB,KAAK,GAAG,SAAS,CAAC;IAClB,QAAQ,GAAG,SAAS,CAAC;IACrB,MAAM,CAAC,KAAK,EAAE,CAAC;IACf,aAAa,GAAG,SAAS,CAAC;IAC1B,aAAa,GAAG,SAAS,CAAC;AAC5B,CAAC;AAED,MAAM,UAAU,WAAW;IACzB,IAAI,KAAK;QAAE,OAAO,KAAK,CAAC;IACxB,MAAM,EAAE,GAAG,gBAAgB,EAAE,CAAC;IAC9B,MAAM,IAAI,GAAG,EAAE,CAAC,OAAO,CACrB,iEAAiE,CAClE,CAAC,GAAG,EAA0F,CAAC;IAEhG,IAAI,KAAK,GAAG,CAAC,CAAC;IACd,KAAK,MAAM,CAAC,IAAI,IAAI;QAAE,IAAI,CAAC,CAAC,EAAE,GAAG,KAAK;YAAE,KAAK,GAAG,CAAC,CAAC,EAAE,CAAC;IAErD,MAAM,GAAG,GAAG,IAAI,UAAU,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC;IACtC,MAAM,OAAO,GAAsB,IAAI,KAAK,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IACnE,MAAM,OAAO,GAAa,IAAI,KAAK,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;IACxD,IAAI,KAAK,GAAG,CAAC,CAAC;IACd,KAAK,MAAM,CAAC,IAAI,IAAI,EAAE,CAAC;QACrB,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC;QAClB,OAAO,CAAC,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,WAAW,CAAC;QAC9B,OAAO,CAAC,CAAC,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC,iBAAiB,CAAC;QACpC,KAAK,IAAI,CAAC,CAAC,GAAG,CAAC;IACjB,CAAC;IAED,KAAK,GAAG;QACN,KAAK;QACL,GAAG;QACH,OAAO;QACP,OAAO;QACP,KAAK,EAAE,IAAI,CAAC,MAAM;QAClB,KAAK,EAAE,IAAI,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,GAAG,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC;KACjD,CAAC;IACF,OAAO,KAAK,CAAC;AACf,CAAC;AAED,MAAM,UAAU,cAAc;IAC5B,IAAI,QAAQ;QAAE,OAAO,QAAQ,CAAC;IAC9B,MAAM,EAAE,GAAG,gBAAgB,EAAE,CAAC;IAC9B,MAAM,IAAI,GAAG,EAAE,CAAC,OAAO,CACrB,2DAA2D,CAC5D,CAAC,GAAG,EAA8E,CAAC;IAEpF,MAAM,GAAG,GAAG,IAAI,CAAC,CAAC,CAAC,EAAE,GAAG,CAAC,MAAM,IAAI,CAAC,CAAC;IACrC,MAAM,KAAK,GAAG,IAAI,CAAC,MAAM,CAAC;IAC1B,MAAM,IAAI,GAAG,IAAI,SAAS,CAAC,KAAK,GAAG,GAAG,CAAC,CAAC;IACxC,MAAM,KAAK,GAAG,IAAI,YAAY,CAAC,KAAK,CAAC,CAAC;IACtC,MAAM,SAAS,GAAG,IAAI,YAAY,CAAC,KAAK,CAAC,CAAC;IAC1C,MAAM,MAAM,GAAG,IAAI,UAAU,CAAC,KAAK,CAAC,CAAC;IACrC,MAAM,QAAQ,GAAG,IAAI,UAAU,CAAC,KAAK,CAAC,CAAC;IAEvC,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,KAAK,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC;QAClC,MAAM,CAAC,GAAG,IAAI,CAAC,CAAC,CAAE,CAAC;QACnB,0EAA0E;QAC1E,6CAA6C;QAC7C,IAAI,CAAC,CAAC,GAAG,CAAC,MAAM,KAAK,GAAG;YAAE,SAAS;QACnC,IAAI,CAAC,GAAG,CAAC,IAAI,SAAS,CAAC,CAAC,CAAC,GAAG,CAAC,MAAM,EAAE,CAAC,CAAC,GAAG,CAAC,UAAU,EAAE,GAAG,CAAC,EAAE,CAAC,GAAG,GAAG,CAAC,CAAC;QACtE,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC;QAClB,MAAM,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,MAAM,CAAC;QACrB,QAAQ,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,WAAW,CAAC;QAC5B,IAAI,IAAI,GAAG,CAAC,CAAC;QACb,KAAK,IAAI,CAAC,GAAG,WAAW,EAAE,CAAC,GAAG,GAAG,EAAE,CAAC,IAAI,CAAC,EAAE,CAAC;YAC1C,MAAM,CAAC,GAAG,IAAI,CAAC,CAAC,GAAG,GAAG,GAAG,CAAC,CAAE,CAAC;YAC7B,IAAI,IAAI,CAAC,GAAG,CAAC,CAAC;QAChB,CAAC;QACD,SAAS,CAAC,CAAC,CAAC,GAAG,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IACjC,CAAC;IAED,QAAQ,GAAG,EAAE,GAAG,EAAE,KAAK,EAAE,IAAI,EAAE,KAAK,EAAE,SAAS,EAAE,MAAM,EAAE,QAAQ,EAAE,CAAC;IACpE,OAAO,QAAQ,CAAC;AAClB,CAAC;AAED,MAAM,MAAM,GAAG,IAAI,GAAG,EAAsB,CAAC;AAE7C;;;;;;;GAOG;AACH,MAAM,UAAU,SAAS,CACvB,IAAc,EACd,IAAuD;IAEvD,IAAI,CAAC,IAAI,CAAC,UAAU,IAAI,CAAC,IAAI,CAAC,eAAe;QAAE,OAAO,IAAI,CAAC;IAC3D,MAAM,GAAG,GAAG,GAAG,IAAI,CAAC,UAAU,IAAI,EAAE,IAAI,IAAI,CAAC,eAAe,IAAI,EAAE,EAAE,CAAC;IACrE,MAAM,GAAG,GAAG,MAAM,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;IAC5B,IAAI,GAAG;QAAE,OAAO,GAAG,CAAC;IAEpB,MAAM,IAAI,GAAG,IAAI,UAAU,CAAC,IAAI,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC;IAC5C,KAAK,IAAI,EAAE,GAAG,CAAC,EAAE,EAAE,IAAI,IAAI,CAAC,KAAK,EAAE,EAAE,IAAI,CAAC,EAAE,CAAC;QAC3C,IAAI,IAAI,CAAC,UAAU,IAAI,IAAI,CAAC,OAAO,CAAC,EAAE,CAAC,KAAK,IAAI,CAAC,UAAU;YAAE,SAAS;QACtE,IAAI,IAAI,CAAC,eAAe,IAAI,IAAI,CAAC,OAAO,CAAC,EAAE,CAAC,KAAK,IAAI,CAAC,eAAe;YAAE,SAAS;QAChF,IAAI,CAAC,EAAE,CAAC,GAAG,CAAC,CAAC;IACf,CAAC;IACD,MAAM,CAAC,GAAG,CAAC,GAAG,EAAE,IAAI,CAAC,CAAC;IACtB,OAAO,IAAI,CAAC;AACd,CAAC;AAED,IAAI,aAAuC,CAAC;AAC5C,IAAI,aAAqC,CAAC;AAE1C;;;;;;;GAOG;AACH,MAAM,UAAU,YAAY,CAAC,KAAa;IACxC,IAAI,CAAC,aAAa,IAAI,aAAa,CAAC,MAAM,GAAG,KAAK,GAAG,CAAC,EAAE,CAAC;QACvD,aAAa,GAAG,IAAI,YAAY,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC,IAAI,CAAC,WAAW,CAAC,CAAC;QAC9D,aAAa,GAAG,IAAI,UAAU,CAAC,KAAK,GAAG,CAAC,CAAC,CAAC;IAC5C,CAAC;IACD,OAAO,EAAE,KAAK,EAAE,aAAa,EAAE,KAAK,EAAE,aAAc,EAAE,CAAC;AACzD,CAAC;AAED,sEAAsE;AACtE,MAAM,CAAC,MAAM,WAAW,GAAG,CAAC,CAAC,CAAC"}

package/dist/transcripts/chunk.d.ts

@@ -0,0 +1,26 @@
+/** Bump on ANY boundary change — forces stored vectors to be re-derived. */
+export declare const CHUNKER_VERSION = 1;
+/** Validated by the paraphrase-gate prototype. Not a tuning knob: changing
+ *  these invalidates the experiment that justified the semantic layer. */
+export declare const WINDOW_CHARS = 240;
+export declare const STRIDE_CHARS = 120;
+/** Input-size guard, not a tuning knob (C2). A pathological event must not be
+ *  able to emit unbounded rows; the tail is dropped and stays lexically
+ *  searchable via BM25, which indexes the whole text regardless. */
+export declare const MAX_WINDOWS_PER_EVENT = 64;
+export type Chunk = {
+    index: number;
+    /** UTF-16 offsets into the source text, so a hit can aim a snippet (C6). */
+    start: number;
+    end: number;
+    text: string;
+};
+/**
+ * Split `text` into overlapping windows.
+ *
+ * Events at or under the window emit exactly one chunk covering the whole
+ * text — the median event is ~205 chars, so most events produce a single row
+ * and the vector count stays proportional to content, not to event count.
+ */
+export declare function chunkText(text: string): Chunk[];
+//# sourceMappingURL=chunk.d.ts.map

package/dist/transcripts/chunk.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"chunk.d.ts","sourceRoot":"","sources":["../../src/transcripts/chunk.ts"],"names":[],"mappings":"AAYA,4EAA4E;AAC5E,eAAO,MAAM,eAAe,IAAI,CAAC;AAEjC;0EAC0E;AAC1E,eAAO,MAAM,YAAY,MAAM,CAAC;AAChC,eAAO,MAAM,YAAY,MAAM,CAAC;AAEhC;;oEAEoE;AACpE,eAAO,MAAM,qBAAqB,KAAK,CAAC;AAExC,MAAM,MAAM,KAAK,GAAG;IAClB,KAAK,EAAE,MAAM,CAAC;IACd,4EAA4E;IAC5E,KAAK,EAAE,MAAM,CAAC;IACd,GAAG,EAAE,MAAM,CAAC;IACZ,IAAI,EAAE,MAAM,CAAC;CACd,CAAC;AAWF;;;;;;GAMG;AACH,wBAAgB,SAAS,CAAC,IAAI,EAAE,MAAM,GAAG,KAAK,EAAE,CAc/C"}

package/dist/transcripts/chunk.js

@@ -0,0 +1,54 @@
+// Chunking for the semantic layer (VERSION-3 item 6, Phase B).
+//
+// One vector per event demonstrably FAILS the paraphrase gate: mean-pooling a
+// 1,936-char event dilutes the one sentence that answers the question down to
+// ~5% of the vector, and it never surfaces. Splitting events into overlapping
+// windows and scoring each event by its BEST window fixed that — the whole
+// reason this file exists (council outcome 9b0595ad, spec §11 amendment).
+//
+// Boundaries here are a compatibility surface: they are pinned by golden tests
+// (C1), and CHUNKER_VERSION must be bumped if any of it changes, because
+// stored vectors are only meaningful against the chunker that produced them.
+/** Bump on ANY boundary change — forces stored vectors to be re-derived. */
+export const CHUNKER_VERSION = 1;
+/** Validated by the paraphrase-gate prototype. Not a tuning knob: changing
+ *  these invalidates the experiment that justified the semantic layer. */
+export const WINDOW_CHARS = 240;
+export const STRIDE_CHARS = 120;
+/** Input-size guard, not a tuning knob (C2). A pathological event must not be
+ *  able to emit unbounded rows; the tail is dropped and stays lexically
+ *  searchable via BM25, which indexes the whole text regardless. */
+export const MAX_WINDOWS_PER_EVENT = 64;
+/** Never split a surrogate pair: a lone surrogate would corrupt normalization
+ *  and produce a vector for text that does not exist. Boundaries move forward
+ *  off a low surrogate, which keeps the result deterministic. */
+function safeBoundary(text, i) {
+    if (i <= 0 || i >= text.length)
+        return i;
+    const c = text.charCodeAt(i);
+    return c >= 0xdc00 && c <= 0xdfff ? i + 1 : i;
+}
+/**
+ * Split `text` into overlapping windows.
+ *
+ * Events at or under the window emit exactly one chunk covering the whole
+ * text — the median event is ~205 chars, so most events produce a single row
+ * and the vector count stays proportional to content, not to event count.
+ */
+export function chunkText(text) {
+    if (text.length === 0)
+        return [];
+    if (text.length <= WINDOW_CHARS) {
+        return [{ index: 0, start: 0, end: text.length, text }];
+    }
+    const out = [];
+    for (let raw = 0; raw < text.length && out.length < MAX_WINDOWS_PER_EVENT; raw += STRIDE_CHARS) {
+        const start = safeBoundary(text, raw);
+        const end = safeBoundary(text, Math.min(start + WINDOW_CHARS, text.length));
+        out.push({ index: out.length, start, end, text: text.slice(start, end) });
+        if (end >= text.length)
+            break;
+    }
+    return out;
+}
+//# sourceMappingURL=chunk.js.map

package/dist/transcripts/chunk.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"chunk.js","sourceRoot":"","sources":["../../src/transcripts/chunk.ts"],"names":[],"mappings":"AAAA,+DAA+D;AAC/D,EAAE;AACF,8EAA8E;AAC9E,8EAA8E;AAC9E,8EAA8E;AAC9E,2EAA2E;AAC3E,0EAA0E;AAC1E,EAAE;AACF,+EAA+E;AAC/E,yEAAyE;AACzE,6EAA6E;AAE7E,4EAA4E;AAC5E,MAAM,CAAC,MAAM,eAAe,GAAG,CAAC,CAAC;AAEjC;0EAC0E;AAC1E,MAAM,CAAC,MAAM,YAAY,GAAG,GAAG,CAAC;AAChC,MAAM,CAAC,MAAM,YAAY,GAAG,GAAG,CAAC;AAEhC;;oEAEoE;AACpE,MAAM,CAAC,MAAM,qBAAqB,GAAG,EAAE,CAAC;AAUxC;;iEAEiE;AACjE,SAAS,YAAY,CAAC,IAAY,EAAE,CAAS;IAC3C,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,IAAI,CAAC,MAAM;QAAE,OAAO,CAAC,CAAC;IACzC,MAAM,CAAC,GAAG,IAAI,CAAC,UAAU,CAAC,CAAC,CAAC,CAAC;IAC7B,OAAO,CAAC,IAAI,MAAM,IAAI,CAAC,IAAI,MAAM,CAAC,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;AAChD,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,SAAS,CAAC,IAAY;IACpC,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,EAAE,CAAC;IACjC,IAAI,IAAI,CAAC,MAAM,IAAI,YAAY,EAAE,CAAC;QAChC,OAAO,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE,GAAG,EAAE,IAAI,CAAC,MAAM,EAAE,IAAI,EAAE,CAAC,CAAC;IAC1D,CAAC;IAED,MAAM,GAAG,GAAY,EAAE,CAAC;IACxB,KAAK,IAAI,GAAG,GAAG,CAAC,EAAE,GAAG,GAAG,IAAI,CAAC,MAAM,IAAI,GAAG,CAAC,MAAM,GAAG,qBAAqB,EAAE,GAAG,IAAI,YAAY,EAAE,CAAC;QAC/F,MAAM,KAAK,GAAG,YAAY,CAAC,IAAI,EAAE,GAAG,CAAC,CAAC;QACtC,MAAM,GAAG,GAAG,YAAY,CAAC,IAAI,EAAE,IAAI,CAAC,GAAG,CAAC,KAAK,GAAG,YAAY,EAAE,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC;QAC5E,GAAG,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,GAAG,CAAC,MAAM,EAAE,KAAK,EAAE,GAAG,EAAE,IAAI,EAAE,IAAI,CAAC,KAAK,CAAC,KAAK,EAAE,GAAG,CAAC,EAAE,CAAC,CAAC;QAC1E,IAAI,GAAG,IAAI,IAAI,CAAC,MAAM;YAAE,MAAM;IAChC,CAAC;IACD,OAAO,GAAG,CAAC;AACb,CAAC"}

package/dist/transcripts/embed.d.ts

@@ -0,0 +1,58 @@
+type Header = {
+    format: string;
+    upstream: {
+        model_id: string;
+        revision: string;
+    };
+    vocab_size: number;
+    dim: number;
+    inference: {
+        add_special_tokens: boolean;
+        pooling: string;
+        normalize: boolean;
+        per_token_weights: unknown;
+    };
+    tokenizer: {
+        strip_accents_effective: boolean;
+    };
+};
+type Model = {
+    header: Header;
+    dim: number;
+    /** Row-major int8 table, vocab_size * dim. */
+    table: Int8Array;
+    /** One scale per row: row_f32 = table[row] * scales[row]. */
+    scales: Float32Array;
+    vocab: Map<string, number>;
+    unkId: number | undefined;
+    prefix: string;
+    maxWordChars: number;
+    lowercase: boolean;
+    stripAccents: boolean;
+    cleanText: boolean;
+    handleChineseChars: boolean;
+};
+/** Token ids for `text`, with no special tokens added (the shipped contract). */
+export declare function tokenizeForEmbedding(text: string, model?: Model): number[];
+/**
+ * Embed `text` into a unit-length vector.
+ *
+ * Gather each token's row, dequantize it by that row's scale, mean-pool, then
+ * L2-normalize. Text with no tokens (empty or punctuation-free-of-vocab)
+ * yields the zero vector rather than NaN — upstream's behaviour, and the
+ * reason the epsilon is in the denominator.
+ */
+export declare function embed(text: string, model?: Model): Float32Array;
+/** Cosine similarity of two unit vectors — a dot product, given the inputs. */
+export declare function cosine(a: Float32Array, b: Float32Array): number;
+/** True when the model package resolves and its payload passes validation. */
+export declare function embeddingsAvailable(): boolean;
+/** Upstream model id + revision of the loaded payload, for recall provenance. */
+export declare function modelProvenance(): {
+    model_id: string;
+    revision: string;
+};
+/** Drop the cached table. Tests only — production loads once per process. */
+export declare function resetModelCache(): void;
+export {};
+//# sourceMappingURL=embed.d.ts.map

package/dist/transcripts/embed.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"embed.d.ts","sourceRoot":"","sources":["../../src/transcripts/embed.ts"],"names":[],"mappings":"AA6BA,KAAK,MAAM,GAAG;IACZ,MAAM,EAAE,MAAM,CAAC;IACf,QAAQ,EAAE;QAAE,QAAQ,EAAE,MAAM,CAAC;QAAC,QAAQ,EAAE,MAAM,CAAA;KAAE,CAAC;IACjD,UAAU,EAAE,MAAM,CAAC;IACnB,GAAG,EAAE,MAAM,CAAC;IACZ,SAAS,EAAE;QACT,kBAAkB,EAAE,OAAO,CAAC;QAC5B,OAAO,EAAE,MAAM,CAAC;QAChB,SAAS,EAAE,OAAO,CAAC;QACnB,iBAAiB,EAAE,OAAO,CAAC;KAC5B,CAAC;IACF,SAAS,EAAE;QAAE,uBAAuB,EAAE,OAAO,CAAA;KAAE,CAAC;CACjD,CAAC;AAEF,KAAK,KAAK,GAAG;IACX,MAAM,EAAE,MAAM,CAAC;IACf,GAAG,EAAE,MAAM,CAAC;IACZ,8CAA8C;IAC9C,KAAK,EAAE,SAAS,CAAC;IACjB,6DAA6D;IAC7D,MAAM,EAAE,YAAY,CAAC;IACrB,KAAK,EAAE,GAAG,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IAC3B,KAAK,EAAE,MAAM,GAAG,SAAS,CAAC;IAC1B,MAAM,EAAE,MAAM,CAAC;IACf,YAAY,EAAE,MAAM,CAAC;IACrB,SAAS,EAAE,OAAO,CAAC;IACnB,YAAY,EAAE,OAAO,CAAC;IACtB,SAAS,EAAE,OAAO,CAAC;IACnB,kBAAkB,EAAE,OAAO,CAAC;CAC7B,CAAC;AA4IF,iFAAiF;AACjF,wBAAgB,oBAAoB,CAAC,IAAI,EAAE,MAAM,EAAE,KAAK,QAAS,GAAG,MAAM,EAAE,CAU3E;AA4FD;;;;;;;GAOG;AACH,wBAAgB,KAAK,CAAC,IAAI,EAAE,MAAM,EAAE,KAAK,QAAS,GAAG,YAAY,CAqBhE;AAED,+EAA+E;AAC/E,wBAAgB,MAAM,CAAC,CAAC,EAAE,YAAY,EAAE,CAAC,EAAE,YAAY,GAAG,MAAM,CAI/D;AAED,8EAA8E;AAC9E,wBAAgB,mBAAmB,IAAI,OAAO,CAE7C;AAED,iFAAiF;AACjF,wBAAgB,eAAe,IAAI;IAAE,QAAQ,EAAE,MAAM,CAAC;IAAC,QAAQ,EAAE,MAAM,CAAA;CAAE,CAGxE;AAED,6EAA6E;AAC7E,wBAAgB,eAAe,IAAI,IAAI,CAEtC"}