@equationalapplications/core-llm-wiki 4.21.0 → 4.23.0

package/dist/index.d.mts

@@ -1,5 +1,5 @@
-import { M as MemoryBundle, F as FormatContextOptions, G as GraphNeighborhood, a as MemoryDump, b as FormattedMemoryDump, R as ReadOptions, S as SQLiteAdapter, W as WikiOptions, c as WikiMemory } from './testing-i91HR1TG.mjs';
-export { E as EntityStatus, d as ExtractedFact, e as ExtractedFactEdge, f as ExtractedFactWithOntology, g as ExtractedTask, h as GraphTraversalOptions, H as HOOK_TIMEOUT_MARKER, L as LLMProvider, O as ONTOLOGY_BACKFILL_BATCH_SIZE, i as ONTOLOGY_BACKFILL_MAX_PROMPT_CHARS, j as ONTOLOGY_BACKFILL_RECHECK_MS, k as OntologyBackfillResult, l as OntologyConfig, m as OntologyEdgeType, n as OntologyManifest, o as OntologyMode, p as OntologyNodeType, q as OntologyPromptContext, r as OntologyUpdates, P as PromptOverrides, s as PromptService, t as PrunePartialFailureError, V as VectorRanker, u as VectorRankerFallback, v as VectorRankerRankArgs, w as VectorRankerSemanticResult, x as WikiBusyError, y as WikiBusyOperation, z as WikiCheckpoint, A as WikiConfig, B as WikiEdge, C as WikiEvent, D as WikiFact, I as WikiMemoryTestAccess, J as WikiOutboxEvent, K as WikiTask, N as WikiTransactionError } from './testing-i91HR1TG.mjs';
+import { M as MemoryBundle, F as FormatContextOptions, G as GraphNeighborhood, a as MemoryDump, b as FormattedMemoryDump, O as OntologyManifest, R as ReadOptions, S as SQLiteAdapter, W as WikiOptions, c as WikiMemory } from './testing-CBjAuTSl.mjs';
+export { E as EntityStatus, d as ExtractedFact, e as ExtractedFactEdge, f as ExtractedFactWithOntology, g as ExtractedTask, h as GraphTraversalOptions, H as HOOK_TIMEOUT_MARKER, L as LLMProvider, i as ONTOLOGY_BACKFILL_BATCH_SIZE, j as ONTOLOGY_BACKFILL_MAX_PROMPT_CHARS, k as ONTOLOGY_BACKFILL_RECHECK_MS, l as OntologyBackfillResult, m as OntologyConfig, n as OntologyEdgeType, o as OntologyMode, p as OntologyNodeType, q as OntologyPromptContext, r as OntologyUpdates, P as PromptOverrides, s as PromptService, t as PrunePartialFailureError, V as VectorRanker, u as VectorRankerFallback, v as VectorRankerRankArgs, w as VectorRankerSemanticResult, x as WikiBusyError, y as WikiBusyOperation, z as WikiCheckpoint, A as WikiConfig, B as WikiEdge, C as WikiEvent, D as WikiFact, I as WikiMemoryTestAccess, J as WikiOutboxEvent, K as WikiTask, N as WikiTransactionError } from './testing-CBjAuTSl.mjs';
 import { OkfFile } from '@equationalapplications/core-okf';
 import 'minisearch';
 
@@ -26,6 +26,8 @@ declare function parseOkfBundle(entityId: string, files: OkfFile[], options?: Ok
 
 declare function parseEmbedding(blob: Uint8Array | null | undefined, text: string | null | undefined): Float32Array | null;
 
+declare function validateManifest(manifest: OntologyManifest): void;
+
 type GetRandomValues = (bytes: Uint8Array) => void;
 /**
  * Inject a platform-specific `getRandomValues` implementation.
@@ -62,4 +64,4 @@ declare const ONTOLOGY_BACKFILL_SYSTEM_PROMPT = "You are a knowledge classificat
 
 declare function createWiki(db: SQLiteAdapter, options: WikiOptions): WikiMemory;
 
-export { DEFAULT_LIBRARIAN_SYNTHESIS_PROMPT, FormatContextOptions, FormattedMemoryDump, GraphNeighborhood, type LibrarianOptions, type LibrarianPromptVariables, MemoryBundle, MemoryDump, ONTOLOGY_BACKFILL_SYSTEM_PROMPT, type OkfImportOptions, ReadOptions, SQLiteAdapter, WikiMemory, WikiOptions, configureRandomSource, createWiki, formatContext, formatGraphContext, formatMemoryDump, formatOkfBundle, hydrateLibrarianPrompt, mapLibrarianOptionsToReadOptions, parseEmbedding, parseOkfBundle, validateLibrarianPromptTemplate };
+export { DEFAULT_LIBRARIAN_SYNTHESIS_PROMPT, FormatContextOptions, FormattedMemoryDump, GraphNeighborhood, type LibrarianOptions, type LibrarianPromptVariables, MemoryBundle, MemoryDump, ONTOLOGY_BACKFILL_SYSTEM_PROMPT, type OkfImportOptions, OntologyManifest, ReadOptions, SQLiteAdapter, WikiMemory, WikiOptions, configureRandomSource, createWiki, formatContext, formatGraphContext, formatMemoryDump, formatOkfBundle, hydrateLibrarianPrompt, mapLibrarianOptionsToReadOptions, parseEmbedding, parseOkfBundle, validateLibrarianPromptTemplate, validateManifest };

package/dist/index.d.ts

@@ -1,5 +1,5 @@
-import { M as MemoryBundle, F as FormatContextOptions, G as GraphNeighborhood, a as MemoryDump, b as FormattedMemoryDump, R as ReadOptions, S as SQLiteAdapter, W as WikiOptions, c as WikiMemory } from './testing-i91HR1TG.js';
-export { E as EntityStatus, d as ExtractedFact, e as ExtractedFactEdge, f as ExtractedFactWithOntology, g as ExtractedTask, h as GraphTraversalOptions, H as HOOK_TIMEOUT_MARKER, L as LLMProvider, O as ONTOLOGY_BACKFILL_BATCH_SIZE, i as ONTOLOGY_BACKFILL_MAX_PROMPT_CHARS, j as ONTOLOGY_BACKFILL_RECHECK_MS, k as OntologyBackfillResult, l as OntologyConfig, m as OntologyEdgeType, n as OntologyManifest, o as OntologyMode, p as OntologyNodeType, q as OntologyPromptContext, r as OntologyUpdates, P as PromptOverrides, s as PromptService, t as PrunePartialFailureError, V as VectorRanker, u as VectorRankerFallback, v as VectorRankerRankArgs, w as VectorRankerSemanticResult, x as WikiBusyError, y as WikiBusyOperation, z as WikiCheckpoint, A as WikiConfig, B as WikiEdge, C as WikiEvent, D as WikiFact, I as WikiMemoryTestAccess, J as WikiOutboxEvent, K as WikiTask, N as WikiTransactionError } from './testing-i91HR1TG.js';
+import { M as MemoryBundle, F as FormatContextOptions, G as GraphNeighborhood, a as MemoryDump, b as FormattedMemoryDump, O as OntologyManifest, R as ReadOptions, S as SQLiteAdapter, W as WikiOptions, c as WikiMemory } from './testing-CBjAuTSl.js';
+export { E as EntityStatus, d as ExtractedFact, e as ExtractedFactEdge, f as ExtractedFactWithOntology, g as ExtractedTask, h as GraphTraversalOptions, H as HOOK_TIMEOUT_MARKER, L as LLMProvider, i as ONTOLOGY_BACKFILL_BATCH_SIZE, j as ONTOLOGY_BACKFILL_MAX_PROMPT_CHARS, k as ONTOLOGY_BACKFILL_RECHECK_MS, l as OntologyBackfillResult, m as OntologyConfig, n as OntologyEdgeType, o as OntologyMode, p as OntologyNodeType, q as OntologyPromptContext, r as OntologyUpdates, P as PromptOverrides, s as PromptService, t as PrunePartialFailureError, V as VectorRanker, u as VectorRankerFallback, v as VectorRankerRankArgs, w as VectorRankerSemanticResult, x as WikiBusyError, y as WikiBusyOperation, z as WikiCheckpoint, A as WikiConfig, B as WikiEdge, C as WikiEvent, D as WikiFact, I as WikiMemoryTestAccess, J as WikiOutboxEvent, K as WikiTask, N as WikiTransactionError } from './testing-CBjAuTSl.js';
 import { OkfFile } from '@equationalapplications/core-okf';
 import 'minisearch';
 
@@ -26,6 +26,8 @@ declare function parseOkfBundle(entityId: string, files: OkfFile[], options?: Ok
 
 declare function parseEmbedding(blob: Uint8Array | null | undefined, text: string | null | undefined): Float32Array | null;
 
+declare function validateManifest(manifest: OntologyManifest): void;
+
 type GetRandomValues = (bytes: Uint8Array) => void;
 /**
  * Inject a platform-specific `getRandomValues` implementation.
@@ -62,4 +64,4 @@ declare const ONTOLOGY_BACKFILL_SYSTEM_PROMPT = "You are a knowledge classificat
 
 declare function createWiki(db: SQLiteAdapter, options: WikiOptions): WikiMemory;
 
-export { DEFAULT_LIBRARIAN_SYNTHESIS_PROMPT, FormatContextOptions, FormattedMemoryDump, GraphNeighborhood, type LibrarianOptions, type LibrarianPromptVariables, MemoryBundle, MemoryDump, ONTOLOGY_BACKFILL_SYSTEM_PROMPT, type OkfImportOptions, ReadOptions, SQLiteAdapter, WikiMemory, WikiOptions, configureRandomSource, createWiki, formatContext, formatGraphContext, formatMemoryDump, formatOkfBundle, hydrateLibrarianPrompt, mapLibrarianOptionsToReadOptions, parseEmbedding, parseOkfBundle, validateLibrarianPromptTemplate };
+export { DEFAULT_LIBRARIAN_SYNTHESIS_PROMPT, FormatContextOptions, FormattedMemoryDump, GraphNeighborhood, type LibrarianOptions, type LibrarianPromptVariables, MemoryBundle, MemoryDump, ONTOLOGY_BACKFILL_SYSTEM_PROMPT, type OkfImportOptions, OntologyManifest, ReadOptions, SQLiteAdapter, WikiMemory, WikiOptions, configureRandomSource, createWiki, formatContext, formatGraphContext, formatMemoryDump, formatOkfBundle, hydrateLibrarianPrompt, mapLibrarianOptionsToReadOptions, parseEmbedding, parseOkfBundle, validateLibrarianPromptTemplate, validateManifest };

package/dist/index.js

@@ -707,6 +707,45 @@ var EntryRepository = class extends BaseRepository {
     );
     return rows.map(mapRowToFact);
   }
+  /**
+   * Fetch live, mutable entries for an entity — everything heal is allowed to
+   * downgrade or delete. Heal previously loaded every row via
+   * findAllByEntityId and filtered in JS, which on a document-heavy corpus
+   * meant loading 2560 rows to keep 31.
+   */
+  async findHealCandidatesByEntityId(entityId, tx) {
+    const executor = this.getExecutor(tx);
+    const rows = await executor.getAllAsync(
+      `SELECT * FROM ${this.prefix}entries
+       WHERE entity_id = ? AND deleted_at IS NULL AND source_type != 'immutable_document'
+       ORDER BY updated_at DESC`,
+      [entityId]
+    );
+    return rows.map(mapRowToFact);
+  }
+  /**
+   * Resolve search hits to document anchors. The MiniSearch index holds every
+   * fact, not only immutable_document rows, so the source-type restriction has
+   * to be applied here, after retrieval.
+   */
+  async findAnchorRowsByIds(entityId, ids, tx) {
+    if (ids.length === 0) return [];
+    const executor = this.getExecutor(tx);
+    const rows = [];
+    for (let i = 0; i < ids.length; i += this.chunkSize) {
+      const chunk = ids.slice(i, i + this.chunkSize);
+      const placeholders = chunk.map(() => "?").join(", ");
+      const chunkRows = await executor.getAllAsync(
+        `SELECT id, title, source_ref FROM ${this.prefix}entries
+         WHERE entity_id = ? AND deleted_at IS NULL
+           AND source_type = 'immutable_document'
+           AND id IN (${placeholders})`,
+        [entityId, ...chunk]
+      );
+      rows.push(...chunkRows);
+    }
+    return rows;
+  }
   /**
    * Fetch recent non-deleted entries for an entity (limited), ordered by updated_at DESC.
    * Used by MaintenanceService.doRunLibrarian().
@@ -1781,10 +1820,13 @@ function resolveNodeType(raw, manifest) {
   const hit = manifest.node_types.find((n) => n.type.toLowerCase() === slug.toLowerCase());
   return hit?.type ?? null;
 }
-function resolveEdgeDefinition(rawEdgeType, manifest) {
+function resolveEdgeDefinitions(rawEdgeType, manifest) {
   const slug = rawEdgeType.trim();
-  if (!slug) return null;
-  return manifest.edge_types.find((e) => e.type.toLowerCase() === slug.toLowerCase()) ?? null;
+  if (!slug) return [];
+  return manifest.edge_types.filter((e) => e.type.toLowerCase() === slug.toLowerCase());
+}
+function edgeTripleKey(type, sourceType, targetType) {
+  return `${type.trim().toLowerCase()}|${sourceType.trim().toLowerCase()}|${targetType.trim().toLowerCase()}`;
 }
 function validateManifest(manifest) {
   const nodeSlugs = /* @__PURE__ */ new Set();
@@ -1795,25 +1837,35 @@ function validateManifest(manifest) {
     if (nodeSlugs.has(key)) throw new Error(`Duplicate node type: ${type}`);
     nodeSlugs.add(key);
   }
-  const edgeSlugs = /* @__PURE__ */ new Set();
+  const edgeKeys = /* @__PURE__ */ new Set();
+  const edgeNames = /* @__PURE__ */ new Map();
   for (const edge of manifest.edge_types ?? []) {
     const edgeType = edge.type?.trim();
     const sourceType = edge.source_type?.trim();
     const targetType = edge.target_type?.trim();
     if (!edgeType) throw new Error("Ontology edge type slug must be non-empty");
-    const edgeKey = edgeType.toLowerCase();
-    if (edgeSlugs.has(edgeKey)) throw new Error(`Duplicate edge type: ${edgeType}`);
-    edgeSlugs.add(edgeKey);
     if (!sourceType || !targetType || !nodeSlugs.has(sourceType.toLowerCase()) || !nodeSlugs.has(targetType.toLowerCase())) {
       throw new Error(`Edge type ${edgeType} references unknown node type`);
     }
+    const edgeKey = edgeTripleKey(edgeType, sourceType, targetType);
+    if (edgeKeys.has(edgeKey)) {
+      throw new Error(`Duplicate edge definition: ${edgeType} (${sourceType} \u2192 ${targetType})`);
+    }
+    edgeKeys.add(edgeKey);
+    const canonical = edgeNames.get(edgeType.toLowerCase());
+    if (canonical === void 0) {
+      edgeNames.set(edgeType.toLowerCase(), edgeType);
+    } else if (canonical !== edgeType) {
+      throw new Error(`Inconsistent casing for edge type: ${edgeType} conflicts with ${canonical}`);
+    }
   }
 }
 function mergeOntologyUpdates(current, updates) {
   const node_types = [...current.node_types];
   const edge_types = [...current.edge_types];
   const nodeSlugs = new Set(node_types.map((n) => n.type.trim().toLowerCase()));
-  const edgeSlugs = new Set(edge_types.map((e) => e.type.trim().toLowerCase()));
+  const edgeKeys = new Set(edge_types.map((e) => edgeTripleKey(e.type, e.source_type, e.target_type)));
+  const edgeNames = new Map(edge_types.map((e) => [e.type.trim().toLowerCase(), e.type.trim()]));
   for (const node of updates.node_types ?? []) {
     const type = node?.type?.trim();
     if (!type) continue;
@@ -1823,20 +1875,22 @@ function mergeOntologyUpdates(current, updates) {
     nodeSlugs.add(key);
   }
   for (const edge of updates.edge_types ?? []) {
-    const edgeType = edge?.type?.trim();
+    const rawEdgeType = edge?.type?.trim();
     const sourceType = edge?.source_type?.trim();
     const targetType = edge?.target_type?.trim();
-    if (!edgeType || !sourceType || !targetType) continue;
-    const edgeKey = edgeType.toLowerCase();
-    if (edgeSlugs.has(edgeKey)) continue;
+    if (!rawEdgeType || !sourceType || !targetType) continue;
+    const edgeType = edgeNames.get(rawEdgeType.toLowerCase()) ?? rawEdgeType;
+    const edgeKey = edgeTripleKey(edgeType, sourceType, targetType);
+    if (edgeKeys.has(edgeKey)) continue;
     if (!nodeSlugs.has(sourceType.toLowerCase()) || !nodeSlugs.has(targetType.toLowerCase())) continue;
+    edgeNames.set(edgeType.toLowerCase(), edgeType);
     edge_types.push({
       type: edgeType,
       source_type: sourceType,
       target_type: targetType,
       description: String(edge.description ?? "")
     });
-    edgeSlugs.add(edgeKey);
+    edgeKeys.add(edgeKey);
   }
   return { node_types, edge_types };
 }
@@ -1845,10 +1899,10 @@ function validateInlineEdges(sourceType, _targetType, edges, manifest) {
   const valid = [];
   for (const edge of edges) {
     if (typeof edge?.edge_type !== "string" || typeof edge?.target_title !== "string") continue;
-    const def = resolveEdgeDefinition(edge.edge_type, manifest);
-    if (!def) continue;
-    if (def.source_type.toLowerCase() !== sourceType.toLowerCase()) continue;
-    valid.push({ edge_type: def.type, target_title: edge.target_title });
+    const defs = resolveEdgeDefinitions(edge.edge_type, manifest);
+    const match = defs.find((d) => d.source_type.toLowerCase() === sourceType.toLowerCase());
+    if (!match) continue;
+    valid.push({ edge_type: match.type, target_title: edge.target_title });
   }
   return valid;
 }
@@ -2048,9 +2102,23 @@ var _SearchService = class _SearchService {
     this.entryRepo = entryRepo;
     this.miniSearchEntryIdsByEntity = /* @__PURE__ */ new Map();
     this.vectorCache = /* @__PURE__ */ new Map();
+    /**
+     * Serializes rebuilds. `rebuildIndex` awaits a repository read between
+     * snapshotting the previous id set and discarding it, so two concurrent
+     * sync() calls for one entity can interleave: a slow, stale read lands last
+     * and discards documents the fresh read just added. Chaining also keeps
+     * discard()/addAll() out of each other's way, which is what accrued the
+     * auto-vacuum debt behind the TypeError in #64.
+     */
+    this.syncChain = Promise.resolve();
     this.miniSearch = new MiniSearch__default.default({
       fields: ["title", "body", "tags"],
       storeFields: ["entity_id"],
+      // Vacuuming is driven explicitly at the end of each serialized rebuild
+      // (see sync). Auto-vacuum fires on its own schedule, asynchronously with
+      // respect to the caller, and traversing the tree mid-rebuild is what
+      // threw the uncaught TypeError in MiniSearch.performVacuuming (#64).
+      autoVacuum: false,
       searchOptions: {
         boost: { title: 2 },
         fuzzy: 0.2,
@@ -2061,10 +2129,26 @@ var _SearchService = class _SearchService {
   /**
    * Rebuilds the search index and clears the vector cache for a given entity.
    * A direct replacement for manually syncing state after a DB transaction.
+   *
+   * Rebuilds are serialized per instance and never reject: the MiniSearch index
+   * is a rebuildable cache over SQLite, so degraded keyword search is the
+   * correct failure mode and killing the host process is not.
    */
   async sync(entityId) {
-    await this.rebuildIndex(entityId);
-    this.evictCache(entityId);
+    const work = this.syncChain.then(async () => {
+      try {
+        try {
+          await this.rebuildIndex(entityId);
+          await this.miniSearch.vacuum();
+        } finally {
+          this.evictCache(entityId);
+        }
+      } catch (err) {
+        console.warn(`[WikiMemory] search index rebuild failed for ${entityId ?? "*"}:`, err);
+      }
+    });
+    this.syncChain = work;
+    return work;
   }
   /**
    * Clears the parsed vector cache. Useful for mid-loop flush guarantees
@@ -3052,12 +3136,122 @@ var IngestionService = class {
   }
 };
 
+// src/services/BoundedLlmCall.ts
+var DEFAULT_BATCH_SIZE = 10;
+var ESTIMATED_OUTPUT_TOKENS_PER_ITEM = 150;
+var OUTPUT_BUDGET_FRACTION = 0.8;
+var TRUNCATION_PATTERNS = [
+  /truncat/i,
+  /token limit/i,
+  /max(imum)?[ _-]?tokens?/i,
+  /output limit/i,
+  /length limit/i,
+  /finish[_ ]?reason/i
+];
+var EXCEEDS_LIMIT_PATTERN = /exceed[a-z]*[^.]{0,40}\b(model|context)?[ _-]?limit/i;
+function isTruncationError(err) {
+  const message = err instanceof Error ? err.message : String(err ?? "");
+  if (EXCEEDS_LIMIT_PATTERN.test(message)) return false;
+  return TRUNCATION_PATTERNS.some((pattern) => pattern.test(message));
+}
+function initialBatchSize(maxOutputTokens) {
+  if (!maxOutputTokens || !Number.isFinite(maxOutputTokens) || maxOutputTokens <= 0) {
+    return DEFAULT_BATCH_SIZE;
+  }
+  const estimate = Math.floor(
+    maxOutputTokens * OUTPUT_BUDGET_FRACTION / ESTIMATED_OUTPUT_TOKENS_PER_ITEM
+  );
+  return Math.max(DEFAULT_BATCH_SIZE, estimate);
+}
+var promptLength = (prompts) => prompts.systemPrompt.length + prompts.userPrompt.length;
+async function runBatched(args) {
+  const { items, buildPrompt, call, parse, maxPromptChars, maxOutputTokens, onSkip } = args;
+  const results = [];
+  const skipped = [];
+  let batches = 0;
+  let batchSize = initialBatchSize(maxOutputTokens);
+  const trim = async (candidate) => {
+    const whole = await buildPrompt(candidate);
+    if (candidate.length <= 1 || promptLength(whole) <= maxPromptChars) {
+      return { batch: candidate, prompts: whole };
+    }
+    let low = 2;
+    let high = candidate.length - 1;
+    let best;
+    let bestPrompts;
+    while (low <= high) {
+      const mid = Math.floor((low + high) / 2);
+      const batch = candidate.slice(0, mid);
+      const prompts = await buildPrompt(batch);
+      if (promptLength(prompts) <= maxPromptChars) {
+        best = batch;
+        bestPrompts = prompts;
+        low = mid + 1;
+      } else {
+        high = mid - 1;
+      }
+    }
+    if (best && bestPrompts) return { batch: best, prompts: bestPrompts };
+    const single = candidate.slice(0, 1);
+    return { batch: single, prompts: await buildPrompt(single) };
+  };
+  const onFailure = async (batch, err) => {
+    if (batch.length <= 1) {
+      if (batch.length === 1) {
+        skipped.push(batch[0]);
+        onSkip?.(batch[0], err);
+      }
+      return;
+    }
+    const mid = Math.ceil(batch.length / 2);
+    if (mid < batchSize) batchSize = mid;
+    let i = 0;
+    while (i < batch.length) {
+      const size = Math.min(batchSize, batch.length - i);
+      const trimmed = await trim(batch.slice(i, i + size));
+      await attempt(trimmed.batch, trimmed.prompts);
+      i += trimmed.batch.length;
+    }
+  };
+  const attempt = async (batch, prebuilt) => {
+    if (batch.length === 0) return;
+    const prompts = prebuilt ?? await buildPrompt(batch);
+    batches++;
+    let responseText;
+    try {
+      responseText = await call(prompts);
+    } catch (err) {
+      if (!isTruncationError(err)) throw err;
+      await onFailure(batch, err);
+      return;
+    }
+    let result;
+    try {
+      result = parse(responseText, batch);
+    } catch (err) {
+      await onFailure(batch, err);
+      return;
+    }
+    results.push(result);
+  };
+  let index = 0;
+  while (index < items.length) {
+    const { batch, prompts } = await trim(items.slice(index, index + batchSize));
+    index += batch.length;
+    await attempt(batch, prompts);
+  }
+  return { results, skipped, batches };
+}
+
 // src/services/MaintenanceService.ts
 var FUZZY_THRESHOLD = 0.5;
 var MIN_TOKENS_TO_QUALIFY = 3;
 var ONTOLOGY_BACKFILL_BATCH_SIZE = 25;
 var ONTOLOGY_BACKFILL_MAX_PROMPT_CHARS = 4e4;
 var ONTOLOGY_BACKFILL_RECHECK_MS = 7 * 24 * 60 * 60 * 1e3;
+var HEAL_MAX_ANCHORS = 50;
+var HEAL_ANCHOR_SEARCH_OVERFETCH = 4;
+var HEAL_MAX_PROMPT_CHARS = 4e4;
 var MaintenanceService = class {
   constructor(db, prefix, options, entryRepo, taskRepo, eventRepo, metadataRepo, searchService, jobManager, embeddingService, promptService, ontologyService) {
     this.db = db;
@@ -3440,30 +3634,56 @@ var MaintenanceService = class {
         console.warn(`[WikiMemory] onEmbeddingPersisted hook failed during heal orphan pass for ${factId}:`, hookErr);
       }
     }
-    const allFactsRows = await this.entryRepo.findAllByEntityId(entityId);
+    const healCandidates = await this.entryRepo.findHealCandidatesByEntityId(entityId);
     const allTasks = await this.taskRepo.findAllPending([entityId]);
     const recentEvents = await this.eventRepo.getRecent(entityId, 20);
-    const healCandidates = allFactsRows.filter((f) => f.source_type !== "immutable_document");
-    const documentAnchors = allFactsRows.filter((f) => f.source_type === "immutable_document").map(({ id, title, source_ref }) => ({ id, title, source_ref }));
-    const healCandidatesForPrompt = healCandidates.map((f) => {
+    const toPromptShape = (f) => {
       const { embedding: _embedding, embedding_blob: _blob, ...rest } = f;
       return { ...rest, tags: typeof rest.tags === "string" ? JSON.parse(rest.tags) : rest.tags };
+    };
+    const anchorCache = /* @__PURE__ */ new Map();
+    const outcome = await runBatched({
+      items: healCandidates,
+      buildPrompt: async (batch) => {
+        const documentAnchors = await this._selectHealAnchors(entityId, batch, anchorCache);
+        return this.promptService.buildHealPrompt(
+          batch.map(toPromptShape),
+          documentAnchors,
+          allTasks,
+          recentEvents,
+          promptOverride
+        );
+      },
+      call: (prompts) => this.options.llmProvider.generateText(prompts),
+      parse: (responseText, batch) => {
+        const result = parseJsonResponse(responseText);
+        return {
+          batch,
+          downgraded: Array.isArray(result.downgraded) ? result.downgraded : [],
+          deleted: Array.isArray(result.deleted) ? result.deleted : [],
+          newFacts: Array.isArray(result.newFacts) ? result.newFacts : []
+        };
+      },
+      maxOutputTokens: this.options.llmProvider.maxOutputTokens,
+      maxPromptChars: HEAL_MAX_PROMPT_CHARS,
+      onSkip: (fact, err) => {
+        console.warn(
+          `[WikiMemory] heal skipped ${entityId}/${fact.id}: response could not be bounded`,
+          err
+        );
+      }
     });
-    const { systemPrompt, userPrompt } = this.promptService.buildHealPrompt(
-      healCandidatesForPrompt,
-      documentAnchors,
-      allTasks,
-      recentEvents,
-      promptOverride
-    );
-    const responseText = await this.options.llmProvider.generateText({ systemPrompt, userPrompt });
-    const result = parseJsonResponse(responseText);
-    const mutableIds = new Set(healCandidates.map((f) => f.id));
-    const downgraded = Array.isArray(result.downgraded) ? result.downgraded : [];
-    const deleted = Array.isArray(result.deleted) ? result.deleted : [];
-    const newFacts = Array.isArray(result.newFacts) ? result.newFacts : [];
-    const safeDowngraded = Array.from(new Set(downgraded.filter((id) => mutableIds.has(id))));
-    const safeDeleted = Array.from(new Set(deleted.filter((id) => mutableIds.has(id))));
+    const safeDowngradedSet = /* @__PURE__ */ new Set();
+    const safeDeletedSet = /* @__PURE__ */ new Set();
+    const newFacts = [];
+    for (const batchResult of outcome.results) {
+      const mutableIds = new Set(batchResult.batch.map((f) => f.id));
+      for (const id of batchResult.downgraded) if (mutableIds.has(id)) safeDowngradedSet.add(id);
+      for (const id of batchResult.deleted) if (mutableIds.has(id)) safeDeletedSet.add(id);
+      newFacts.push(...batchResult.newFacts);
+    }
+    const safeDowngraded = Array.from(safeDowngradedSet);
+    const safeDeleted = Array.from(safeDeletedSet);
     const validNewFacts = newFacts.map(validateFact).filter((f) => f !== null);
     const insertedFacts = [];
     const uniqueDeletedFactIds = Array.from(new Set(safeDeleted));
@@ -3530,7 +3750,7 @@ var MaintenanceService = class {
     }
     const now = Date.now();
     const recheckCutoff = now - ONTOLOGY_BACKFILL_RECHECK_MS;
-    const zeroed = { scanned: 0, typed: 0, failedValidation: 0, edgesAdded: 0 };
+    const zeroed = { scanned: 0, typed: 0, failedValidation: 0, edgesAdded: 0, skipped: 0 };
     const ontologyService = this.ontologyService;
     if (!ontologyService) {
       return { ...zeroed, remaining: 0, deferred: 0 };
@@ -3552,18 +3772,79 @@ var MaintenanceService = class {
       options?.promptOverride,
       ontologyContext
     );
-    const batch = [candidates[0]];
-    let built = buildPrompt(batch);
-    for (let i = 1; i < candidates.length; i++) {
-      const next = buildPrompt([...batch, candidates[i]]);
-      if (next.systemPrompt.length + next.userPrompt.length > ONTOLOGY_BACKFILL_MAX_PROMPT_CHARS) break;
-      batch.push(candidates[i]);
-      built = next;
-    }
-    const { systemPrompt, userPrompt } = built;
-    const responseText = await this.options.llmProvider.generateText({ systemPrompt, userPrompt });
-    const parsed = parseJsonResponse(responseText);
-    const classifications = Array.isArray(parsed.classifications) ? parsed.classifications : [];
+    const outcome = await runBatched({
+      items: candidates,
+      buildPrompt,
+      call: (prompts) => this.options.llmProvider.generateText(prompts),
+      parse: (responseText, batch) => {
+        const parsed = parseJsonResponse(responseText);
+        return {
+          batch,
+          classifications: Array.isArray(parsed.classifications) ? parsed.classifications : [],
+          ontologyUpdates: parsed.ontology_updates
+        };
+      },
+      maxOutputTokens: this.options.llmProvider.maxOutputTokens,
+      maxPromptChars: ONTOLOGY_BACKFILL_MAX_PROMPT_CHARS,
+      onSkip: (fact, err) => {
+        console.warn(
+          `[WikiMemory] ontology backfill skipped ${entityId}/${fact.id}: response could not be bounded`,
+          err
+        );
+      }
+    });
+    let typed = 0;
+    let failedValidation = 0;
+    let edgesAdded = 0;
+    let scanned = 0;
+    let abortedOntologyOff = false;
+    for (const batchResult of outcome.results) {
+      const applied = await this._applyOntologyBackfillBatch(entityId, batchResult, now);
+      if (applied.abortedOntologyOff) {
+        abortedOntologyOff = true;
+        break;
+      }
+      typed += applied.typed;
+      failedValidation += applied.failedValidation;
+      edgesAdded += applied.edgesAdded;
+      scanned += batchResult.batch.length;
+    }
+    if (abortedOntologyOff) {
+      const counts2 = await this.entryRepo.countUntypedByEntityId(entityId, recheckCutoff);
+      return {
+        scanned,
+        typed,
+        failedValidation,
+        edgesAdded,
+        skipped: outcome.skipped.length,
+        remaining: 0,
+        deferred: counts2.deferred
+      };
+    }
+    if (outcome.skipped.length > 0) {
+      await this.entryRepo.markOntologyChecked(outcome.skipped.map((f) => f.id), entityId, now, this.db);
+    }
+    this.searchService.evictCache(entityId);
+    const counts = await this.entryRepo.countUntypedByEntityId(entityId, recheckCutoff);
+    return {
+      scanned,
+      typed,
+      failedValidation,
+      edgesAdded,
+      skipped: outcome.skipped.length,
+      remaining: counts.eligible,
+      deferred: counts.deferred
+    };
+  }
+  /**
+   * Applies one parsed backfill batch in its own transaction. Per-batch rather
+   * than one transaction for the pass, so mergeEmergentUpdates semantics and
+   * the mid-flight `mode === 'off'` abort check keep the shape they had when a
+   * pass was a single call.
+   */
+  async _applyOntologyBackfillBatch(entityId, batchResult, now) {
+    const ontologyService = this.ontologyService;
+    const { batch, classifications, ontologyUpdates } = batchResult;
     let typed = 0;
     let failedValidation = 0;
     let edgesAdded = 0;
@@ -3574,8 +3855,8 @@ var MaintenanceService = class {
         abortedOntologyOff = true;
         return;
       }
-      if (txMode === "emergent" && parsed.ontology_updates) {
-        manifest = await ontologyService.mergeEmergentUpdates(entityId, parsed.ontology_updates, tx);
+      if (txMode === "emergent" && ontologyUpdates) {
+        manifest = await ontologyService.mergeEmergentUpdates(entityId, ontologyUpdates, tx);
       }
       const titleRows = await this.entryRepo.findTitleIndexByEntityId(entityId, tx);
       const titleIndex = /* @__PURE__ */ new Map();
@@ -3629,19 +3910,58 @@ var MaintenanceService = class {
       }
       await this.entryRepo.markOntologyChecked(batch.map((f) => f.id), entityId, now, tx);
     });
-    if (abortedOntologyOff) {
-      const counts2 = await this.entryRepo.countUntypedByEntityId(entityId, recheckCutoff);
-      return { ...zeroed, remaining: 0, deferred: counts2.deferred };
-    }
-    this.searchService.evictCache(entityId);
-    const counts = await this.entryRepo.countUntypedByEntityId(entityId, recheckCutoff);
-    return { scanned: batch.length, typed, failedValidation, edgesAdded, remaining: counts.eligible, deferred: counts.deferred };
+    return { typed, failedValidation, edgesAdded, abortedOntologyOff };
   }
   _validatePruneDuration(value, name) {
     if (value !== null && value !== void 0 && (typeof value !== "number" || !isFinite(value) || value < 0)) {
       throw new Error(`Invalid ${name}: must be a non-negative finite number or null`);
     }
   }
+  /**
+   * Anchors relevant to one batch of heal candidates.
+   *
+   * Heal used to pass every immutable_document fact for the entity — 2560 rows
+   * against 31 candidates on the corpus behind #63 — which is what blew the
+   * output ceiling. Anchors are now retrieved by keyword relevance to the batch
+   * and capped.
+   *
+   * The MiniSearch index holds all facts, not only anchors, so hits are
+   * overfetched and the source_type restriction is applied after retrieval, in
+   * SQL. Search rank order is preserved through the filter.
+   *
+   * Accepted tradeoff: an anchor that contradicts a candidate while sharing no
+   * vocabulary with it is now missed. Exhaustive-but-broken traded for
+   * relevance-scoped-and-working.
+   *
+   * `cache` is keyed by the derived query rather than by the batch, so two
+   * batches that reduce to the same query share one lookup. Caller-owned and
+   * per-pass — see the call site in doRunHeal.
+   */
+  async _selectHealAnchors(entityId, batch, cache) {
+    const query = batch.map((f) => f.title).join(" ").trim();
+    if (!query) return [];
+    const cached = cache?.get(query);
+    if (cached) return cached;
+    const hits = this.searchService.searchKeyword(
+      query,
+      [entityId],
+      HEAL_MAX_ANCHORS * HEAL_ANCHOR_SEARCH_OVERFETCH
+    );
+    const hitIds = hits.map((h) => h.id);
+    const anchors = [];
+    if (hitIds.length > 0) {
+      const rows = await this.entryRepo.findAnchorRowsByIds(entityId, hitIds);
+      const byId = new Map(rows.map((r) => [r.id, r]));
+      for (const id of hitIds) {
+        const row = byId.get(id);
+        if (!row) continue;
+        anchors.push(row);
+        if (anchors.length >= HEAL_MAX_ANCHORS) break;
+      }
+    }
+    cache?.set(query, anchors);
+    return anchors;
+  }
   _sanitizeRankerError(err) {
     return sanitizeRankerError(err, this.options.sanitizeRankerErrors);
   }
@@ -4854,7 +5174,8 @@ var WriteService = class {
 var FACT_ONTOLOGY_FIELDS = `
 Each fact may optionally include:
 - "okf_type": string \u2014 must be one of the node_types in the manifest
-- "edges": [{ "edge_type": string, "target_title": string }] \u2014 target_title must match another fact's title in this response or existing memory`;
+- "edges": [{ "edge_type": string, "target_title": string }] \u2014 target_title must match another fact's title in this response or existing memory
+- An edge_type may appear in the manifest multiple times with different source_type/target_type; use the row whose types match your fact and target.`;
 var EMERGENT_EXTRA = `
 You may also return "ontology_updates" to propose new types:
 "ontology_updates": {
@@ -4939,12 +5260,15 @@ var OntologyService = class {
     if (!sourceType || edges.length === 0) return 0;
     let persisted = 0;
     for (const edge of edges) {
-      const def = resolveEdgeDefinition(edge.edge_type, manifest);
-      if (!def || def.source_type.toLowerCase() !== sourceType.toLowerCase()) continue;
+      const candidates = resolveEdgeDefinitions(edge.edge_type, manifest).filter((d) => d.source_type.toLowerCase() === sourceType.toLowerCase());
+      if (candidates.length === 0) continue;
       const targetKey = normalizeTitleKey(edge.target_title);
       const target = titleIndex.get(targetKey);
       if (!target) continue;
-      if (def.target_type.toLowerCase() !== (target.okf_type ?? "").toLowerCase()) continue;
+      const def = candidates.find(
+        (d) => d.target_type.toLowerCase() === (target.okf_type ?? "").toLowerCase()
+      );
+      if (!def) continue;
       const wikiEdge = {
         id: generateId(),
         entity_id: entityId,
@@ -6073,5 +6397,6 @@ exports.mapLibrarianOptionsToReadOptions = mapLibrarianOptionsToReadOptions;
 exports.parseEmbedding = parseEmbedding;
 exports.parseOkfBundle = parseOkfBundle;
 exports.validateLibrarianPromptTemplate = validateLibrarianPromptTemplate;
+exports.validateManifest = validateManifest;
 //# sourceMappingURL=index.js.map
 //# sourceMappingURL=index.js.map