@gamaze/hicortex 0.10.0 → 0.11.0

package/dist/consolidate.d.ts

@@ -4,9 +4,49 @@
  * Ported from hicortex/consolidate/ (stages.py, __init__.py, budget.py).
  */
 import type Database from "better-sqlite3";
-import type { ConsolidationReport } from "./types.js";
+import type { Memory, ConsolidationReport } from "./types.js";
 import type { LlmClient } from "./llm.js";
 import type { EmbedFn } from "./retrieval.js";
+import { type DomainDef } from "./domain-classify.js";
+/**
+ * Minimum COSINE similarity for a link candidate.
+ *
+ * Calibration (2026-07): measured top-10 neighbor cosine histogram on the
+ * 2945-memory production corpus (bedrock). Typical top-1 neighbor cosine:
+ * median 0.823, p10 0.743, p90 0.902. Threshold 0.75 combined with the
+ * top-3 cap yields ≈ 2.2 candidate links/memory. The previous value (0.55)
+ * lived on an accidental 1−L2 scale where it required cosine > 0.90 — a
+ * near-duplicate detector that linked only 12% of memories.
+ */
+export declare const CONSOLIDATE_LINK_THRESHOLD = 0.75;
+/** Max link candidates kept per memory (highest-cosine neighbors first). */
+export declare const CONSOLIDATE_LINK_TOP_K = 3;
+/**
+ * Minimum COSINE similarity for a CROSS-PROJECT link candidate.
+ *
+ * A 672-link audit (17 LLM judges, 2026-07) found cross-project links were 65%
+ * wrong-link vs 6% for same-project, and that strength (cosine) predicts quality
+ * (wrong-link 42% → 6% across strength quartiles). Cross-project pairs must clear
+ * a much higher bar than the same-project 0.75 to survive discovery. Same-project
+ * links keep CONSOLIDATE_LINK_THRESHOLD (0.75).
+ */
+export declare const CROSS_PROJECT_LINK_THRESHOLD = 0.8;
+/**
+ * l2ToCosine moved to retrieval.ts (#145) — consolidate.ts already imports
+ * from retrieval, so retrieval is the circular-dependency-safe home.
+ * Re-exported here so pre-#145 importers and tests keep working unchanged.
+ */
+export { l2ToCosine } from "./retrieval.js";
+/**
+ * Minimum TRUE cosine similarity between a new lesson and an existing one
+ * for the existing lesson to count as a contradiction-check candidate
+ * (stageReflection). 0.80 = "strongly similar lesson" — the original intent.
+ * Before #145 the check was `1 − L2 > 0.80`, which required cosine > 0.98,
+ * so lesson-contradiction suppression effectively never fired.
+ */
+export declare const REFLECTION_CONTRADICTION_MIN_COSINE = 0.8;
+/** True when an L2 neighbor distance clears the contradiction-check bar. */
+export declare function isContradictionCandidate(distance: number): boolean;
 export declare class BudgetTracker {
     maxCalls: number;
     callsUsed: number;
@@ -21,10 +61,102 @@ export declare class BudgetTracker {
  * Parse JSON from LLM output, tolerating markdown fences and indexed formats.
  */
 export declare function parseJsonLenient<T>(text: string, fallback: T): T;
+/**
+ * Rebuild moduleIndex from the configured domain set + live DB counts, and
+ * persist it. Shared by the nightly stage and `hicortex classify-domains`.
+ * `projects` is left empty (content domains don't map to projects); the lesson
+ * selector's same-domain boost instead keys off memory.domain directly (it
+ * still reads the field). Descriptions are carried through for /index.
+ */
+export declare function rebuildContentModuleIndex(db: Database.Database, domains: DomainDef[], stateDir?: string): {
+    domains: number;
+};
+/** A candidate link discovered by vector similarity, pending classification. */
+export interface LinkCandidate {
+    source: Memory;
+    target: Memory & {
+        distance: number;
+    };
+    similarity: number;
+    heuristicType: string;
+}
+/**
+ * Discovery: find link candidates for one memory given its embedding.
+ * Top-10 vector neighbors (excluding self), keep the CONSOLIDATE_LINK_TOP_K
+ * highest-cosine neighbors above CONSOLIDATE_LINK_THRESHOLD, pre-compute the
+ * heuristic relationship type.
+ *
+ * Shared between the nightly `stageLinks` (which embeds via embedFn) and
+ * `hicortex relink` (which reuses stored embeddings from memory_vectors).
+ */
+export declare function discoverLinkCandidates(db: Database.Database, mem: Memory, embedding: Float32Array): LinkCandidate[];
+/**
+ * Classification: assign a relationship type to each candidate link.
+ *
+ * HEURISTIC-ONLY (2026-07). LLM edge classification was retired after the
+ * 672-link audit (see the Stage 3 header) found the LLM-classified UPPERCASE
+ * types near-useless (CONTRADICTS 4% acceptable). Every candidate now takes its
+ * pre-computed `heuristicType` (only `extends` or `relates_to` — see
+ * classifyRelationship). No LLM call is made.
+ *
+ * Signature stability: `llm` and `budget` are RETAINED but intentionally
+ * ignored so the callers (nightly `stageLinks`, `hicortex relink`) and the
+ * tests that import this need no change to their call sites. The return shape
+ * is unchanged; `llmClassified` is always 0 now and `heuristicFallback` counts
+ * every candidate. Do NOT re-add an LLM path here without a classifier that
+ * passes the audit harness at >= 70% acceptable.
+ *
+ * Shared between the nightly `stageLinks` and `hicortex relink`.
+ * Returns one relationship type per candidate (same order as input).
+ */
+export declare function classifyLinkCandidates(candidates: LinkCandidate[], _llm: LlmClient | null, _budget: BudgetTracker): Promise<{
+    types: string[];
+    llmClassified: number;
+    heuristicFallback: number;
+}>;
+/**
+ * Classify the relationship between two memories.
+ *
+ * TWO-LABEL heuristic (2026-07). The 672-link audit (see the Stage 3 header)
+ * showed only `extends` (57% acceptable) and `relates_to` (53%) held up; the
+ * emitted vocabulary is collapsed to exactly those two. The retired labels
+ * `updates` and `derives` (~31% acceptable) and all UPPERCASE LLM types are no
+ * longer produced. They remain in VALID_RELATIONSHIP_TYPES so pre-existing rows
+ * still validate.
+ *
+ * Rule: same-project (both projects non-null and equal) AND higher cosine
+ * (> CONSOLIDATE_LINK_THRESHOLD) → `extends`; everything else → `relates_to`.
+ *
+ * `similarity` is COSINE similarity (see l2ToCosine); the CONSOLIDATE_LINK_THRESHOLD
+ * boundary from the l2ToCosine calibration is preserved.
+ */
+export declare function classifyRelationship(source: Memory, target: Memory, similarity: number): string;
 /**
  * Run the full consolidation pipeline. Returns a structured report.
  */
-export declare function runConsolidation(db: Database.Database, llm: LlmClient, embedFn: EmbedFn, dryRun?: boolean, skipReflection?: boolean, stateDir?: string): Promise<ConsolidationReport>;
+/**
+ * Options controlling how the domain-assignment stage runs.
+ *
+ * When `domains` is a non-empty list, the pipeline uses content-based
+ * classification (config-owned) INSTEAD of project grouping — provided the
+ * classification endpoint (classify tier when configured, else reflect) passed
+ * pre-flight (`contentDomainsReady`). If that endpoint is unreachable, the
+ * caller sets `contentDomainsReady: false` and the stage is SKIPPED entirely
+ * (strict — no fall-back to a weak model or to project grouping). When
+ * `domains` is absent/empty, the legacy project-grouping curation runs
+ * unchanged.
+ */
+export interface DomainStageOptions {
+    domains?: DomainDef[] | null;
+    contentDomainsReady?: boolean;
+    /**
+     * Weak-primary floor for the no-fit path (see nofit.ts). Resolved by the
+     * caller from config (`weakPrimaryFloor`); defaults to
+     * DEFAULT_WEAK_PRIMARY_FLOOR when absent.
+     */
+    weakPrimaryFloor?: number;
+}
+export declare function runConsolidation(db: Database.Database, llm: LlmClient, embedFn: EmbedFn, dryRun?: boolean, skipReflection?: boolean, stateDir?: string, domainOptions?: DomainStageOptions): Promise<ConsolidationReport>;
 /**
  * Calculate milliseconds until the next occurrence of a given hour (local time).
  */
@@ -32,5 +164,10 @@ export declare function msUntilHour(hour: number): number;
 /**
  * Schedule the consolidation pipeline to run nightly.
  * Returns a cleanup function to cancel the timer.
+ *
+ * NOTE: currently unused (nightly.ts drives consolidation directly). Any future
+ * caller MUST read config.domains and thread `domainOptions` into runConsolidation
+ * when content domains are configured — otherwise it silently falls back to the
+ * legacy project-grouping path even when a domain list is set.
  */
 export declare function scheduleConsolidation(db: Database.Database, llm: LlmClient, embedFn: EmbedFn, hour?: number): () => void;

package/dist/consolidate.js

@@ -38,12 +38,16 @@ var __importStar = (this && this.__importStar) || (function () {
     };
 })();
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.BudgetTracker = void 0;
+exports.BudgetTracker = exports.REFLECTION_CONTRADICTION_MIN_COSINE = exports.l2ToCosine = exports.CROSS_PROJECT_LINK_THRESHOLD = exports.CONSOLIDATE_LINK_TOP_K = exports.CONSOLIDATE_LINK_THRESHOLD = void 0;
+exports.isContradictionCandidate = isContradictionCandidate;
 exports.parseJsonLenient = parseJsonLenient;
+exports.rebuildContentModuleIndex = rebuildContentModuleIndex;
+exports.discoverLinkCandidates = discoverLinkCandidates;
+exports.classifyLinkCandidates = classifyLinkCandidates;
+exports.classifyRelationship = classifyRelationship;
 exports.runConsolidation = runConsolidation;
 exports.msUntilHour = msUntilHour;
 exports.scheduleConsolidation = scheduleConsolidation;
-const types_js_1 = require("./types.js");
 const retrieval_js_1 = require("./retrieval.js");
 const storage = __importStar(require("./storage.js"));
 const prompts_js_1 = require("./prompts.js");
@@ -51,10 +55,54 @@ const node_crypto_1 = require("node:crypto");
 const features_js_1 = require("./features.js");
 const graph_js_1 = require("./graph.js");
 const state_js_1 = require("./state.js");
+const domain_classify_js_1 = require("./domain-classify.js");
+const schema_prototypes_js_1 = require("./schema-prototypes.js");
+const nofit_js_1 = require("./nofit.js");
 // Default config constants (matching Python config.py)
 const CONSOLIDATE_MAX_LLM_CALLS = 200;
 const CONSOLIDATE_PRUNE_MIN_AGE_DAYS = 90;
-const CONSOLIDATE_LINK_THRESHOLD = 0.55;
+/**
+ * Minimum COSINE similarity for a link candidate.
+ *
+ * Calibration (2026-07): measured top-10 neighbor cosine histogram on the
+ * 2945-memory production corpus (bedrock). Typical top-1 neighbor cosine:
+ * median 0.823, p10 0.743, p90 0.902. Threshold 0.75 combined with the
+ * top-3 cap yields ≈ 2.2 candidate links/memory. The previous value (0.55)
+ * lived on an accidental 1−L2 scale where it required cosine > 0.90 — a
+ * near-duplicate detector that linked only 12% of memories.
+ */
+exports.CONSOLIDATE_LINK_THRESHOLD = 0.75;
+/** Max link candidates kept per memory (highest-cosine neighbors first). */
+exports.CONSOLIDATE_LINK_TOP_K = 3;
+/**
+ * Minimum COSINE similarity for a CROSS-PROJECT link candidate.
+ *
+ * A 672-link audit (17 LLM judges, 2026-07) found cross-project links were 65%
+ * wrong-link vs 6% for same-project, and that strength (cosine) predicts quality
+ * (wrong-link 42% → 6% across strength quartiles). Cross-project pairs must clear
+ * a much higher bar than the same-project 0.75 to survive discovery. Same-project
+ * links keep CONSOLIDATE_LINK_THRESHOLD (0.75).
+ */
+exports.CROSS_PROJECT_LINK_THRESHOLD = 0.8;
+/**
+ * l2ToCosine moved to retrieval.ts (#145) — consolidate.ts already imports
+ * from retrieval, so retrieval is the circular-dependency-safe home.
+ * Re-exported here so pre-#145 importers and tests keep working unchanged.
+ */
+var retrieval_js_2 = require("./retrieval.js");
+Object.defineProperty(exports, "l2ToCosine", { enumerable: true, get: function () { return retrieval_js_2.l2ToCosine; } });
+/**
+ * Minimum TRUE cosine similarity between a new lesson and an existing one
+ * for the existing lesson to count as a contradiction-check candidate
+ * (stageReflection). 0.80 = "strongly similar lesson" — the original intent.
+ * Before #145 the check was `1 − L2 > 0.80`, which required cosine > 0.98,
+ * so lesson-contradiction suppression effectively never fired.
+ */
+exports.REFLECTION_CONTRADICTION_MIN_COSINE = 0.8;
+/** True when an L2 neighbor distance clears the contradiction-check bar. */
+function isContradictionCandidate(distance) {
+    return (0, retrieval_js_1.l2ToCosine)(distance) > exports.REFLECTION_CONTRADICTION_MIN_COSINE;
+}
 // ---------------------------------------------------------------------------
 // BudgetTracker
 // ---------------------------------------------------------------------------
@@ -273,11 +321,11 @@ async function stageReflection(db, memories, llm, budget, embedFn, dryRun) {
                 // If a very similar lesson exists, ask the LLM whether the new one
                 // contradicts it. If yes, suppress the new lesson to prevent the
                 // "false coherence" failure mode (wrong lessons reinforcing themselves).
+                // TRUE cosine > 0.80 (#145): the old `1 − n.distance > 0.80` sat on
+                // the accidental 1−L2 scale and required cosine > 0.98 — the check
+                // effectively never fired. See isContradictionCandidate.
                 const similarLessons = storage.vectorSearch(db, embedding, 3)
-                    .filter((n) => {
-                    const sim = 1.0 - n.distance;
-                    return sim > 0.80 && n.memory_type === "lesson";
-                });
+                    .filter((n) => isContradictionCandidate(n.distance) && n.memory_type === "lesson");
                 let contradicted = false;
                 if (similarLessons.length > 0 && budget.use("contradiction_check")) {
                     const existingText = similarLessons[0].content.slice(0, 300);
@@ -317,7 +365,142 @@ async function stageReflection(db, memories, llm, budget, embedFn, dryRun) {
     }
 }
 // ---------------------------------------------------------------------------
-// Stage 2.7: Domain Curation (MODULE_INDEX)
+// Stage 2.7a: Content-based Domain Classification (config-owned domains)
+// ---------------------------------------------------------------------------
+//
+// Active ONLY when config.json carries a `domains` list. Each memory is filed
+// into one configured life-sphere by its CONTENT (via the reflect model),
+// replacing the project-grouping path. Only NULL or stale-domain rows are
+// (re)classified, so re-runs are cheap and a config change re-files affected
+// rows. moduleIndex becomes {configured domains + live per-domain counts} so
+// /index and the lesson selector keep working.
+/**
+ * Rebuild moduleIndex from the configured domain set + live DB counts, and
+ * persist it. Shared by the nightly stage and `hicortex classify-domains`.
+ * `projects` is left empty (content domains don't map to projects); the lesson
+ * selector's same-domain boost instead keys off memory.domain directly (it
+ * still reads the field). Descriptions are carried through for /index.
+ */
+function rebuildContentModuleIndex(db, domains, stateDir) {
+    const memRows = db
+        .prepare(`SELECT domain, COUNT(*) AS cnt FROM memories WHERE domain IS NOT NULL GROUP BY domain`)
+        .all();
+    const lessonRows = db
+        .prepare(`SELECT domain, COUNT(*) AS cnt FROM memories
+       WHERE domain IS NOT NULL AND memory_type = 'lesson' GROUP BY domain`)
+        .all();
+    const memByDomain = new Map(memRows.map((r) => [r.domain, r.cnt]));
+    const lessonByDomain = new Map(lessonRows.map((r) => [r.domain, r.cnt]));
+    const moduleDomains = domains.map((d) => ({
+        name: d.name,
+        projects: [],
+        memoryCount: memByDomain.get(d.name) ?? 0,
+        lessonCount: lessonByDomain.get(d.name) ?? 0,
+        keywords: [],
+        description: d.description,
+    }));
+    const totalMemories = moduleDomains.reduce((s, d) => s + d.memoryCount, 0);
+    const totalLessons = moduleDomains.reduce((s, d) => s + d.lessonCount, 0);
+    const moduleIndex = {
+        domains: moduleDomains,
+        projectSetHash: (0, domain_classify_js_1.domainSetHash)(domains),
+        curatedAt: new Date().toISOString(),
+        totalMemories,
+        totalLessons,
+        mode: "content",
+    };
+    (0, state_js_1.updateState)((s) => { s.moduleIndex = moduleIndex; }, stateDir);
+    return { domains: moduleDomains.length };
+}
+async function stageContentDomains(db, domains, llm, budget, embedFn, dryRun, stateDir, weakPrimaryFloor = nofit_js_1.DEFAULT_WEAK_PRIMARY_FLOOR) {
+    // Rows needing (re)classification:
+    //   - domain IS NULL (never classified), OR
+    //   - domain NOT IN the current vocabulary (a rename/removal re-files), OR
+    //   - no memory_tags rows yet (single-domain memories from feat/content-domains
+    //     that have a primary but no tag set — backfill them to multi-tag).
+    const placeholders = domains.map(() => "?").join(", ");
+    const rows = db
+        .prepare(`SELECT id, content, project FROM memories
+       WHERE domain IS NULL
+          OR domain NOT IN (${placeholders})
+          OR id NOT IN (SELECT DISTINCT memory_id FROM memory_tags)`)
+        .all(...domains.map((d) => d.name));
+    if (dryRun) {
+        return { curated: false, domains: domains.length, classified: 0, reason: `dry_run (${rows.length} would classify)` };
+    }
+    const getEmbedFn = async () => embedFn;
+    const compartments = (0, schema_prototypes_js_1.compartmentSet)(domains);
+    let classified = 0;
+    let weakPrimary = 0;
+    let noAssociationDecayed = 0;
+    if (rows.length > 0) {
+        // Prototypes at run start — newly classified memories get their weights
+        // from the CURRENT prototypes (the post-classification recompute below
+        // refreshes everything from the updated tag sets anyway).
+        const { prototypes: startPrototypes } = await (0, schema_prototypes_js_1.computeDomainPrototypes)(db, domains, getEmbedFn);
+        for (const row of rows) {
+            if (budget.exhausted || !budget.use("content_domain")) {
+                console.warn(`[hicortex] content-domain: budget exhausted after ${classified} classified`);
+                break;
+            }
+            // classifyMemoryTags returns null ONLY on infra error (throws after retry) —
+            // skip that memory, leaving domain/tags/strength untouched so a later
+            // run retries it (issue #150: never file or decay on infra errors).
+            const result = await (0, domain_classify_js_1.classifyMemoryTags)(row.content, row.project, domains, llm);
+            if (result === null) {
+                console.warn(`[hicortex] content-domain: infra error classifying ${row.id} — skipped (will retry)`);
+                continue;
+            }
+            if (result.tags.length === 0) {
+                // Genuine no-fit (owner amendment 07.07): weak primary from the
+                // prototype argmax when it clears the floor, else accelerated decay.
+                // Each row appears exactly once in `rows`, so a run never
+                // double-halves.
+                const resolution = (0, nofit_js_1.resolveNoFit)(db, row.id, domains, startPrototypes, weakPrimaryFloor);
+                if (resolution.kind === "weak_primary") {
+                    (0, nofit_js_1.applyWeakPrimary)(db, row.id, resolution.domain, resolution.weight, compartments);
+                    weakPrimary++;
+                }
+                else {
+                    (0, nofit_js_1.applyNoAssociationDecay)(db, row.id);
+                    noAssociationDecayed++;
+                }
+                continue;
+            }
+            const weights = (0, schema_prototypes_js_1.computeTagWeights)(db, row.id, result.tags, startPrototypes);
+            storage.setMemoryTags(db, row.id, result.tags, { weights, compartments });
+            classified++;
+        }
+    }
+    // Graded-schema reconsolidation pass — runs EVERY nightly, including when
+    // nothing new was classified: prototypes drift with the data, so weights and
+    // derived primaries must follow (spec: "recomputed for ALL memory_tags rows
+    // each nightly"). Order: prototypes (from the post-classification tag sets)
+    // → all weights → derived primaries → moduleIndex counts from the refreshed
+    // primaries. No LLM calls — embeddings only.
+    const { prototypes, stats } = await (0, schema_prototypes_js_1.computeDomainPrototypes)(db, domains, getEmbedFn);
+    const { updated: weightsRecomputed } = (0, schema_prototypes_js_1.recomputeAllTagWeights)(db, prototypes);
+    const { updated: primariesUpdated } = (0, schema_prototypes_js_1.refreshPrimaries)(db, domains);
+    const seeded = stats.filter((s) => s.seeded).length;
+    const { domains: domainCount } = rebuildContentModuleIndex(db, domains, stateDir);
+    console.log(`[hicortex] Graded tags: ${classified} classified, ${weakPrimary} weak-primary, ` +
+        `${noAssociationDecayed} no-association decayed, ${prototypes.size} prototypes ` +
+        `(${seeded} description-seeded), ${weightsRecomputed} weights recomputed, ` +
+        `${primariesUpdated} primaries updated, ${domainCount} domains indexed`);
+    return {
+        curated: rows.length > 0,
+        domains: domainCount,
+        classified,
+        prototypes: prototypes.size,
+        weights_recomputed: weightsRecomputed,
+        primaries_updated: primariesUpdated,
+        weak_primary: weakPrimary,
+        no_association_decayed: noAssociationDecayed,
+        ...(rows.length === 0 ? { reason: "nothing_stale" } : {}),
+    };
+}
+// ---------------------------------------------------------------------------
+// Stage 2.7b: Domain Curation (MODULE_INDEX) — project grouping (legacy path)
 // ---------------------------------------------------------------------------
 async function stageDomainCuration(db, llm, budget, dryRun, stateDir) {
     // Gather all projects with memory and lesson counts
@@ -481,28 +664,80 @@ async function stageDomainCuration(db, llm, budget, dryRun, stateDir) {
     console.log(`[hicortex] Domain curation: ${domains.length} domains from ${projectRows.length} projects`);
     return { curated: true, domains: domains.length };
 }
-/** Batch size for LLM edge classification calls. */
-const EDGE_CLASSIFICATION_BATCH_SIZE = 8;
-/** Valid relationship type set for fast lookup. */
-const VALID_REL_SET = new Set(types_js_1.VALID_RELATIONSHIP_TYPES);
+/**
+ * Discovery: find link candidates for one memory given its embedding.
+ * Top-10 vector neighbors (excluding self), keep the CONSOLIDATE_LINK_TOP_K
+ * highest-cosine neighbors above CONSOLIDATE_LINK_THRESHOLD, pre-compute the
+ * heuristic relationship type.
+ *
+ * Shared between the nightly `stageLinks` (which embeds via embedFn) and
+ * `hicortex relink` (which reuses stored embeddings from memory_vectors).
+ */
+function discoverLinkCandidates(db, mem, embedding) {
+    const neighbors = storage.vectorSearch(db, embedding, 10, [mem.id]);
+    const candidates = [];
+    // vectorSearch orders by L2 distance ascending (`ORDER BY distance` in
+    // storage.ts), and cosine is monotonically decreasing in L2 distance for
+    // normalized vectors — so iterating in order and stopping at TOP_K keeps
+    // exactly the highest-cosine neighbors.
+    for (const neighbor of neighbors) {
+        if (candidates.length >= exports.CONSOLIDATE_LINK_TOP_K)
+            break;
+        // sqlite-vec vec0 `distance` is L2, not a similarity. Embeddings are
+        // L2-normalized (embedder.ts, normalize: true), so cos = 1 − d²/2.
+        // The old `1 − distance` formula silently required cosine > 0.90.
+        const similarity = (0, retrieval_js_1.l2ToCosine)(neighbor.distance);
+        // Cross-project guard (2026-07 audit): cross-project links were 65%
+        // wrong-link vs 6% same-project. A candidate whose source/target belong to
+        // DIFFERENT projects must clear the higher CROSS_PROJECT_LINK_THRESHOLD;
+        // same-project keeps CONSOLIDATE_LINK_THRESHOLD. A memory with no project
+        // (null) is treated as same-project — the guard only fires on two distinct
+        // non-null project names.
+        const crossProject = mem.project != null &&
+            neighbor.project != null &&
+            mem.project !== neighbor.project;
+        const threshold = crossProject
+            ? exports.CROSS_PROJECT_LINK_THRESHOLD
+            : exports.CONSOLIDATE_LINK_THRESHOLD;
+        if (similarity > threshold) {
+            const heuristicType = classifyRelationship(mem, neighbor, similarity);
+            candidates.push({ source: mem, target: neighbor, similarity, heuristicType });
+        }
+    }
+    return candidates;
+}
+/**
+ * Classification: assign a relationship type to each candidate link.
+ *
+ * HEURISTIC-ONLY (2026-07). LLM edge classification was retired after the
+ * 672-link audit (see the Stage 3 header) found the LLM-classified UPPERCASE
+ * types near-useless (CONTRADICTS 4% acceptable). Every candidate now takes its
+ * pre-computed `heuristicType` (only `extends` or `relates_to` — see
+ * classifyRelationship). No LLM call is made.
+ *
+ * Signature stability: `llm` and `budget` are RETAINED but intentionally
+ * ignored so the callers (nightly `stageLinks`, `hicortex relink`) and the
+ * tests that import this need no change to their call sites. The return shape
+ * is unchanged; `llmClassified` is always 0 now and `heuristicFallback` counts
+ * every candidate. Do NOT re-add an LLM path here without a classifier that
+ * passes the audit harness at >= 70% acceptable.
+ *
+ * Shared between the nightly `stageLinks` and `hicortex relink`.
+ * Returns one relationship type per candidate (same order as input).
+ */
+async function classifyLinkCandidates(candidates, _llm, _budget) {
+    const types = candidates.map((c) => c.heuristicType);
+    return { types, llmClassified: 0, heuristicFallback: candidates.length };
+}
 async function stageLinks(db, memories, embedFn, dryRun, llm, budget) {
     let autoLinked = 0;
-    let llmClassified = 0;
-    let heuristicFallback = 0;
     let failed = 0;
     // Phase A: Discovery — collect candidates via vector similarity
     const candidates = [];
     for (const mem of memories) {
         try {
             const embedding = await embedFn(mem.content);
-            const neighbors = storage.vectorSearch(db, embedding, 10, [mem.id]);
-            for (const neighbor of neighbors) {
-                const similarity = 1.0 - neighbor.distance;
-                if (similarity > CONSOLIDATE_LINK_THRESHOLD) {
-                    const heuristicType = classifyRelationship(mem, neighbor, similarity);
-                    candidates.push({ source: mem, target: neighbor, similarity, heuristicType });
-                }
-            }
+            candidates.push(...discoverLinkCandidates(db, mem, embedding));
         }
         catch {
             failed++;
@@ -511,48 +746,8 @@ async function stageLinks(db, memories, embedFn, dryRun, llm, budget) {
     if (candidates.length === 0) {
         return { auto_linked: 0, llm_classified: 0, heuristic_fallback: 0, failed };
     }
-    // Phase B: LLM batch classification
-    // Build batches and classify with LLM where budget allows
-    const classifiedTypes = new Array(candidates.length);
-    for (let i = 0; i < candidates.length; i += EDGE_CLASSIFICATION_BATCH_SIZE) {
-        const batch = candidates.slice(i, i + EDGE_CLASSIFICATION_BATCH_SIZE);
-        // Attempt LLM classification if budget allows
-        if (budget.use("edge_classification")) {
-            try {
-                const pairsBlock = batch.map((c, idx) => {
-                    const srcContent = c.source.content.slice(0, 200);
-                    const tgtContent = c.target.content.slice(0, 200);
-                    return `[${idx}] SOURCE: ${c.source.memory_type} | ${c.source.project ?? "global"} | ${srcContent}\n    TARGET: ${c.target.memory_type} | ${c.target.project ?? "global"} | ${tgtContent}\n    similarity: ${c.similarity.toFixed(2)}`;
-                }).join("\n\n");
-                const prompt = (0, prompts_js_1.edgeClassification)(pairsBlock);
-                const raw = await llm.completeFast(prompt, 512);
-                const parsed = parseJsonLenient(raw, []);
-                if (Array.isArray(parsed) && parsed.length > 0) {
-                    for (let j = 0; j < batch.length; j++) {
-                        const llmType = parsed[j];
-                        if (typeof llmType === "string" && VALID_REL_SET.has(llmType)) {
-                            classifiedTypes[i + j] = llmType;
-                            llmClassified++;
-                        }
-                        else {
-                            // Invalid type from LLM — fall back to heuristic
-                            classifiedTypes[i + j] = batch[j].heuristicType;
-                            heuristicFallback++;
-                        }
-                    }
-                    continue;
-                }
-            }
-            catch {
-                // LLM call failed — fall through to heuristic for this batch
-            }
-        }
-        // Budget exhausted or LLM failed — use heuristic for entire batch
-        for (let j = 0; j < batch.length; j++) {
-            classifiedTypes[i + j] = batch[j].heuristicType;
-            heuristicFallback++;
-        }
-    }
+    // Phase B: LLM batch classification (heuristic fallback inside)
+    const { types: classifiedTypes, llmClassified, heuristicFallback } = await classifyLinkCandidates(candidates, llm, budget);
     // Phase C: Store all classified links
     for (let i = 0; i < candidates.length; i++) {
         const c = candidates[i];
@@ -573,24 +768,26 @@ async function stageLinks(db, memories, embedFn, dryRun, llm, budget) {
     return { auto_linked: autoLinked, llm_classified: llmClassified, heuristic_fallback: heuristicFallback, failed };
 }
 /**
- * Classify the relationship between two memories based on type, temporal ordering, and similarity.
+ * Classify the relationship between two memories.
+ *
+ * TWO-LABEL heuristic (2026-07). The 672-link audit (see the Stage 3 header)
+ * showed only `extends` (57% acceptable) and `relates_to` (53%) held up; the
+ * emitted vocabulary is collapsed to exactly those two. The retired labels
+ * `updates` and `derives` (~31% acceptable) and all UPPERCASE LLM types are no
+ * longer produced. They remain in VALID_RELATIONSHIP_TYPES so pre-existing rows
+ * still validate.
+ *
+ * Rule: same-project (both projects non-null and equal) AND higher cosine
+ * (> CONSOLIDATE_LINK_THRESHOLD) → `extends`; everything else → `relates_to`.
+ *
+ * `similarity` is COSINE similarity (see l2ToCosine); the CONSOLIDATE_LINK_THRESHOLD
+ * boundary from the l2ToCosine calibration is preserved.
  */
 function classifyRelationship(source, target, similarity) {
-    // Lesson derived from episode(s)
-    if (source.memory_type === "lesson" && target.memory_type === "episode")
-        return "derives";
-    if (target.memory_type === "lesson" && source.memory_type === "episode")
-        return "derives";
-    // Same type + very high similarity + different timestamps → newer updates older
-    if (source.memory_type === target.memory_type &&
-        similarity > 0.8 &&
-        source.created_at !== target.created_at) {
-        return "updates";
-    }
-    // Same project, moderate similarity → extends
+    // Same project + above the link threshold → extends
     if (source.project && target.project &&
         source.project === target.project &&
-        similarity > 0.55 && similarity <= 0.8) {
+        similarity > exports.CONSOLIDATE_LINK_THRESHOLD) {
         return "extends";
     }
     return "relates_to";
@@ -656,13 +853,7 @@ function stageDecayPrune(db, dryRun) {
     }
     return { candidates, pruned, failed };
 }
-// ---------------------------------------------------------------------------
-// Full pipeline
-// ---------------------------------------------------------------------------
-/**
- * Run the full consolidation pipeline. Returns a structured report.
- */
-async function runConsolidation(db, llm, embedFn, dryRun = false, skipReflection = false, stateDir) {
+async function runConsolidation(db, llm, embedFn, dryRun = false, skipReflection = false, stateDir, domainOptions) {
     const start = new Date();
     const report = {
         started_at: start.toISOString(),
@@ -708,8 +899,27 @@ async function runConsolidation(db, llm, embedFn, dryRun = false, skipReflection
         else {
             report.stages.reflection = await stageReflection(db, precheck.newMemories, llm, budget, embedFn, dryRun);
         }
-        // Stage 2.7: Domain Curation
-        report.stages.domain_curation = await stageDomainCuration(db, llm, budget, dryRun, stateDir);
+        // Stage 2.7: Domain assignment.
+        // Content-based (config-owned domains) REPLACES project grouping when a
+        // domain list is configured AND the reflect endpoint passed pre-flight.
+        // If the list is configured but the reflect endpoint is down, SKIP the
+        // stage (strict) — do not fall back to project grouping.
+        const cfgDomains = domainOptions?.domains;
+        if (cfgDomains && cfgDomains.length > 0) {
+            if (domainOptions?.contentDomainsReady === false) {
+                report.stages.domain_curation = {
+                    curated: false,
+                    domains: cfgDomains.length,
+                    reason: "reflect_endpoint_offline",
+                };
+            }
+            else {
+                report.stages.domain_curation = await stageContentDomains(db, cfgDomains, llm, budget, embedFn, dryRun, stateDir, domainOptions?.weakPrimaryFloor ?? nofit_js_1.DEFAULT_WEAK_PRIMARY_FLOOR);
+            }
+        }
+        else {
+            report.stages.domain_curation = await stageDomainCuration(db, llm, budget, dryRun, stateDir);
+        }
         // Stage 3: Link Discovery (with LLM-assisted edge classification)
         report.stages.links = await stageLinks(db, precheck.newMemories, embedFn, dryRun, llm, budget);
         // Stage 3.5: Hub Detection — boost highly-connected memories
@@ -754,6 +964,11 @@ const ONE_DAY_MS = 24 * 60 * 60 * 1000;
 /**
  * Schedule the consolidation pipeline to run nightly.
  * Returns a cleanup function to cancel the timer.
+ *
+ * NOTE: currently unused (nightly.ts drives consolidation directly). Any future
+ * caller MUST read config.domains and thread `domainOptions` into runConsolidation
+ * when content domains are configured — otherwise it silently falls back to the
+ * legacy project-grouping path even when a domain list is set.
  */
 function scheduleConsolidation(db, llm, embedFn, hour = 2) {
     let timeout = null;