recallnest 2.3.0 → 2.6.0

package/src/pivot-apply.ts

@@ -0,0 +1,399 @@
+/**
+ * Pivot Apply — dedicated structured write channel for reviewed pivot candidates.
+ *
+ * Purpose: apply LLM-distilled pivot candidates (reportId-keyed, reviewed via
+ * allowlist) into scope `memory:pivot` with deterministic dispositions.
+ *
+ * Why not the normal store_memory path (mutual-review P0, 2026-08-04):
+ * - persistMemory routes preferences through matchPreference (LLM semantic
+ *   merge) → polymorphic dispositions the apply ledger cannot predict;
+ * - A-2 supersede detection calls the LLM per write (re-judging candidates is
+ *   out of contract for apply);
+ * - default source=manual mislabels distilled content as 0.9/direct confidence;
+ * - kind/anchor/evidence/reportId/candidateHash/batchId are dropped;
+ * - the implicit preference dual-write (LME-1) is fire-and-forget, so its id
+ *   can never reach the ledger.
+ *
+ * This channel therefore:
+ * - keeps: PII scan+redact (trust boundary), rule-based admission (minus rate
+ *   limiting — bulk writes always trip it and it is not a quality signal),
+ *   canonicalKey supersede semantics via writeDurableEntry, audit logging,
+ *   synchronous post-write verification;
+ * - drops: every LLM call, F2 interference marking (must not silently mutate
+ *   existing memories), async KG extraction, and the implicit preference
+ *   dual-write (apply's contract is to write exactly the allowlisted entries
+ *   and nothing else — a derived preference is not on the allowlist).
+ */
+
+import { z } from "zod";
+
+import { checkAdmission } from "./admission-control.js";
+import {
+  buildStructuredMetadata,
+  writeDurableEntry,
+  type PersistMemoryDeps,
+} from "./capture-engine.js";
+import { assignDefaultConfidence } from "./confidence-tracker.js";
+import { detectLang, tokenizeFts } from "./language-hook.js";
+import { scanForPII, redactSecrets } from "./pii-detector.js";
+
+export const PIVOT_APPLY_SOURCE = "session_distill";
+export const PIVOT_APPLY_CAPTURE = "pivot_apply_v1";
+export const PIVOT_SCOPE = "memory:pivot";
+
+const SHA256_HEX = /^[0-9a-f]{64}$/;
+
+export const PivotKindSchema = z.enum([
+  "judgment_shift",
+  "decision",
+  "preference_rule",
+  "case",
+]);
+export type PivotKind = z.infer<typeof PivotKindSchema>;
+
+/** The only kind→category pairs that exist in the corpus (verified over all 8,971 candidates). */
+const KIND_CATEGORY: Record<PivotKind, "cases" | "events" | "preferences"> = {
+  case: "cases",
+  decision: "events",
+  judgment_shift: "events",
+  preference_rule: "preferences",
+};
+
+export const PivotCandidateInputSchema = z
+  .object({
+    kind: PivotKindSchema,
+    text: z.string().min(10).max(4000),
+    anchor: z.string().min(1).max(2000),
+    canonicalKey: z.string().min(1).max(120),
+    evidence: z.array(z.string().min(1).max(4000)).max(16),
+    proposedScope: z.literal(PIVOT_SCOPE),
+    proposedCategory: z.enum(["cases", "events", "preferences"]),
+    // candidate tags max out at 4 in the corpus; the channel appends 2 more and
+    // buildStructuredMetadata stores tags verbatim, so cap input at 6 to stay
+    // within the schema-wide 8-tag budget.
+    tags: z.array(z.string().min(1).max(40)).min(1).max(6),
+    sessionId: z.string().min(8).max(200),
+    reportId: z.string().regex(SHA256_HEX, "reportId must be 64-char sha256 hex"),
+    candidateHash: z.string().regex(SHA256_HEX, "candidateHash must be 64-char sha256 hex"),
+    batchId: z.string().min(1).max(64),
+    importance: z.number().min(0).max(1),
+  })
+  .refine((c) => KIND_CATEGORY[c.kind] === c.proposedCategory, {
+    message: "proposedCategory does not match kind (corpus-verified mapping violated)",
+  });
+export type PivotCandidateInput = z.infer<typeof PivotCandidateInputSchema>;
+
+export type PivotApplyDisposition =
+  | "stored"
+  | "updated"
+  | "deduped"
+  | "promoted"
+  | "conflict"
+  | "rejected";
+
+/**
+ * The stock embedder only retries connection-class errors; HTTP 429/5xx throw
+ * straight through (embedder.ts isTransientEmbeddingError). A bulk apply makes
+ * thousands of calls, so rate-limit blips must not void a whole batch — retry
+ * with backoff here, and only here (single-write callers keep stock behavior).
+ */
+async function embedWithRetry(
+  embedder: PersistMemoryDeps["embedder"],
+  text: string,
+  delaysMs: readonly number[] = [1000, 5000, 15000],
+): Promise<number[]> {
+  let lastErr: unknown;
+  for (let attempt = 0; attempt <= delaysMs.length; attempt++) {
+    try {
+      return await embedder.embedPassage(text);
+    } catch (err) {
+      lastErr = err;
+      const msg = String((err as Error)?.message ?? err);
+      const retriable = /\b(429|500|502|503|504)\b|rate.?limit|overloaded/i.test(msg);
+      if (!retriable || attempt === delaysMs.length) throw err;
+      await new Promise((r) => setTimeout(r, delaysMs[attempt]));
+    }
+  }
+  throw lastErr;
+}
+
+export interface PivotApplyOutcome {
+  candidateHash: string;
+  batchId: string;
+  /** null when the write never reached the store (admission rejection). */
+  memoryId: string | null;
+  disposition: PivotApplyDisposition;
+  conflictId?: string;
+  rejectionReason?: string;
+  supersededOldId?: string;
+  piiRedactedCount: number;
+  canonicalKey: string;
+  category: "cases" | "events" | "preferences";
+  verified: boolean;
+  verifyIssues: string[];
+}
+
+export interface PivotBatchReport {
+  batchId: string;
+  total: number;
+  written: number;
+  halted: boolean;
+  haltReason?: string;
+  outcomes: PivotApplyOutcome[];
+}
+
+/**
+ * Persist one reviewed pivot candidate. Deterministic: no LLM calls, no writes
+ * outside the candidate itself (plus the audit log line).
+ */
+export async function persistPivotCandidate(
+  deps: PersistMemoryDeps,
+  rawInput: unknown,
+): Promise<PivotApplyOutcome> {
+  const input = PivotCandidateInputSchema.parse(rawInput);
+
+  // F-3b ordering (same as persistMemory): scan first so piiWarning reflects
+  // the raw text, then redact before the text reaches embedding/storage.
+  const piiScan = scanForPII(input.text);
+  const redaction = redactSecrets(input.text);
+  const text = redaction.redacted > 0 ? redaction.text : input.text;
+
+  // Rule-based admission only. No rateLimiter on purpose: a bulk apply always
+  // exceeds per-scope write rates, which says nothing about entry quality.
+  const admission = checkAdmission(text, input.importance, PIVOT_SCOPE, undefined, deps.admissionConfig);
+  if (admission.verdict === "rejected") {
+    try {
+      deps.auditLogger?.log({
+        operation: "reject",
+        scope: PIVOT_SCOPE,
+        actor: PIVOT_APPLY_SOURCE,
+        details: `pivot-apply ${input.kind} ${input.candidateHash.slice(0, 12)}: ${admission.reason ?? "unknown"}`,
+      });
+    } catch {
+      // Audit must never block the outcome report.
+    }
+    return {
+      candidateHash: input.candidateHash,
+      batchId: input.batchId,
+      memoryId: null,
+      disposition: "rejected",
+      rejectionReason: admission.reason,
+      piiRedactedCount: redaction.redacted,
+      canonicalKey: input.canonicalKey,
+      category: input.proposedCategory,
+      verified: false,
+      verifyIssues: [],
+    };
+  }
+
+  const vector = await embedWithRetry(deps.embedder, text);
+  // The store layer never validates vector shape (only importEntry does), so an
+  // empty/short vector would land silently and poison retrieval — assert here.
+  if (!Array.isArray(vector) || vector.length === 0) {
+    throw new Error(`embedding returned an empty vector for candidate ${input.candidateHash.slice(0, 12)}`);
+  }
+  const language = detectLang(text);
+  const fts_text = tokenizeFts(text, language);
+
+  let metadata = buildStructuredMetadata({
+    source: PIVOT_APPLY_SOURCE,
+    tags: [...input.tags, "pivot-apply", `batch:${input.batchId}`],
+    capture: PIVOT_APPLY_CAPTURE,
+    category: input.proposedCategory,
+    canonicalKey: input.canonicalKey,
+    extra: {
+      anchor: input.anchor,
+      pivotApply: {
+        kind: input.kind,
+        evidence: input.evidence,
+        sessionId: input.sessionId,
+        reportId: input.reportId,
+        candidateHash: input.candidateHash,
+        batchId: input.batchId,
+      },
+    },
+  });
+  {
+    // Same post-build patch pattern persistMemory uses for confidence/PII.
+    const parsed = JSON.parse(metadata) as Record<string, unknown>;
+    parsed.confidence = assignDefaultConfidence(PIVOT_APPLY_SOURCE);
+    if (piiScan.hasPII) {
+      parsed.piiWarning = {
+        summary: piiScan.summary,
+        detections: piiScan.detections.length,
+        severity: piiScan.detections.some((d) => d.severity === "high") ? "high" : "medium",
+        ...(redaction.redacted > 0 ? { redacted: redaction.redacted } : {}),
+      };
+    }
+    metadata = JSON.stringify(parsed);
+  }
+
+  const { entry, disposition, conflictId } = await writeDurableEntry(deps, {
+    text,
+    vector,
+    category: input.proposedCategory,
+    scope: PIVOT_SCOPE,
+    importance: input.importance,
+    metadata,
+    canonicalKey: input.canonicalKey,
+    source: PIVOT_APPLY_SOURCE,
+    language,
+    fts_text,
+  });
+
+  try {
+    deps.auditLogger?.log({
+      operation: "store",
+      scope: PIVOT_SCOPE,
+      memoryId: entry.id,
+      actor: PIVOT_APPLY_SOURCE,
+      details: `pivot-apply ${input.kind} ${input.candidateHash.slice(0, 12)} → ${disposition}`,
+    });
+  } catch {
+    // Audit must never block the outcome report.
+  }
+
+  // Post-write verification, synchronous on purpose: a bulk apply needs the
+  // verdict inside the ledger row, not on a console later. Deliberately NOT
+  // verifyWrite(): its checkEntry expects scope/importance inside metadata,
+  // which buildStructuredMetadata-shaped rows never carry (a pre-existing
+  // mismatch its async caller silently swallows). This check asserts the claim
+  // that matters instead: the row at entry.id holds THIS candidate's content.
+  let verified = true;
+  let verifyIssues: string[] = [];
+  if ((disposition === "stored" || disposition === "updated") && deps.store.get) {
+    verifyIssues = await verifyPivotRow(deps.store.get.bind(deps.store), entry.id, {
+      candidateHash: input.candidateHash,
+      text,
+    });
+    verified = verifyIssues.length === 0;
+  }
+
+  return {
+    candidateHash: input.candidateHash,
+    batchId: input.batchId,
+    memoryId: entry.id,
+    disposition,
+    ...(conflictId ? { conflictId } : {}),
+    piiRedactedCount: redaction.redacted,
+    canonicalKey: input.canonicalKey,
+    category: input.proposedCategory,
+    verified,
+    verifyIssues,
+  };
+}
+
+function normalizeForCompare(text: string): string {
+  return text.replace(/\s+/g, " ").trim();
+}
+
+/** Read the row back and assert it is this candidate's write, not merely a well-formed row. */
+async function verifyPivotRow(
+  get: (id: string) => Promise<{ text: string; scope: string; metadata: string } | null>,
+  id: string,
+  expected: { candidateHash: string; text: string },
+): Promise<string[]> {
+  const issues: string[] = [];
+  let row: { text: string; scope: string; metadata: string } | null = null;
+  try {
+    row = await get(id);
+  } catch (err) {
+    return [`readback threw: ${(err as Error).message}`];
+  }
+  if (!row) return ["row missing on readback"];
+  if (row.scope !== PIVOT_SCOPE) issues.push(`scope is "${row.scope}", expected "${PIVOT_SCOPE}"`);
+  if (normalizeForCompare(row.text) !== normalizeForCompare(expected.text)) {
+    issues.push("text on readback differs from written text");
+  }
+  try {
+    const meta = JSON.parse(row.metadata || "{}") as Record<string, unknown>;
+    const pivotApply = meta.pivotApply as Record<string, unknown> | undefined;
+    if (pivotApply?.candidateHash !== expected.candidateHash) {
+      issues.push("pivotApply.candidateHash on readback does not match this candidate");
+    }
+  } catch {
+    issues.push("metadata unparseable on readback");
+  }
+  return issues;
+}
+
+export interface PivotBatchOptions {
+  /** Dispositions that do NOT halt the batch. Default: only "stored". */
+  expectedDispositions?: readonly PivotApplyDisposition[];
+}
+
+/**
+ * Persist a physical batch sequentially, fail-closed.
+ *
+ * Pre-screening guarantees the incoming batch has no key collisions with the
+ * live pool (R5) and no identical-text duplicates (R3 folding), so every
+ * disposition other than "stored" is unexpected by default and halts the batch
+ * immediately. Rows already written before the halt stay in the store — the
+ * caller reconciles via the returned outcomes (ledger) and the table snapshot
+ * taken before the batch.
+ */
+export async function persistPivotBatch(
+  deps: PersistMemoryDeps,
+  batchId: string,
+  rawInputs: readonly unknown[],
+  options?: PivotBatchOptions,
+): Promise<PivotBatchReport> {
+  const expected = new Set<PivotApplyDisposition>(options?.expectedDispositions ?? ["stored"]);
+  const outcomes: PivotApplyOutcome[] = [];
+  let written = 0;
+
+  for (const raw of rawInputs) {
+    // Halt before the write when the row belongs to a different batch: the
+    // ledger keys every side effect by batchId, so a mixed batch would corrupt
+    // snapshot/rollback bookkeeping.
+    const claimed =
+      typeof raw === "object" && raw !== null && "batchId" in raw
+        ? String((raw as Record<string, unknown>).batchId)
+        : "(missing)";
+    if (claimed !== batchId) {
+      return {
+        batchId,
+        total: rawInputs.length,
+        written,
+        halted: true,
+        haltReason: `candidate claims batchId "${claimed}" but batch is "${batchId}"`,
+        outcomes,
+      };
+    }
+
+    const outcome = await persistPivotCandidate(deps, raw);
+    outcomes.push(outcome);
+    if (outcome.memoryId !== null && (outcome.disposition === "stored" || outcome.disposition === "updated")) {
+      written += 1;
+    }
+
+    if (!expected.has(outcome.disposition)) {
+      return {
+        batchId,
+        total: rawInputs.length,
+        written,
+        halted: true,
+        haltReason: `unexpected disposition "${outcome.disposition}" for candidate ${outcome.candidateHash.slice(0, 12)}`,
+        outcomes,
+      };
+    }
+    if (!outcome.verified && (outcome.disposition === "stored" || outcome.disposition === "updated")) {
+      return {
+        batchId,
+        total: rawInputs.length,
+        written,
+        halted: true,
+        haltReason: `post-write verification failed for ${outcome.candidateHash.slice(0, 12)}: ${outcome.verifyIssues.join("; ")}`,
+        outcomes,
+      };
+    }
+  }
+
+  return {
+    batchId,
+    total: rawInputs.length,
+    written,
+    halted: false,
+    outcomes,
+  };
+}

package/src/preference-matcher.ts

@@ -15,6 +15,9 @@ import type { MemoryStore, MemoryEntry } from "./store.js";
 import type { Embedder } from "./embedder.js";
 import type { LLMClient } from "./llm-client.js";
 import { logInfo } from "./stderr-log.js";
+import { isActiveMemory, parseEvolution } from "./memory-evolution.js";
+import { archiveBeliefVersion, buildSupersedingBeliefMetadata } from "./belief-history.js";
+import { detectLang, tokenizeFts } from "./language-hook.js";
 
 // ============================================================================
 // Configuration
@@ -77,8 +80,13 @@ export async function matchPreference(
     [scope],
   );
 
-  // Filter to only preferences category
-  const prefMatches = candidates.filter(c => c.entry.category === "preferences");
+  // Filter to only preferences category. store.vectorSearch() filters by scope only, so
+  // superseded belief-history rows come back here too — matching against one would either
+  // silently drop the incoming preference as a duplicate of a belief no longer held, or
+  // merge new text into an archived version.
+  const prefMatches = candidates.filter(
+    c => c.entry.category === "preferences" && isActiveMemory(c.entry.metadata),
+  );
 
   if (prefMatches.length === 0) {
     return { action: "create", reason: "no-similar-preferences" };
@@ -132,6 +140,7 @@ export async function applyPreferenceMatch(
   result: PreferenceMatchResult,
   store: MemoryStore,
   scope: string,
+  deps: { embedder: Pick<Embedder, "embedPassage"> },
 ): Promise<{ handled: boolean; entry?: MemoryEntry }> {
   if (result.action === "skip") {
     logInfo(`[INFO] Preference skipped (duplicate): ${result.reason}`);
@@ -139,9 +148,34 @@ export async function applyPreferenceMatch(
   }
 
   if (result.action === "merge" && result.mergeTargetId && result.mergedText) {
+    const existing = await store.getById(result.mergeTargetId);
+    if (!existing) return { handled: false };
+
+    // Merging rewrites the target's text, which is a belief change like any other —
+    // archive the version being replaced before overwriting it.
+    const now = Date.now();
+    const previousEvo = parseEvolution(existing.metadata, existing.timestamp);
+    const { historyId } = await archiveBeliefVersion({ store, embedder: deps.embedder }, existing, { now });
+
+    // The old code updated `text` alone, leaving vector and fts_text describing the
+    // pre-merge wording — so a merged preference stayed unfindable by its own new text.
+    const language = detectLang(result.mergedText);
     const updated = await store.update(
       result.mergeTargetId,
-      { text: result.mergedText },
+      {
+        text: result.mergedText,
+        vector: await deps.embedder.embedPassage(result.mergedText),
+        metadata: buildSupersedingBeliefMetadata(
+          existing.metadata,
+          previousEvo,
+          historyId,
+          now,
+          `preference-merge:${result.reason}`,
+        ),
+        timestamp: now,
+        language,
+        fts_text: tokenizeFts(result.mergedText, language),
+      },
       [scope],
     );
     if (updated) {

package/src/query-expander.ts

@@ -10,8 +10,8 @@
  * v1.3: P0.3 — User-level alias-map (data/alias-map.json) for short query expansion.
  */
 
-import { existsSync, readFileSync } from "fs";
-import { join } from "path";
+import { existsSync, mkdirSync, readFileSync, writeFileSync } from "fs";
+import { dirname, join } from "path";
 import { resolveQueryEntities } from "./entity-resolver.js";
 
 interface SynonymEntry {
@@ -89,9 +89,14 @@ export function resetAliasMapCache(): void {
   aliasMapCache = null;
 }
 
+/** Resolve the alias-map file path currently in effect. */
+export function aliasMapFilePath(): string {
+  return aliasMapPath ?? join(process.cwd(), "data", "alias-map.json");
+}
+
 function loadAliasMap(): AliasEntry[] {
   if (aliasMapCache !== null) return aliasMapCache;
-  const filePath = aliasMapPath ?? join(process.cwd(), "data", "alias-map.json");
+  const filePath = aliasMapFilePath();
   try {
     if (existsSync(filePath)) {
       const raw = JSON.parse(readFileSync(filePath, "utf-8"));
@@ -112,6 +117,75 @@ function loadAliasMap(): AliasEntry[] {
   return aliasMapCache;
 }
 
+// ---------------------------------------------------------------------------
+// User alias management (P1-B manage_alias tool 后端)
+//
+// 这些别名只作用于 BM25 通道(expandQuery 在 hybridRetrieval 内部被调用),
+// 刻意不进 vector embedding——随手加的用户别名若改变语义向量,排查
+// retrieval drift 会非常困难。全通道的 builtin 规则在 src/aliases.ts。
+// ---------------------------------------------------------------------------
+
+/** 写入校验:trigger 最短 2 字符(防高频泛词),expansions 最多 8 个。 */
+export const USER_ALIAS_TRIGGER_MIN = 2;
+export const USER_ALIAS_EXPANSIONS_MAX = 8;
+
+export type AliasUpsertResult =
+  | { ok: true; action: "added" | "updated"; entry: AliasEntry }
+  | { ok: false; reason: string };
+
+/** 新增/更新用户别名(写 alias-map.json;校验后缓存即时失效)。 */
+export function upsertUserAlias(trigger: string, expansions: string[]): AliasUpsertResult {
+  const t = trigger.trim();
+  const exp = [...new Set(expansions.map(e => e.trim()).filter(e => e.length > 0))];
+
+  if (t.length < USER_ALIAS_TRIGGER_MIN) {
+    return { ok: false, reason: `trigger 太短(<${USER_ALIAS_TRIGGER_MIN} 字符)——高频泛词会污染所有查询` };
+  }
+  if (exp.length === 0) {
+    return { ok: false, reason: "expansions 为空" };
+  }
+  if (exp.length > USER_ALIAS_EXPANSIONS_MAX) {
+    return { ok: false, reason: `expansions 超过 ${USER_ALIAS_EXPANSIONS_MAX} 个——过长扩展会冲淡 BM25 原查询词权重` };
+  }
+  if (exp.every(e => e === t)) {
+    return { ok: false, reason: "expansions 与 trigger 相同,无扩展意义" };
+  }
+
+  const rules = loadAliasMap();
+  const idx = rules.findIndex(r => r.trigger === t);
+  const entry: AliasEntry = { trigger: t, expansions: exp };
+  const next = idx >= 0 ? rules.map((r, i) => (i === idx ? entry : r)) : [...rules, entry];
+
+  const path = aliasMapFilePath();
+  mkdirSync(dirname(path), { recursive: true });
+  writeFileSync(path, JSON.stringify(next, null, 2) + "\n", "utf-8");
+  aliasMapCache = null;
+  return { ok: true, action: idx >= 0 ? "updated" : "added", entry };
+}
+
+/** 删除用户别名;返回是否真的删了。 */
+export function removeUserAlias(trigger: string): boolean {
+  const rules = loadAliasMap();
+  const next = rules.filter(r => r.trigger !== trigger.trim());
+  if (next.length === rules.length) return false;
+  const path = aliasMapFilePath();
+  mkdirSync(dirname(path), { recursive: true });
+  writeFileSync(path, JSON.stringify(next, null, 2) + "\n", "utf-8");
+  aliasMapCache = null;
+  return true;
+}
+
+/** 列出用户别名规则。 */
+export function listUserAliases(): AliasEntry[] {
+  return [...loadAliasMap()];
+}
+
+/** 检视:query 命中了哪些用户别名(BM25-only 通道)。 */
+export function explainUserAliases(query: string): AliasEntry[] {
+  const lower = query.toLowerCase();
+  return loadAliasMap().filter(e => lower.includes(e.trigger.toLowerCase()));
+}
+
 /**
  * Expand a query by appending synonym terms from the dictionary.
  * Returns the original query with additional terms appended.
@@ -123,7 +197,8 @@ export function expandQuery(query: string): string {
   // Entity resolution pre-pass: normalize aliases before synonym expansion
   const resolved = resolveQueryEntities(query);
   const lower = resolved.toLowerCase();
-  const additions = new Set<string>();
+  const synonymAdditions = new Set<string>();
+  const aliasAdditions = new Set<string>();
 
   // Static synonym map (built-in)
   for (const entry of SYNONYM_MAP) {
@@ -138,7 +213,7 @@ export function expandQuery(query: string): string {
     if (cnMatch || enMatch) {
       for (const exp of entry.expansions) {
         if (!lower.includes(exp.toLowerCase())) {
-          additions.add(exp);
+          synonymAdditions.add(exp);
         }
       }
     }
@@ -151,15 +226,21 @@ export function expandQuery(query: string): string {
     if (lower.includes(triggerLower)) {
       for (const exp of alias.expansions) {
         if (!lower.includes(exp.toLowerCase())) {
-          additions.add(exp);
+          aliasAdditions.add(exp);
         }
       }
     }
   }
 
-  if (additions.size === 0) return resolved;
+  if (aliasAdditions.size === 0 && synonymAdditions.size === 0) return resolved;
 
-  // Cap expansion terms to prevent query bloat
-  const limited = [...additions].slice(0, MAX_EXPANSION_TERMS);
+  // Cap expansion terms to prevent query bloat. User aliases rank first inside
+  // the cap — 用户为这个 trigger 显式配置的 canonical 名,不应被泛化同义词
+  // 挤出 MAX_EXPANSION_TERMS(P1-B:同义词"挂"曾把 alias 的实体名挤掉)。
+  const merged = [
+    ...aliasAdditions,
+    ...[...synonymAdditions].filter(t => !aliasAdditions.has(t)),
+  ];
+  const limited = merged.slice(0, MAX_EXPANSION_TERMS);
   return `${resolved} ${limited.join(" ")}`;
 }