@remnic/core 9.3.597 → 9.3.598

package/src/direct-answer.ts

@@ -18,12 +18,9 @@
  * Not wired into retrieval yet — see slice 3.
  */
 
-import type { MemoryFile, MemoryStatus } from "./types.js";
+import { normalizeRecallTokenSet, normalizeRecallTokens } from "./recall-tokenization.js";
 import type { TrustZoneName } from "./trust-zones.js";
-import {
-  countRecallTokenOverlap,
-  normalizeRecallTokens,
-} from "./recall-tokenization.js";
+import type { MemoryFile, MemoryStatus } from "./types.js";
 
 /**
  * Caller-supplied candidate.
@@ -105,9 +102,246 @@ export const FILTER_LABELS = {
   belowTokenOverlapFloor: "below-token-overlap-floor",
 } as const;
 
+const PROMPT_RECALL_WORDS = new Set([
+  "what",
+  "who",
+  "where",
+  "when",
+  "why",
+  "how",
+  "is",
+  "are",
+  "was",
+  "were",
+  "do",
+  "does",
+  "did",
+  "find",
+  "get",
+  "show",
+  "search",
+  "lookup",
+  "recall",
+  "remember",
+  "list",
+  "status",
+  "include",
+  "tell",
+  "me",
+  "give",
+  "about",
+  "please",
+  "the",
+  "and",
+  "for",
+  "with",
+  "from",
+  "into",
+  "that",
+  "this",
+  "найди",
+  "найти",
+  "поиск",
+  "покажи",
+  "статус",
+  "включи",
+]);
+
 interface ScoredCandidate {
   candidate: DirectAnswerCandidate;
   tokenOverlap: number;
+  requiredTokenMismatch: boolean;
+}
+
+function hasUnsegmentableRecallChar(token: string): boolean {
+  if (token.includes("ー") || token.includes("ｰ")) return true;
+  return /[\p{Script=Han}\p{Script=Hiragana}\p{Script=Katakana}]/u.test(token);
+}
+
+function requiredCjkPhraseTokens(query: string): string[] {
+  const phrases = new Set<string>();
+  let segment = "";
+
+  const addPhrase = (phrase: string) => {
+    if ([...phrase].length >= 4) {
+      phrases.add(phrase);
+    }
+  };
+
+  const flushSegment = () => {
+    let buffered = "";
+    for (const run of segment.split(/\s+/)) {
+      if (!run) continue;
+      if ([...run].length >= 4) {
+        addPhrase(buffered);
+        buffered = "";
+        addPhrase(run);
+        continue;
+      }
+      buffered += run;
+    }
+    addPhrase(buffered);
+    segment = "";
+  };
+
+  for (const ch of query.toLowerCase().normalize("NFC")) {
+    if (hasUnsegmentableRecallChar(ch)) {
+      segment += ch;
+      continue;
+    }
+    if (/\p{M}/u.test(ch) && segment.length > 0 && !/\s$/u.test(segment)) {
+      segment += ch;
+      continue;
+    }
+    if (/\s/u.test(ch)) {
+      segment += " ";
+      continue;
+    }
+    flushSegment();
+  }
+  flushSegment();
+
+  return [...phrases];
+}
+
+function requiredMixedScriptTokens(query: string): string[] {
+  const required = new Set<string>();
+  const parts: string[] = [];
+  let segment = "";
+
+  const flushSegment = () => {
+    if (segment) {
+      parts.push(segment);
+    }
+    segment = "";
+  };
+
+  const segmentableRecallTokens = (value: string) => {
+    const tokens = new Set<string>();
+    let segment = "";
+    const flushSegmentableSegment = () => {
+      for (const token of normalizeRecallTokenSet(segment, [], { minTokenLength: 1 })) {
+        tokens.add(token);
+      }
+      segment = "";
+    };
+
+    for (const ch of value) {
+      if (/[\p{L}\p{N}]/u.test(ch) && !hasUnsegmentableRecallChar(ch)) {
+        segment += ch;
+        continue;
+      }
+      if (/\p{M}/u.test(ch) && segment.length > 0) {
+        segment += ch;
+        continue;
+      }
+      flushSegmentableSegment();
+    }
+    flushSegmentableSegment();
+    return tokens;
+  };
+
+  const hasRequiredSegmentableToken = (value: string) => {
+    return segmentableRecallTokens(value).size > 0;
+  };
+
+  const hasBoundarySegmentableToken = (value: string) => {
+    for (const token of segmentableRecallTokens(value)) {
+      if (PROMPT_RECALL_WORDS.has(token)) {
+        continue;
+      }
+      const hasNonAsciiCodepoint = [...token].some((ch) => (ch.codePointAt(0) ?? 0) > 0x7f);
+      if (token.length >= 3 || /\p{N}/u.test(token) || hasNonAsciiCodepoint) {
+        return true;
+      }
+    }
+    return false;
+  };
+
+  const addRequiredTokens = (value: string) => {
+    for (const token of normalizeRecallTokenSet(value, [], { minTokenLength: 1 })) {
+      required.add(token);
+    }
+  };
+
+  for (const ch of query.toLowerCase().normalize("NFC")) {
+    if (/[\p{L}\p{N}\p{M}]/u.test(ch) || hasUnsegmentableRecallChar(ch)) {
+      segment += ch;
+      continue;
+    }
+    flushSegment();
+  }
+  flushSegment();
+
+  for (const part of parts) {
+    if (hasUnsegmentableRecallChar(part) && hasRequiredSegmentableToken(part)) {
+      addRequiredTokens(part);
+    }
+  }
+
+  for (let i = 0; i < parts.length - 1; i += 1) {
+    const current = parts[i];
+    const next = parts[i + 1];
+    const currentHasUnsegmentable = hasUnsegmentableRecallChar(current);
+    const nextHasUnsegmentable = hasUnsegmentableRecallChar(next);
+    if (currentHasUnsegmentable === nextHasUnsegmentable) {
+      continue;
+    }
+    const segmentablePart = currentHasUnsegmentable ? next : current;
+    if (!hasBoundarySegmentableToken(segmentablePart)) {
+      continue;
+    }
+    addRequiredTokens(current);
+    addRequiredTokens(next);
+  }
+
+  for (let i = 1; i < parts.length - 1; i += 1) {
+    const prev = parts[i - 1];
+    const current = parts[i];
+    const next = parts[i + 1];
+    if (!hasUnsegmentableRecallChar(prev) || hasUnsegmentableRecallChar(current) || !hasUnsegmentableRecallChar(next)) {
+      continue;
+    }
+    if (!hasRequiredSegmentableToken(current)) {
+      continue;
+    }
+    addRequiredTokens(prev);
+    addRequiredTokens(current);
+    addRequiredTokens(next);
+  }
+
+  return [...required];
+}
+
+function requiredSegmentableUnicodeTokens(queryTokens: Set<string>): string[] {
+  const segmentableTokens = [...queryTokens].filter((token) => !hasUnsegmentableRecallChar(token));
+  const hasSegmentableUnicodeToken = segmentableTokens.some((token) =>
+    [...token].some((ch) => (ch.codePointAt(0) ?? 0) > 0x7f)
+  );
+  if (!hasSegmentableUnicodeToken) {
+    return [];
+  }
+  return segmentableTokens.filter((token) => {
+    if (PROMPT_RECALL_WORDS.has(token)) {
+      return false;
+    }
+    const hasNonAsciiCodepoint = [...token].some((ch) => (ch.codePointAt(0) ?? 0) > 0x7f);
+    if (hasNonAsciiCodepoint) {
+      return true;
+    }
+    if (!/^[a-z0-9]+$/u.test(token)) {
+      return false;
+    }
+    return token.length >= 3 || /\p{N}/u.test(token);
+  });
+}
+
+function countTokenOverlap(queryTokens: Set<string>, valueTokens: Set<string>): number {
+  let matches = 0;
+  for (const token of queryTokens) {
+    if (valueTokens.has(token)) matches += 1;
+  }
+  return matches;
 }
 
 /**
@@ -123,9 +357,7 @@ interface ScoredCandidate {
  *   6. top two candidates within ambiguityMargin → "ambiguous"
  *   7. otherwise → "eligible"
  */
-export function isDirectAnswerEligible(
-  input: DirectAnswerInput,
-): DirectAnswerResult {
+export function isDirectAnswerEligible(input: DirectAnswerInput): DirectAnswerResult {
   const { query, candidates, config, queryEntityRefs } = input;
 
   if (!config.enabled) {
@@ -164,17 +396,13 @@ export function isDirectAnswerEligible(
     return status === "active";
   });
 
-  working = applyFilter(working, filteredBy, FILTER_LABELS.notTrustedZone, (c) =>
-    c.trustZone === "trusted",
-  );
+  working = applyFilter(working, filteredBy, FILTER_LABELS.notTrustedZone, (c) => c.trustZone === "trusted");
 
   working = applyFilter(
     working,
     filteredBy,
     FILTER_LABELS.ineligibleTaxonomyBucket,
-    (c) =>
-      c.taxonomyBucket !== null &&
-      config.eligibleTaxonomyBuckets.includes(c.taxonomyBucket),
+    (c) => c.taxonomyBucket !== null && config.eligibleTaxonomyBuckets.includes(c.taxonomyBucket)
   );
 
   working = applyFilter(working, filteredBy, FILTER_LABELS.belowImportanceFloor, (c) => {
@@ -201,13 +429,24 @@ export function isDirectAnswerEligible(
   }
 
   const scored: ScoredCandidate[] = working.map((candidate) => {
-    const searchable =
-      `${candidate.memory.frontmatter.tags?.join(" ") ?? ""} ${candidate.memory.content}`.trim();
-    const matches = countRecallTokenOverlap(queryTokens, searchable);
-    return { candidate, tokenOverlap: matches / queryTokens.size };
+    const searchable = `${candidate.memory.frontmatter.tags?.join(" ") ?? ""} ${candidate.memory.content}`.trim();
+    const searchableTokens = normalizeRecallTokenSet(searchable);
+    const requiredSearchableTokens = normalizeRecallTokenSet(searchable, [], { minTokenLength: 1 });
+    const requiredPhrases = requiredCjkPhraseTokens(query);
+    const requiredMixedTokens = requiredMixedScriptTokens(query);
+    const requiredUnicodeTokens = requiredSegmentableUnicodeTokens(queryTokens);
+    const hasRequiredPhrase =
+      requiredPhrases.length === 0 || requiredPhrases.every((token) => searchableTokens.has(token));
+    const hasRequiredMixedTokens =
+      requiredMixedTokens.length === 0 || requiredMixedTokens.every((token) => requiredSearchableTokens.has(token));
+    const hasRequiredUnicodeTokens =
+      requiredUnicodeTokens.length === 0 || requiredUnicodeTokens.every((token) => searchableTokens.has(token));
+    const requiredTokenMismatch = !hasRequiredPhrase || !hasRequiredMixedTokens || !hasRequiredUnicodeTokens;
+    const matches = requiredTokenMismatch ? 0 : countTokenOverlap(queryTokens, searchableTokens);
+    return { candidate, tokenOverlap: matches / queryTokens.size, requiredTokenMismatch };
   });
 
-  const overlapSurvivors = scored.filter((s) => s.tokenOverlap >= config.tokenOverlapFloor);
+  const overlapSurvivors = scored.filter((s) => !s.requiredTokenMismatch && s.tokenOverlap >= config.tokenOverlapFloor);
   if (overlapSurvivors.length < scored.length) {
     filteredBy.push(FILTER_LABELS.belowTokenOverlapFloor);
   }
@@ -252,7 +491,7 @@ function applyFilter(
   working: DirectAnswerCandidate[],
   filteredBy: string[],
   label: string,
-  keep: (c: DirectAnswerCandidate) => boolean,
+  keep: (c: DirectAnswerCandidate) => boolean
 ): DirectAnswerCandidate[] {
   const before = working.length;
   const next = working.filter(keep);

package/src/recall-query-policy.ts

@@ -1,3 +1,5 @@
+import { normalizeRecallTokens } from "./recall-tokenization.js";
+
 export type RecallPromptShape = "standard" | "instruction_heavy";
 export type CronConversationRecallMode = "auto" | "always" | "never";
 export type RecallBudgetMode = "full" | "minimal";
@@ -16,7 +18,7 @@ export interface RecallQueryPolicyResult {
   retrievalBudgetMode: RecallBudgetMode;
 }
 
-const DEFAULT_STOPWORDS = new Set([
+const DEFAULT_STOPWORDS = [
   "the",
   "and",
   "for",
@@ -72,7 +74,9 @@ const DEFAULT_STOPWORDS = new Set([
   "data",
   "gathering",
   "context",
-]);
+];
+const MAX_COMPACT_TOKENS_PER_SOURCE_TERM = 8;
+const COMPACT_IDENTIFIER_RE = /^[a-z0-9]+(?:[:_-][a-z0-9]+)+$/i;
 
 function collapseWhitespace(text: string): string {
   return text.replace(/\s+/g, " ").trim();
@@ -85,6 +89,24 @@ function stripFilesystemLikePaths(text: string): string {
     .replace(/(?:^|\s)([A-Za-z]:\\[^\s)]+)(?=\s|$)/g, " ");
 }
 
+function trimCompactSourceTerm(term: string): string {
+  return term.replace(/^[^\p{L}\p{N}\p{M}]+|[^\p{L}\p{N}\p{M}]+$/gu, "");
+}
+
+function splitCompactSourceTerm(term: string): string[] {
+  return term
+    .split(/[^\p{L}\p{N}\p{M}:_-]+/gu)
+    .map((part) => trimCompactSourceTerm(part))
+    .filter((part) => part.length > 0);
+}
+
+function capCompactSourceTermTokens(tokens: string[]): string[] {
+  if (tokens.length <= MAX_COMPACT_TOKENS_PER_SOURCE_TERM) return tokens;
+  const last = tokens.at(-1);
+  if (!last) return tokens.slice(0, MAX_COMPACT_TOKENS_PER_SOURCE_TERM);
+  return [...tokens.slice(0, MAX_COMPACT_TOKENS_PER_SOURCE_TERM - 1), last];
+}
+
 function isBulletOrNumberedLine(line: string): boolean {
   if (line.startsWith("-") || line.startsWith("*")) {
     return true;
@@ -112,16 +134,14 @@ function scoreInstructionHeavyShape(prompt: string): number {
   const headingLineCount = lines.filter(
     (line) =>
       /^(goal|output format|tone rules|grounding rules|data gathering|date computation|crm context|follow-up|social|current time|return)\b/i.test(
-        line,
-      ) || /^[A-Z][A-Z\s/-]{4,}:$/.test(line),
+        line
+      ) || /^[A-Z][A-Z\s/-]{4,}:$/.test(line)
   ).length;
   const bulletLineCount = lines.filter((line) => isBulletOrNumberedLine(line)).length;
   const longLineCount = lines.filter((line) => line.length >= 180).length;
-  const hasPathDensity =
-    (prompt.match(/(?:~\/|\/Users\/|[A-Za-z]:\\)/g)?.length ?? 0) >= 2;
+  const hasPathDensity = (prompt.match(/(?:~\/|\/Users\/|[A-Za-z]:\\)/g)?.length ?? 0) >= 2;
   const hasImperativeDensity =
-    (prompt.match(/\b(run|extract|read|parse|determine|include|omit|skip)\b/gi)?.length ?? 0) >=
-    8;
+    (prompt.match(/\b(run|extract|read|parse|determine|include|omit|skip)\b/gi)?.length ?? 0) >= 8;
 
   let score = 0;
   if (lineCount >= 24) score += 2;
@@ -139,27 +159,29 @@ export function classifyRecallPromptShape(prompt: string): RecallPromptShape {
 }
 
 function tokenizeForCompactQuery(text: string): string[] {
-  const raw = text
-    .toLowerCase()
-    .replace(/[^a-z0-9\s:_-]+/g, " ")
-    .split(/\s+/)
-    .filter((token) => token.length >= 3);
-  const deduped: string[] = [];
+  const tokens: string[] = [];
   const seen = new Set<string>();
-  for (const token of raw) {
-    if (DEFAULT_STOPWORDS.has(token)) continue;
-    if (seen.has(token)) continue;
+  const addToken = (token: string) => {
+    if (token.length === 0 || seen.has(token)) return;
     seen.add(token);
-    deduped.push(token);
+    tokens.push(token);
+  };
+
+  for (const rawTerm of text.split(/\s+/)) {
+    for (const sourceTerm of splitCompactSourceTerm(trimCompactSourceTerm(rawTerm))) {
+      if (COMPACT_IDENTIFIER_RE.test(sourceTerm)) {
+        addToken(sourceTerm.toLowerCase());
+        continue;
+      }
+      for (const token of capCompactSourceTermTokens(normalizeRecallTokens(sourceTerm, DEFAULT_STOPWORDS))) {
+        addToken(token);
+      }
+    }
   }
-  return deduped;
+  return tokens;
 }
 
-function buildInstructionHeavyQuery(
-  prompt: string,
-  tokenCap: number,
-  maxChars: number,
-): string {
+function buildInstructionHeavyQuery(prompt: string, tokenCap: number, maxChars: number): string {
   const cleaned = stripFilesystemLikePaths(prompt);
   const tokens = tokenizeForCompactQuery(cleaned).slice(0, Math.max(8, tokenCap));
   const joined = tokens.join(" ");
@@ -182,7 +204,7 @@ function buildStandardQuery(prompt: string, maxChars: number): string {
 export function buildRecallQueryPolicy(
   prompt: string,
   sessionKey: string | undefined,
-  cfg: RecallQueryPolicyConfig,
+  cfg: RecallQueryPolicyConfig
 ): RecallQueryPolicyResult {
   const normalizedPrompt = collapseWhitespace(prompt);
   const isCron = (sessionKey ?? "").includes(":cron:");
@@ -207,8 +229,8 @@ export function buildRecallQueryPolicy(
     cfg.cronConversationRecallMode === "never"
       ? true
       : cfg.cronConversationRecallMode === "always"
-      ? false
-      : promptShape === "instruction_heavy";
+        ? false
+        : promptShape === "instruction_heavy";
 
   const retrievalBudgetMode = promptShape === "instruction_heavy" ? "minimal" : "full";
 

package/src/recall-tokenization.ts

@@ -1,32 +1,142 @@
-export function normalizeRecallTokens(value: string, extraStopWords: string[] = []): string[] {
-  const stopWords = new Set([
-    "the",
-    "and",
-    "for",
-    "with",
-    "from",
-    "into",
-    "that",
-    "this",
-    "why",
-    "did",
-    ...extraStopWords,
-  ]);
-
-  return value
+export interface NormalizeRecallTokenOptions {
+  minTokenLength?: number;
+}
+
+const DEFAULT_RECALL_STOP_WORDS = ["the", "and", "for", "with", "from", "into", "that", "this", "why", "did"];
+
+function isUnsegmentableRecallChar(char: string): boolean {
+  if (char === "ー" || char === "ｰ") return true;
+  return /[\p{Script=Han}\p{Script=Hiragana}\p{Script=Katakana}]/u.test(char);
+}
+
+function isRecallCombiningMark(char: string): boolean {
+  return /\p{M}/u.test(char);
+}
+
+function buildRecallStopWords(extraStopWords: string[]): Set<string> {
+  return new Set([...DEFAULT_RECALL_STOP_WORDS, ...extraStopWords.map((word) => word.toLowerCase())]);
+}
+
+function shouldKeepRecallToken(token: string, minTokenLength: number, stopWords: Set<string>): boolean {
+  if (stopWords.has(token)) return false;
+  if (token.length >= minTokenLength) return true;
+  const hasNonAsciiCodepoint = [...token].some((ch) => (ch.codePointAt(0) ?? 0) > 0x7f);
+  return token.length >= 2 && hasNonAsciiCodepoint && /\p{L}/u.test(token);
+}
+
+function addUnsegmentableRecallSegment(tokens: Set<string>, segment: string, stopWords: Set<string>): void {
+  const chars = [...segment].filter((ch) => /[\p{L}\p{N}\p{M}]/u.test(ch) || isUnsegmentableRecallChar(ch));
+  for (const ch of chars) {
+    if (!stopWords.has(ch)) tokens.add(ch);
+  }
+  for (const size of [2, 3, 4]) {
+    if (chars.length < size) continue;
+    for (let index = 0; index <= chars.length - size; index += 1) {
+      const token = chars.slice(index, index + size).join("");
+      if (!stopWords.has(token)) tokens.add(token);
+    }
+  }
+  const whole = chars.join("");
+  if (whole.length > 3 && !stopWords.has(whole)) {
+    tokens.add(whole);
+  }
+}
+
+function isUnsegmentableRecallToken(token: string): boolean {
+  const chars = [...token].filter((ch) => /[\p{L}\p{N}\p{M}]/u.test(ch) || isUnsegmentableRecallChar(ch));
+  return (
+    chars.length > 0 &&
+    chars.some(isUnsegmentableRecallChar) &&
+    chars.every((ch) => isUnsegmentableRecallChar(ch) || isRecallCombiningMark(ch))
+  );
+}
+
+function addBridgedUnsegmentableRecallSegments(tokens: Set<string>, cleaned: string, stopWords: Set<string>): void {
+  let segment = "";
+  const flushSegment = () => {
+    addUnsegmentableRecallSegment(tokens, segment, stopWords);
+    segment = "";
+  };
+
+  for (const token of cleaned.split(/\s+/)) {
+    if (isUnsegmentableRecallToken(token)) {
+      segment += token;
+    } else {
+      flushSegment();
+    }
+  }
+  flushSegment();
+}
+
+export function normalizeRecallTokenSet(
+  value: string,
+  extraStopWords: string[] = [],
+  options: NormalizeRecallTokenOptions = {}
+): Set<string> {
+  const minTokenLength = Math.max(1, Math.floor(options.minTokenLength ?? 3));
+  const stopWords = buildRecallStopWords(extraStopWords);
+  const cleaned = value
     .toLowerCase()
-    .split(/[^a-z0-9]+/)
-    .map((token) => token.trim())
-    .filter((token) => token.length >= 3 && !stopWords.has(token));
+    .normalize("NFC")
+    .replace(/[^\p{L}\p{N}\p{M}\u30fc\uff70]+/gu, " ")
+    .trim();
+  if (cleaned.length === 0) return new Set();
+
+  const tokens = new Set<string>();
+  addBridgedUnsegmentableRecallSegments(tokens, cleaned, stopWords);
+  for (const token of cleaned.split(/\s+/)) {
+    if (!token) continue;
+    if ([...token].some(isUnsegmentableRecallChar)) {
+      let segment = "";
+      let unsegmentableSegment = "";
+      const flushSegment = () => {
+        if (shouldKeepRecallToken(segment, minTokenLength, stopWords)) {
+          tokens.add(segment);
+        }
+        segment = "";
+      };
+      const flushUnsegmentableSegment = () => {
+        addUnsegmentableRecallSegment(tokens, unsegmentableSegment, stopWords);
+        unsegmentableSegment = "";
+      };
+      for (const ch of token) {
+        if (!/[\p{L}\p{N}\p{M}]/u.test(ch) && !isUnsegmentableRecallChar(ch)) continue;
+        if (isUnsegmentableRecallChar(ch)) {
+          flushSegment();
+          unsegmentableSegment += ch;
+        } else if (isRecallCombiningMark(ch)) {
+          if (unsegmentableSegment.length > 0) {
+            unsegmentableSegment += ch;
+          } else {
+            segment += ch;
+          }
+        } else {
+          flushUnsegmentableSegment();
+          segment += ch;
+        }
+      }
+      flushUnsegmentableSegment();
+      flushSegment();
+      continue;
+    }
+    if (shouldKeepRecallToken(token, minTokenLength, stopWords)) {
+      tokens.add(token);
+    }
+  }
+  return tokens;
+}
+
+export function normalizeRecallTokens(value: string, extraStopWords: string[] = []): string[] {
+  return Array.from(normalizeRecallTokenSet(value, extraStopWords));
 }
 
 export function countRecallTokenOverlap(
   queryTokens: Set<string>,
   value: string | undefined,
-  extraStopWords: string[] = [],
+  extraStopWords: string[] = []
 ): number {
   if (!value) return 0;
-  const tokens = new Set(normalizeRecallTokens(value, extraStopWords));
+  const tokens = normalizeRecallTokenSet(value, extraStopWords);
   let matches = 0;
   for (const token of queryTokens) {
     if (tokens.has(token)) matches += 1;

package/src/temporal-index.ts

@@ -76,9 +76,7 @@ function tagIndexPath(memoryDir: string): string {
 
 function ensureStateDir(memoryDir: string): void {
   const dir = stateDir(memoryDir);
-  if (!fs.existsSync(dir)) {
-    fs.mkdirSync(dir, { recursive: true });
-  }
+  fs.mkdirSync(dir, { recursive: true });
 }
 
 function readJsonSafe<T>(filePath: string, fallback: T): T {
@@ -187,6 +185,13 @@ function lockOwnerIsRunning(owner: IndexLockOwner): boolean {
   return runningStartedAtMs <= owner.processStartedAtMs + INDEX_PROCESS_START_TOLERANCE_MS;
 }
 
+function lockIsFresh(lockInfo: fs.Stats, owner: IndexLockOwner | null): boolean {
+  const ownerCreatedAtMs =
+    typeof owner?.createdAt === "string" && owner.createdAt.length > 0 ? Date.parse(owner.createdAt) : Number.NaN;
+  const referenceMs = Number.isFinite(ownerCreatedAtMs) ? ownerCreatedAtMs : lockInfo.mtimeMs;
+  return Date.now() - referenceMs < INDEX_LOCK_STALE_MS;
+}
+
 function removeAbandonedIndexLock(lockDir: string): IndexLockCleanupResult {
   try {
     const info = fs.lstatSync(lockDir);
@@ -196,11 +201,14 @@ function removeAbandonedIndexLock(lockDir: string): IndexLockCleanupResult {
       return "removed";
     }
     const owner = readIndexLockOwner(lockDir);
-    if (owner !== null && lockOwnerIsRunning(owner)) return "wait";
-    if (owner === null && Date.now() - info.mtimeMs < INDEX_LOCK_STALE_MS) return "wait";
+    if (owner !== null) {
+      if (lockOwnerIsRunning(owner)) return "wait";
+    }
+    if (owner === null && lockIsFresh(info, null)) return "wait";
     fs.rmSync(lockDir, { recursive: true, force: true });
     return "removed";
-  } catch {
+  } catch (error) {
+    if ((error as NodeJS.ErrnoException)?.code === "ENOENT") return "removed";
     // Fail silently — indexes are advisory only
     return "blocked";
   }
@@ -217,6 +225,15 @@ function withIndexFileLock(filePath: string, update: () => void): void {
       acquired = true;
     } catch (error) {
       const code = (error as NodeJS.ErrnoException)?.code;
+      if (code === "ENOENT") {
+        try {
+          fs.mkdirSync(path.dirname(lockDir), { recursive: true });
+        } catch {
+          return;
+        }
+        sleepSync(INDEX_LOCK_POLL_MS);
+        continue;
+      }
       if (code !== "EEXIST") return;
       const cleanupResult = removeAbandonedIndexLock(lockDir);
       if (cleanupResult === "blocked") return;