mixdog 0.7.18 → 0.8.1

package/src/runtime/memory/lib/memory-recall-store.mjs

@@ -0,0 +1,644 @@
+const __mixdogMemoryStderrWrite = process.stderr.write.bind(process.stderr);
+function __mixdogMemoryLog(...args) {
+  if (process.env.MIXDOG_QUIET_MEMORY_LOG) return true;
+  return __mixdogMemoryStderrWrite(...args);
+}
+
+import { buildFtsQuery } from './memory-text-utils.mjs'
+import { VALID_CATEGORY, embeddingToSql } from './memory.mjs'
+import { freshnessFactor } from './memory-score.mjs'
+import { buildRecallScopeFilter } from './memory-recall-scope-filter.mjs'
+import { recallReadQuery } from './memory-recall-read-query.mjs'
+
+// Per-db cache of mv_hot_active populated state. The main recall path currently
+// uses entries directly; this guard remains for explicit useHotActive callers.
+const _MV_HOT_ACTIVE_TTL_MS = 60_000
+const _mvHotActiveCache = new WeakMap() // db → { populated: boolean, ts: number }
+const SEMANTIC_ONLY_MIN_SIM = 0.72
+const SHORT_QUERY_TOKEN_MAX = 2
+
+function buildExactTerms(query) {
+  const clean = String(query ?? '').replace(/\s+/g, ' ').trim()
+  if (!clean) return []
+  const terms = []
+  const add = (value) => {
+    const term = String(value ?? '').trim()
+      .replace(/^[^\p{L}\p{N}_./:-]+|[^\p{L}\p{N}_./:-]+$/gu, '')
+    if (!term) return
+    const hasIdentifierShape = /[_./:-]/.test(term)
+    if (!hasIdentifierShape && /^[\d\s]+$/.test(term)) return
+    const symbolCount = Array.from(term).length
+    if (!hasIdentifierShape && symbolCount < 2) return
+    terms.push(term.slice(0, 80))
+  }
+  if (clean.length <= 80) add(clean)
+  const tokens = clean.match(/[\p{L}\p{N}_./:-]+/gu) || []
+  for (const token of tokens) add(token)
+  for (let i = 0; i < tokens.length - 1; i++) {
+    add(`${tokens[i]} ${tokens[i + 1]}`)
+  }
+  return [...new Set(terms.map(t => t.toLowerCase()))].slice(0, 12)
+}
+
+function countQueryTokens(query) {
+  const clean = String(query ?? '').replace(/\s+/g, ' ').trim()
+  if (!clean) return 0
+  return (clean.match(/[\p{L}\p{N}_./:-]+/gu) || [])
+    .filter(token => String(token ?? '').trim().length > 0)
+    .length
+}
+
+function hasFullQueryTextMatch(query, row) {
+  const clean = String(query ?? '').replace(/\s+/g, ' ').trim().toLowerCase()
+  if (!clean || clean.length > 160) return false
+  const element = String(row?.element ?? '').toLowerCase()
+  const summary = String(row?.summary ?? '').toLowerCase()
+  const content = String(row?.content ?? '').toLowerCase()
+  return (element && element.includes(clean))
+    || (summary && summary.includes(clean))
+    || (content && content.includes(clean))
+}
+
+function hasQueryTokenCoverage(row, queryTokenCount) {
+  if (queryTokenCount <= SHORT_QUERY_TOKEN_MAX) return true
+  const hits = Number(row?.exact_hits)
+  return Number.isFinite(hits) && hits >= queryTokenCount
+}
+
+function exactTextBoost(query, row, exactHits) {
+  const clean = String(query ?? '').trim().toLowerCase()
+  if (!clean) return 0
+  const element = String(row?.element ?? '').toLowerCase()
+  const summary = String(row?.summary ?? '').toLowerCase()
+  const content = String(row?.content ?? '').toLowerCase()
+  let boost = 0
+  if (element && element.includes(clean)) boost += 0.035
+  if (summary && summary.includes(clean)) boost += 0.025
+  if (content && content.includes(clean)) boost += 0.015
+  const hits = Number(exactHits)
+  if (Number.isFinite(hits) && hits > 0) boost += Math.min(0.03, hits * 0.008)
+  return boost
+}
+
+async function _checkMvHotActivePopulated(db) {
+  const cached = _mvHotActiveCache.get(db)
+  const now = Date.now()
+  if (cached && now - cached.ts < _MV_HOT_ACTIVE_TTL_MS) return cached.populated
+  const r = await recallReadQuery(
+    db,
+    `SELECT relispopulated FROM pg_class WHERE relname = 'mv_hot_active' LIMIT 1`,
+  )
+  if (!r.rows?.length) throw new Error('mv_hot_active not found in pg_class')
+  const populated = Boolean(r.rows[0].relispopulated)
+  _mvHotActiveCache.set(db, { populated, ts: now })
+  return populated
+}
+
+export async function searchRelevantHybrid(db, query, options = {}) {
+  const clean = String(query ?? '').trim()
+  if (!clean) return []
+  // Numeric-only lookup is too broad for text recall ("1" matches nearly
+  // everything through the short ILIKE path). Callers that know an entry id
+  // should use recall's `id` mode instead of query search.
+  if (/^\d+$/.test(clean)) return []
+
+  const limit = Math.max(1, Math.floor(Number(options?.limit ?? 8)))
+  const candidateWindow = Math.max(40, limit * 8)
+  const includeMembers = Boolean(options.includeMembers)
+  const writeBackMemberHits = options.writeBackMemberHits !== false
+  // Pre-filter knobs. Without them, FTS/vec rank the whole tree and a
+  // post-filter time window can wipe the result set.
+  const tsFrom = Number.isFinite(Number(options.ts_from)) ? Number(options.ts_from) : null
+  const tsTo = Number.isFinite(Number(options.ts_to)) ? Number(options.ts_to) : null
+  // Default = empty exclusion. The archive bucket holds the bulk of historical
+  // work (active is reserved for permanent invariants in this design; the
+  // last-week / last-month / "what did I work on previously" recall pattern
+  // depends on archived rows being in the pool). Cycle2 internal sweeps
+  // that genuinely want active-only data must pass excludeStatuses
+  // explicitly.
+  const excludeStatuses = Array.isArray(options.excludeStatuses)
+    ? options.excludeStatuses.filter(s => typeof s === 'string' && s.trim()).map(s => s.trim().toLowerCase())
+    : []
+  // Project scope pre-filter applied to the candidate fetch SQL.
+  // 'common' → project_id IS NULL; specific slug → project_id IS NULL OR = slug;
+  // 'all' or undefined → no filter.
+  const projectScope = typeof options.projectScope === 'string' ? options.projectScope : null
+  const categories = (Array.isArray(options.category) ? options.category : [options.category])
+    .map(c => String(c ?? '').trim().toLowerCase())
+    .filter(c => VALID_CATEGORY.has(c))
+  // Caller can disable freshness decay when the period is calendar-bounded
+  // (yesterday/today/this_week/last_week/specific date). Inside a fixed
+  // window, absolute-age decay misranks early-week vs late-week entries.
+  const applyFreshness = options.applyFreshness !== false
+
+  // ── mv_hot_active fast-path opt-in ──────────────────────────────────────
+  // When useHotActive:true, the dense and sparse CTE legs query mv_hot_active
+  // instead of the full entries table.
+  //
+  // WHEN TO USE:
+  //   - Explicit active-only recall (no archived inclusion, no ts_from/ts_to
+  //     window). The history-first default recall path should keep useHotActive
+  //     false so archived roots and fresh pending work remain eligible.
+  //   - mv_hot_active holds only active roots with embeddings. Its dedicated
+  //     HNSW (mv_hot_active_hnsw) and GIN (mv_hot_active_tsv) indexes are smaller
+  //     than the partial indexes on entries, so ANN and FTS scans are faster.
+  //   - Caller must ensure cycle2 has run at least once. The MV is created WITH NO
+  //     DATA; a never-refreshed MV silently returns 0 rows — primary risk on fresh
+  //     deployments.
+  //
+  // WHEN NOT TO USE:
+  //   - ts_from / ts_to active: MV lacks the ts column; the filter clause would
+  //     reference a non-existent column and the query would error.
+  //   - Archived entries must be included: MV only holds active rows.
+  //   - trgm is the primary signal: MV lacks content and ts, so the trgm leg
+  //     always routes to entries regardless of useHotActive.
+  //
+  // COLUMN GAPS (resolved per CTE leg):
+  //   ts      : missing → trgm short-query ORDER BY ts DESC impossible on MV;
+  //             also makes ts_from/ts_to filter clauses invalid.
+  //   content : missing → trgm similarity/ILIKE impossible on MV.
+  //   Both gaps are intentional; trgm is unconditionally routed to entries.
+  //
+  // The combined/JOIN fetch after the CTE always queries entries by id, so the
+  // final row shape is identical regardless of which path was taken.
+  const hasTsFilter = tsFrom != null || tsTo != null
+  const hasArchivedInclusion = !excludeStatuses.includes('archived')
+  let useHotActive = Boolean(options.useHotActive)
+    && !hasTsFilter
+    && !hasArchivedInclusion
+  // Guard against unrefreshed mv_hot_active (created WITH NO DATA → SQLSTATE
+  // 55000 on read). Cheap pg_class check, cached 60 s per db handle to avoid
+  // per-recall round-trip cost.
+  if (useHotActive) {
+    const populated = await _checkMvHotActivePopulated(db)
+    if (!populated) useHotActive = false
+  }
+
+  // buildFilterClause: pushes ts/status/scope filters INTO candidate SELECTs.
+  // offset = 1-based index of the first bind param it may consume.
+  // Returns { clause: string, params: any[] }; clause begins with AND or is ''.
+  function buildFilterClause(offset) {
+    return buildRecallScopeFilter(offset, {
+      ts_from: tsFrom,
+      ts_to: tsTo,
+      excludeStatuses,
+      category: categories,
+      projectScope,
+    })
+  }
+
+  // Kept for the non-candidate root-lookup inside the member-hit resolution path.
+  function buildScopeClause(offset) {
+    if (projectScope === 'common') {
+      return { clause: 'AND project_id IS NULL', params: [] }
+    } else if (projectScope && projectScope !== 'all') {
+      return { clause: `AND (project_id IS NULL OR project_id = $${offset})`, params: [projectScope] }
+    }
+    return { clause: '', params: [] }
+  }
+
+  // ── Single-round-trip hybrid CTE ─────────────────────────────────────────
+  // Param layout (fixed prefix):
+  //   $1  = halfvec literal  (NULL when no queryVector)
+  //   $2  = tsQuery text     (NULL when short query)
+  //   $3  = cleanText        (trigram term)
+  //   $4  = candidateWindow  (LIMIT for each CTE leg)
+  //   $5+ = filter params (ts_from, ts_to, excludeStatuses..., category..., projectScope slug)
+  //
+  // When a leg is inapplicable its CTE returns no rows; the UNION + LEFT JOINs
+  // handle that cleanly. dense/sparse/trgm legs each re-use the same filter
+  // params starting at $5 since they live in independent CTE scopes.
+
+  const vecSql = (Array.isArray(options.queryVector) && options.queryVector.length > 0)
+    ? embeddingToSql(options.queryVector)
+    : null
+
+  const ftsQuery = clean.length >= 3 ? (buildFtsQuery(clean) ?? null) : null
+  const exactTerms = buildExactTerms(clean)
+  const queryTokenCount = countQueryTokens(clean)
+  const minExactHits = exactTerms.length >= 8 ? 3 : exactTerms.length >= 4 ? 2 : 1
+
+  // For very short queries (< 3 chars) the trigram operator still works but
+  // we relax the server-side threshold via set_limit() — however that requires
+  // a separate round-trip. Instead we fall back to a plain ILIKE scan for
+  // short text (rare edge case; sequential scan is acceptable for < 3 chars).
+  const isShortQuery = clean.length < 3
+
+  // $5 onward are the filter params for the entries legs (non-MV path).
+  // Each CTE leg duplicates the same positional params because they live in
+  // independent SELECT scopes. When useHotActive=true, the trgm leg still uses
+  // these params but at adjusted offsets (see activeBindParams below).
+  const { clause: filterClause, params: filterParams } = buildFilterClause(5)
+
+  // MV-specific filter: only category/projectScope matter (status='active' and
+  // embedding IS NOT NULL are baked into mv_hot_active; ts_from/ts_to are
+  // unavailable since MV lacks the ts column).
+  function buildMvFilterClause(offset) {
+    const clauses = []
+    const params = []
+    let next = offset
+    if (categories.length > 0) {
+      const placeholders = categories.map(() => `$${next++}`).join(', ')
+      clauses.push(`category IN (${placeholders})`)
+      params.push(...categories)
+    }
+    if (projectScope === 'common') {
+      clauses.push('project_id IS NULL')
+    } else if (projectScope && projectScope !== 'all') {
+      clauses.push(`(project_id IS NULL OR project_id = $${next++})`)
+      params.push(projectScope)
+    }
+    return { clause: clauses.length > 0 ? `AND ${clauses.join(' AND ')}` : '', params }
+  }
+  // mvBindParams layout when useHotActive=true:
+  //   $1–$4 : same prefix (vec, fts, clean, window)
+  //   $5+   : mvFilterParams (category filters + optional projectScope slug)
+  //   $5+N+ : trgmFilterParams (ts/status/scope for the entries-only trgm leg)
+  //
+  // The trgm CTE always targets entries and needs the full filter (excludeStatuses,
+  // ts_from, ts_to, category, projectScope). When useHotActive=true, trgm filter params
+  // start AFTER mvFilterParams so positional params align correctly in the
+  // combined bind array.
+  const { clause: mvFilterClause, params: mvFilterParams } = buildMvFilterClause(5)
+  // trgm filter: when useHotActive, build starting at offset 5 + mvFilterParams.length.
+  const trgmFilterOffset = useHotActive ? 5 + mvFilterParams.length : 5
+  const { clause: trgmFilterClause, params: trgmFilterParams } = buildFilterClause(trgmFilterOffset)
+  // activeBindParams is the single array passed to db.query for the full hybrid SQL.
+  // Non-MV path: [vec,fts,clean,window, ...filterParams] (filterClause == trgmFilterClause).
+  // MV path:     [vec,fts,clean,window, ...mvFilterParams, ...trgmFilterParams].
+  const recallScopeOpts = {
+    ts_from: tsFrom,
+    ts_to: tsTo,
+    excludeStatuses,
+    category: categories,
+    projectScope,
+  }
+  const exactTermsParam = useHotActive
+    ? 5 + mvFilterParams.length + trgmFilterParams.length
+    : 5 + filterParams.length
+  const exactFilterClause = buildRecallScopeFilter(
+    useHotActive ? trgmFilterOffset : 5,
+    recallScopeOpts,
+    'ee',
+  ).clause
+  const activeBindParams = useHotActive
+    ? [vecSql, ftsQuery, clean, candidateWindow, ...mvFilterParams, ...trgmFilterParams, ...(exactTerms.length > 0 ? [exactTerms] : [])]
+    : [vecSql, ftsQuery, clean, candidateWindow, ...filterParams, ...(exactTerms.length > 0 ? [exactTerms] : [])]
+
+  // dense CTE: active only when a query vector is supplied.
+  // useHotActive → queries mv_hot_active (smaller HNSW, no ts/content needed).
+  const denseCte = vecSql ? (useHotActive ? `
+dense AS (
+  SELECT id,
+         1 - (embedding <=> $1::halfvec) AS sim,
+         ROW_NUMBER() OVER (ORDER BY embedding <=> $1::halfvec) AS dense_rank
+  FROM mv_hot_active
+  WHERE true
+    ${mvFilterClause}
+  ORDER BY embedding <=> $1::halfvec
+  LIMIT $4
+),` : `
+dense AS (
+  SELECT id,
+         1 - (embedding <=> $1::halfvec) AS sim,
+         ROW_NUMBER() OVER (ORDER BY embedding <=> $1::halfvec) AS dense_rank
+  FROM entries
+  WHERE embedding IS NOT NULL
+    ${filterClause}
+  ORDER BY embedding <=> $1::halfvec
+  LIMIT $4
+),`) : `
+dense AS (SELECT NULL::bigint AS id, NULL::float8 AS sim, NULL::bigint AS dense_rank WHERE $1::halfvec IS NOT NULL AND false),`
+
+  // sparse CTE: active only when ftsQuery is non-null.
+  // useHotActive → queries mv_hot_active GIN index (mv_hot_active_tsv).
+  const sparseCte = ftsQuery ? (useHotActive ? `
+sparse AS (
+  SELECT id,
+         ts_rank_cd(search_tsv, websearch_to_tsquery('simple', $2)) AS lex,
+         ROW_NUMBER() OVER (ORDER BY ts_rank_cd(search_tsv, websearch_to_tsquery('simple', $2)) DESC) AS sparse_rank
+  FROM mv_hot_active
+  WHERE search_tsv @@ websearch_to_tsquery('simple', $2)
+    ${mvFilterClause}
+  ORDER BY lex DESC
+  LIMIT $4
+),` : `
+sparse AS (
+  SELECT id,
+         ts_rank_cd(search_tsv, websearch_to_tsquery('simple', $2)) AS lex,
+         ROW_NUMBER() OVER (ORDER BY ts_rank_cd(search_tsv, websearch_to_tsquery('simple', $2)) DESC) AS sparse_rank
+  FROM entries
+  WHERE search_tsv @@ websearch_to_tsquery('simple', $2)
+    ${filterClause}
+  ORDER BY lex DESC
+  LIMIT $4
+),`) : `
+sparse AS (SELECT NULL::bigint AS id, NULL::float8 AS lex, NULL::bigint AS sparse_rank WHERE $2::text IS NOT NULL AND false),`
+
+  // trgm CTE: pg_trgm similarity path. For short queries (< 3 chars) the %
+  // operator is unreliable (trigrams need at least 3 chars); use ILIKE instead.
+  // NOTE: trgm always queries entries regardless of useHotActive — mv_hot_active
+  // lacks the content column (trgm/ILIKE) and ts column (short-query ORDER BY).
+  // Uses trgmFilterClause whose $N offsets are aligned to activeBindParams.
+  const trgmCte = isShortQuery ? `
+trgm AS (
+  SELECT id,
+         0.5::float8 AS trg_sim,
+         ROW_NUMBER() OVER (ORDER BY ts DESC) AS trgm_rank
+  FROM entries
+  WHERE (content ILIKE '%' || $3 || '%' OR element ILIKE '%' || $3 || '%')
+    ${trgmFilterClause}
+  ORDER BY ts DESC
+  LIMIT $4
+),` : `
+trgm AS (
+  SELECT id,
+          GREATEST(
+            CASE WHEN content ILIKE '%' || $3 || '%' THEN 1.0 ELSE similarity(content, $3) END,
+            CASE WHEN coalesce(element, '') ILIKE '%' || $3 || '%' THEN 1.0 ELSE similarity(coalesce(element, ''), $3) END,
+            CASE WHEN coalesce(summary, '') ILIKE '%' || $3 || '%' THEN 1.0 ELSE similarity(coalesce(summary, ''), $3) END
+          ) AS trg_sim,
+          ROW_NUMBER() OVER (ORDER BY GREATEST(
+            CASE WHEN content ILIKE '%' || $3 || '%' THEN 1.0 ELSE similarity(content, $3) END,
+            CASE WHEN coalesce(element, '') ILIKE '%' || $3 || '%' THEN 1.0 ELSE similarity(coalesce(element, ''), $3) END,
+            CASE WHEN coalesce(summary, '') ILIKE '%' || $3 || '%' THEN 1.0 ELSE similarity(coalesce(summary, ''), $3) END
+          ) DESC) AS trgm_rank
+   FROM entries
+   WHERE (
+       content % $3 OR element % $3 OR summary % $3
+       OR content ILIKE '%' || $3 || '%'
+       OR coalesce(element, '') ILIKE '%' || $3 || '%'
+       OR coalesce(summary, '') ILIKE '%' || $3 || '%'
+     )
+     AND GREATEST(
+       CASE WHEN content ILIKE '%' || $3 || '%' THEN 1.0 ELSE similarity(content, $3) END,
+       CASE WHEN coalesce(element, '') ILIKE '%' || $3 || '%' THEN 1.0 ELSE similarity(coalesce(element, ''), $3) END,
+       CASE WHEN coalesce(summary, '') ILIKE '%' || $3 || '%' THEN 1.0 ELSE similarity(coalesce(summary, ''), $3) END
+     ) >= 0.10
+     ${trgmFilterClause}
+   ORDER BY trg_sim DESC
+   LIMIT $4
+ ),`
+
+  const exactCte = exactTerms.length > 0 ? `
+exact AS (
+  SELECT ee.id,
+         COUNT(*)::float8 AS exact_hits,
+         ROW_NUMBER() OVER (ORDER BY COUNT(*) DESC, ee.ts DESC) AS exact_rank
+  FROM entries ee
+  JOIN LATERAL unnest($${exactTermsParam}::text[]) AS q(term) ON (
+       ee.content ILIKE '%' || q.term || '%'
+       OR coalesce(ee.element, '') ILIKE '%' || q.term || '%'
+       OR coalesce(ee.summary, '') ILIKE '%' || q.term || '%'
+  )
+  WHERE true
+    ${exactFilterClause}
+  GROUP BY ee.id, ee.ts
+  HAVING COUNT(*) >= ${minExactHits}
+  ORDER BY exact_hits DESC, ee.ts DESC
+  LIMIT $4
+),` : `
+exact AS (SELECT NULL::bigint AS id, NULL::float8 AS exact_hits, NULL::bigint AS exact_rank WHERE false),`
+
+  const hybridSql = `
+WITH
+${denseCte}
+${sparseCte}
+${trgmCte}
+${exactCte}
+combined AS (
+  SELECT id FROM dense  WHERE id IS NOT NULL UNION
+  SELECT id FROM sparse WHERE id IS NOT NULL UNION
+  SELECT id FROM trgm   WHERE id IS NOT NULL UNION
+  SELECT id FROM exact  WHERE id IS NOT NULL
+)
+SELECT
+  e.id, e.element, e.summary, e.category, e.status, e.score,
+  e.last_seen_at, e.ts, e.project_id, e.session_id, e.source_ref,
+  e.source_turn, e.content, e.chunk_root, e.is_root,
+  e.role,
+  d.sim        AS dense_sim,
+  d.dense_rank,
+  s.lex        AS sparse_lex,
+  s.sparse_rank,
+  t.trg_sim,
+  t.trgm_rank,
+  x.exact_hits,
+  x.exact_rank
+FROM combined c
+JOIN   entries e ON e.id = c.id
+LEFT JOIN dense  d ON d.id = c.id
+LEFT JOIN sparse s ON s.id = c.id
+LEFT JOIN trgm   t ON t.id = c.id
+LEFT JOIN exact  x ON x.id = c.id`
+
+  let rawRows = []
+  let denseCount = 0
+  let sparseCount = 0
+  let trgmCount = 0
+  let exactCount = 0
+
+  try {
+    const { rows } = await recallReadQuery(db, hybridSql, activeBindParams)
+    rawRows = rows
+    // Count how many rows each leg contributed (a row may appear in multiple legs).
+    for (const r of rawRows) {
+      if (r.dense_rank != null) denseCount++
+      if (r.sparse_rank != null) sparseCount++
+      if (r.trgm_rank != null) trgmCount++
+      if (r.exact_rank != null) exactCount++
+    }
+  } catch (err) {
+    __mixdogMemoryLog(`[recall] hybrid CTE failed: ${err.message}\n`)
+    return []
+  }
+
+  if (rawRows.length === 0) return []
+
+  // ── JS-side RRF merge (unchanged logic) ──────────────────────────────────
+  // K=60 is the standard RRF constant from Cormack et al. (SIGIR 2009).
+  const K = 60
+  const nowMs = Date.now()
+
+  const scoredAll = rawRows.map(row => {
+    const id = Number(row.id)
+    const denseRank = row.dense_rank != null ? Number(row.dense_rank) : null
+    const sparseRank = row.sparse_rank != null ? Number(row.sparse_rank) : null
+    const trgmRank = row.trgm_rank != null ? Number(row.trgm_rank) : null
+    const exactRank = row.exact_rank != null ? Number(row.exact_rank) : null
+    const rrf = (denseRank ? 1 / (K + denseRank) : 0)
+              + (sparseRank ? 1 / (K + sparseRank) : 0)
+              + (trgmRank ? 1 / (K + trgmRank) : 0)
+              + (exactRank ? 1 / (K + exactRank) : 0)
+    const freshness = applyFreshness ? freshnessFactor(row.ts, nowMs) : 1.0
+    const boost = exactTextBoost(clean, row, row.exact_hits)
+    return { id, row, rrf, freshness, retrievalScore: (rrf * freshness) + boost }
+  })
+  let semanticOnlyDropped = 0
+  let weakTextDropped = 0
+  const scored = scoredAll.filter(({ row }) => {
+    const hasTextSupport = row.sparse_rank != null || row.trgm_rank != null || row.exact_rank != null
+    const sim = Number(row.dense_sim)
+    const hasSemanticSupport = Number.isFinite(sim) && sim >= SEMANTIC_ONLY_MIN_SIM
+    if (!hasTextSupport) {
+      if (hasSemanticSupport) return true
+      semanticOnlyDropped += 1
+      return false
+    }
+    // A long query that only shares a weak trigram/exact tail with old rows
+    // should not fill the page with accidental matches. Accept lexical-only
+    // rows when the query is a short keyword/identifier lookup, a full phrase
+    // match, or an FTS hit. Otherwise require semantic support as a second
+    // independent signal.
+    if (categories.length > 0) return true
+    const hasFullPhrase = hasFullQueryTextMatch(clean, row)
+    if (hasFullPhrase) return true
+    if (queryTokenCount <= SHORT_QUERY_TOKEN_MAX) return true
+    if (row.sparse_rank != null) return true
+    if (hasSemanticSupport && hasQueryTokenCoverage(row, queryTokenCount)) return true
+    weakTextDropped += 1
+    return false
+  })
+  if (scored.length === 0) return []
+  scored.sort((a, b) => b.retrievalScore - a.retrievalScore || b.rrf - a.rrf)
+
+  const filtered = scored
+
+  // ── Root resolution + member-hit write-back ───────────────────────────────
+  const byId = new Map(rawRows.map(r => [Number(r.id), r]))
+  const memberHitRootIds = new Set()
+  const rootIdsForReturn = []
+  const seen = new Set()
+
+  // Batch-resolve member-chunk roots in ONE query (was an N+1: a per-row SELECT
+  // inside the loop below). Collect the distinct in-scope chunk_root ids, fetch
+  // all matching roots at once, then resolve each member from rootById.
+  const memberRootIds = []
+  const memberRootSeen = new Set()
+  for (const { id } of filtered) {
+    const r0 = byId.get(id)
+    if (!r0 || r0.is_root === 1) continue
+    if (r0.chunk_root != null && r0.chunk_root !== r0.id) {
+      const rid = Number(r0.chunk_root)
+      if (!memberRootSeen.has(rid)) { memberRootSeen.add(rid); memberRootIds.push(rid) }
+    }
+  }
+  const rootById = new Map()
+  if (memberRootIds.length > 0) {
+    const { clause: rootScopeClause, params: rootScopeParams } = buildScopeClause(2)
+    const { rows: rootRows } = await recallReadQuery(
+      db,
+      `SELECT id, ts, role, content, session_id, source_turn, chunk_root, is_root,
+              element, category, summary, project_id, status, score, last_seen_at
+       FROM entries WHERE id = ANY($1::bigint[]) AND is_root = 1 ${rootScopeClause}`,
+      [memberRootIds, ...rootScopeParams],
+    )
+    for (const rr of rootRows) rootById.set(Number(rr.id), rr)
+  }
+
+  for (const { id, rrf, retrievalScore } of filtered) {
+    const row = byId.get(id)
+    if (!row) continue
+    let targetRow = null
+    if (row.is_root === 1) {
+      targetRow = row
+    } else if (row.chunk_root != null && row.chunk_root !== row.id) {
+      const r = rootById.get(Number(row.chunk_root))
+      if (!r) continue
+      memberHitRootIds.add(r.id)
+      targetRow = r
+    } else {
+      targetRow = row
+    }
+    if (seen.has(targetRow.id)) continue
+    seen.add(targetRow.id)
+    rootIdsForReturn.push({
+      root: targetRow,
+      rrf,
+      retrievalScore,
+      retrievalRank: rootIdsForReturn.length + 1,
+    })
+    if (rootIdsForReturn.length >= limit) break
+  }
+
+  let writeBackCount = 0
+  if (writeBackMemberHits && memberHitRootIds.size > 0) {
+    // Batch UPDATE — single round-trip instead of N (one per member-hit root).
+    const validRootIds = []
+    for (const rootId of memberHitRootIds) {
+      const r = rootIdsForReturn.find(x => x.root.id === rootId)?.root ?? byId.get(rootId)
+      if (r) validRootIds.push(Number(rootId))
+    }
+    if (validRootIds.length > 0) {
+      try {
+        const { rowCount } = await db.query(
+          `UPDATE entries SET last_seen_at = $1::bigint WHERE id = ANY($2::bigint[]) AND is_root = 1
+           AND (last_seen_at IS NULL OR last_seen_at < $1::bigint - 3600000)`,
+          [String(nowMs), validRootIds],
+        )
+        writeBackCount = rowCount ?? validRootIds.length
+      } catch (err) {
+        __mixdogMemoryLog(`[recall] writeback batch failed (count=${validRootIds.length}): ${err.message}\n`)
+      }
+    }
+  }
+
+  // ── Final fetch: full row for each root by id = ANY(bigint[]) ────────────
+  const topIds = rootIdsForReturn.map(x => Number(x.root.id))
+  const { clause: finalFilter, params: finalFilterParams } = buildFilterClause(2)
+  const { rows: finalRows } = await recallReadQuery(
+    db,
+    `SELECT id, ts, role, content, session_id, source_turn, chunk_root, is_root,
+            element, category, summary, project_id, status, score, last_seen_at
+     FROM entries
+     WHERE id = ANY($1::bigint[])
+       ${finalFilter}`,
+    [topIds, ...finalFilterParams],
+  )
+  const finalById = new Map(finalRows.map(r => [Number(r.id), r]))
+
+  // Members: single batch fetch keyed by chunk_root = ANY($1) — one
+  // round-trip vs N. Map to per-root arrays preserving (ts ASC, id ASC).
+  let membersByRoot = new Map()
+  if (includeMembers) {
+    const rootIds = rootIdsForReturn
+      .map(x => Number(finalById.get(Number(x.root.id))?.id ?? x.root.id))
+      .filter(id => {
+        const fr = finalById.get(id) ?? rootIdsForReturn.find(x => Number(x.root.id) === id)?.root
+        return fr && fr.is_root === 1
+      })
+    if (rootIds.length > 0) {
+      const { rows: memberRows } = await recallReadQuery(
+        db,
+        `SELECT id, ts, role, content, session_id, source_turn, project_id, chunk_root
+         FROM entries WHERE chunk_root = ANY($1::bigint[]) AND is_root = 0
+         ORDER BY ts ASC, id ASC`,
+        [rootIds],
+      )
+      for (const m of memberRows) {
+        const k = Number(m.chunk_root)
+        if (!membersByRoot.has(k)) membersByRoot.set(k, [])
+        membersByRoot.get(k).push(m)
+      }
+    }
+  }
+  const results = []
+  for (const { root, rrf, retrievalScore, retrievalRank } of rootIdsForReturn) {
+    // Roots absent from finalById were excluded by the status/time filter on
+    // the final fetch; falling back to the unfiltered `root` would leak
+    // archived / out-of-window rows via member-hit resolution.
+    const finalRoot = finalById.get(Number(root.id))
+    if (!finalRoot) continue
+    const out = { ...finalRoot, rrf, retrievalScore, retrievalRank }
+    if (includeMembers && finalRoot.is_root === 1) {
+      out.members = membersByRoot.get(Number(finalRoot.id)) ?? []
+    }
+    results.push(out)
+  }
+
+  __mixdogMemoryLog(
+    `[recall] dense=${denseCount} sparse=${sparseCount} trgm=${trgmCount} exact=${exactCount} semantic_only_dropped=${semanticOnlyDropped} weak_text_dropped=${weakTextDropped} merged=${results.length} write_back=${writeBackCount}\n`,
+  )
+
+  return results
+}