mixdog 0.7.1

package/src/memory/lib/memory-cycle2.mjs

@@ -0,0 +1,1284 @@
+import { existsSync, readFileSync } from 'fs'
+import { join } from 'path'
+import { homedir as _homedir } from 'os'
+import { resolveMaintenancePreset } from '../../shared/llm/index.mjs'
+import { callBridgeLlm } from './agent-ipc.mjs'
+import {
+  syncRootEmbedding, deleteRootEmbedding, flushEmbeddingDirty,
+} from './memory-embed.mjs'
+import { listCore, backfillCoreEmbeddings, CORE_SUMMARY_MAX } from './core-memory-store.mjs'
+
+export const CYCLE2_ACTIVE_TARGET_CAP = 100
+const TIER1_THRESHOLD = 0.78
+
+const TIER2_LOW = 0.65
+const LLM_JUDGE_CAP = 20
+
+function throwIfAborted(signal) {
+  if (signal?.aborted) throw signal.reason ?? new Error('aborted')
+}
+
+// Status-based verb whitelist. 3-tier policy: pending → active/archived,
+// active → active/archived/update/merge.
+const STATUS_ALLOWED_VERBS = {
+  pending: new Set(['active', 'archived']),
+  active:  new Set(['active', 'archived', 'update', 'merge']),
+}
+const NON_ARCHIVE_VERBS = new Set(['active', 'update', 'merge'])
+// Union of every primary (status) verb across all statuses, plus the two
+// non-verb line kinds. Used by the stray-index shift guard to decide whether
+// a `idx|id|verb` line had a leading row index prepended by the LLM.
+const ALL_PRIMARY_VERBS = new Set(['active', 'archived', 'update', 'merge'])
+const isShiftFollowToken = (tok) => {
+  const v = String(tok ?? '').trim().toLowerCase()
+  return ALL_PRIMARY_VERBS.has(v) || v === 'why' || v === 'core'
+}
+
+function resourceDir() {
+  if (process.env.CLAUDE_PLUGIN_ROOT) return process.env.CLAUDE_PLUGIN_ROOT
+  throw new Error('CLAUDE_PLUGIN_ROOT env var required for prompt loading')
+}
+
+async function invokeLlm(prompt, mode, preset, timeout, llmCall = callBridgeLlm) {
+  return await llmCall({
+    role: 'cycle2-agent',
+    taskType: 'maintenance',
+    mode,
+    preset,
+    timeout,
+    cwd: null,
+  }, prompt)
+}
+
+function buildPidMap(rowSets) {
+  const pids = [...new Set(rowSets.flat().map(r => r.project_id).filter(Boolean))].sort()
+  return new Map(pids.map((p, i) => [p, `P${i + 1}`]))
+}
+
+function formatEntriesForPromotePrompt(rows, pidMap, opts = {}) {
+  if (!rows || rows.length === 0) return '(none)'
+  const map = pidMap ?? buildPidMap([rows])
+  // When numbered, prefix each row with its 1-based prompt-order ordinal so the
+  // gate LLM can echo a row number it can see, instead of inventing one. The
+  // ordinal domain (1..N) and the 5-digit batch-id domain must stay disjoint —
+  // see the ordinalToId invariant in runUnifiedGate.
+  const numbered = opts.numbered === true
+  const lines = rows.map((r, i) => {
+    const tag = r.project_id ? (map.get(r.project_id) ?? 'C') : 'C'
+    const stat = r.status ? `[${r.status}]` : '[?]'
+    const prefix = numbered ? `${i + 1}. ` : '- '
+    return `${prefix}id:${r.id} ${stat} ${tag} ${r.category} s:${r.score ?? 'n'} el:${r.element} sm:${String(r.summary || '').slice(0, 100)}`
+  })
+  if (map.size === 0) return lines.join('\n')
+  const legend = [...map.entries()].map(([p, t]) => `${t}=${p}`).concat('C=COMMON').join(', ')
+  return `# pid: ${legend}\n` + lines.join('\n')
+}
+
+// User-curated rows from core_entries — id-less, no status, no score; the
+// LLM only needs element + summary + project tag to detect overlap with
+// candidate entries below. Format kept terse so the prompt budget stays small.
+function formatUserCoreForPrompt(rows, pidMap) {
+  if (!rows || rows.length === 0) return '(none)'
+  const map = pidMap ?? new Map()
+  return rows.map(r => {
+    const tag = r.project_id ? (map.get(r.project_id) ?? 'C') : 'C'
+    const sm = String(r.summary || '').slice(0, 200)
+    return `- ${tag} ${r.category}: ${r.element}${sm && sm !== r.element ? ` — ${sm}` : ''}`
+  }).join('\n')
+}
+
+// Parse pipe-format unified verdicts. Each line: <id>|<verb> [|...].
+// Verbs validated against the row's current status via STATUS_ALLOWED_VERBS.
+// Returns { actions, rejected } or null when no parseable lines.
+function parseUnifiedFormat(raw, statusById, ordinalToId = null) {
+  if (raw == null) return null
+  const text = String(raw).trim()
+  if (!text) return { actions: [], rejected: new Set() }
+  const lines = text.split('\n')
+  const actions = []
+  const rejected = new Set()
+  const support = new Map()
+  let sawValid = false
+  // Resolve a first-field/merge token to a real batch id. The gate may echo
+  // either the exact 5-digit batch id OR the 1-based row ordinal shown in the
+  // numbered Entries block. The two domains are disjoint (asserted in
+  // runUnifiedGate), so an exact-id hit always wins and an unmatched value
+  // falls back to ordinal lookup; anything else is NaN (line treated invalid).
+  const resolveId = (tok) => {
+    const n = Number(String(tok ?? '').trim())
+    if (!Number.isFinite(n)) return NaN
+    if (statusById.has(n)) return n
+    if (ordinalToId && ordinalToId.has(n)) return ordinalToId.get(n)
+    return NaN
+  }
+  for (const rawLine of lines) {
+    const line = rawLine.trim()
+    if (!line) continue
+    if (line.startsWith('//') || line.startsWith('#')) continue
+    if (line.startsWith('```')) continue
+    const parts = line.split('|')
+    if (parts.length < 2) continue
+    // LLM sometimes prefixes a row index, emitting `idx|id|verdict` instead of
+    // `id|verdict`; parts[0] (the index) is a stray token and the line must be
+    // shifted before parsing. Strict invariant so a real 2-field `id|verdict`
+    // is never shifted into a 1-field line (which would throw on parts[1]):
+    //   parts.length >= 3 AND parts[1] is a known batch id AND parts[2] is a
+    //   valid primary verb / why / core (the shifted verdict slot).
+    // Trigger when EITHER parts[0] is not a known id (classic stray index) OR
+    // parts[0] IS known but parts[1] is not itself a valid verb — that covers
+    // `1|1|active`, where the stray index collides with a real batch id and the
+    // un-shifted reading would verb-reject the wrong row.
+    if (
+      parts.length >= 3 &&
+      statusById.has(Number(parts[1].trim())) &&
+      isShiftFollowToken(parts[2]) &&
+      (!statusById.has(Number(parts[0].trim())) || !isShiftFollowToken(parts[1]))
+    ) {
+      parts.shift()
+    }
+    const entryId = resolveId(parts[0])
+    const action = parts[1].trim().toLowerCase()
+    if (!Number.isFinite(entryId) || !action) continue
+    const status = statusById.get(entryId)
+    if (!status) continue
+    // Only mark as parse-ok when the id is known to the batch; a response
+    // composed entirely of unknown ids would otherwise return parse-ok with
+    // zero actions/rejections, leaving the rows un-reviewed and re-queued.
+    sawValid = true
+    if (action === 'core') {
+      actions.push({ entry_id: entryId, action: 'core', core_summary: parts.slice(2).join('|').trim().slice(0, 120) })
+      continue
+    }
+    if (action === 'why') {
+      const kind = (parts[2] ?? '').trim().toUpperCase()
+      const reason = parts.slice(3).join('|').replace(/\s+/g, ' ').trim().slice(0, 240)
+      if ((kind === 'A' || kind === 'B') && reason) {
+        support.set(entryId, { kind, reason })
+      }
+      continue
+    }
+    const allowed = STATUS_ALLOWED_VERBS[status]
+    if (!allowed || !allowed.has(action)) {
+      process.stderr.write(`[cycle2] verb rejected: id=${entryId} status=${status} verb=${action}\n`)
+      rejected.add(entryId)
+      continue
+    }
+    if (action === 'update') {
+      actions.push({
+        entry_id: entryId, action,
+        element: (parts[2] ?? '').trim(),
+        summary: parts.slice(3).join('|').trim(),
+      })
+    } else if (action === 'merge') {
+      const targetId = resolveId(parts[2])
+      const sourceIds = [...new Set((parts[3] ?? '').split(',').map(s => resolveId(s)).filter(Number.isFinite))]
+      if (!Number.isFinite(targetId) || sourceIds.length === 0) {
+        process.stderr.write(`[cycle2] merge rejected: id=${entryId} invalid target/sources\n`)
+        rejected.add(entryId)
+        continue
+      }
+      if (targetId !== entryId && !sourceIds.includes(entryId)) {
+        process.stderr.write(
+          `[cycle2] merge rejected: id=${entryId} must be target or listed source (target=${targetId} sources=${sourceIds.join(',')})\n`,
+        )
+        rejected.add(entryId)
+        continue
+      }
+      actions.push({
+        entry_id: entryId, action,
+        target_id: targetId,
+        source_ids: sourceIds,
+        element: (parts[4] ?? '').trim(),
+        summary: parts.slice(5).join('|').trim(),
+      })
+    } else {
+      actions.push({ entry_id: entryId, action })
+    }
+  }
+  if (!sawValid && rejected.size === 0) return null
+  return { actions, rejected, support }
+}
+
+// Batch CTE UPDATE for status-only verdicts (active/archived from pending or active rows).
+// Trigger handles score recompute automatically — no app-side score writes.
+async function applyBatchStatusVerdicts(db, batch, nowMs) {
+  if (!batch || batch.length === 0) return { promoted: 0, archived: 0 }
+  const valueRows = batch.map((item, i) => {
+    const base = i * 3
+    return `($${base + 1}::bigint, $${base + 2}::text, $${base + 3}::boolean)`
+  })
+  const params = []
+  for (const item of batch) {
+    params.push(item.entry_id, item.new_status, item.was_pending)
+  }
+  params.push(nowMs)
+  const lastParam = `$${params.length}`
+  const res = await db.query(
+    `WITH actions(entry_id, new_status, was_pending) AS (
+       VALUES ${valueRows.join(', ')}
+     )
+     UPDATE entries
+     SET status = a.new_status::entry_status,
+         last_seen_at = ${lastParam},
+         promoted_at = CASE
+           WHEN a.was_pending AND a.new_status = 'active' THEN ${lastParam}
+           ELSE promoted_at
+         END
+     FROM actions a
+     WHERE entries.id = a.entry_id AND entries.is_root = 1
+     RETURNING entries.id, entries.status, a.was_pending, a.new_status`,
+    params,
+  )
+  let promoted = 0
+  let archived = 0
+  for (const r of (res.rows ?? [])) {
+    if (r.was_pending && r.new_status === 'active') promoted += 1
+    else if (r.new_status === 'archived') archived += 1
+  }
+  return { promoted, archived }
+}
+
+// Generic status update for archived/active terminal transitions.
+export async function applySimpleStatus(db, entryId, nextStatus) {
+  const res = await db.query(
+    `UPDATE entries SET status = $1 WHERE id = $2 AND is_root = 1`,
+    [nextStatus, entryId],
+  )
+  return Number(res.rowCount ?? res.affectedRows ?? 0) > 0
+}
+
+export async function applyUpdate(db, entryId, element, summary, options = {}) {
+  const setClauses = []
+  const params = []
+  let paramIdx = 1
+  const newElement = (typeof element === 'string' && element.trim()) ? element.trim() : null
+  const newSummary = (typeof summary === 'string' && summary.trim()) ? summary.trim() : null
+  if (newElement) {
+    setClauses.push(`element = $${paramIdx++}`); params.push(newElement)
+  }
+  if (newSummary) {
+    setClauses.push(`summary = $${paramIdx++}`); params.push(newSummary)
+    setClauses.push('summary_hash = NULL')
+  }
+  if (setClauses.length === 0) return false
+  params.push(entryId)
+  const res = await db.query(
+    `UPDATE entries SET ${setClauses.join(', ')} WHERE id = $${paramIdx} AND is_root = 1`,
+    params,
+  )
+  if (Number(res.rowCount ?? res.affectedRows ?? 0) === 0) return false
+  await syncRootEmbedding(db, entryId, options)
+  return true
+}
+
+export async function applyMerge(db, targetId, sourceIds, options = {}) {
+  const signal = options?.signal
+  throwIfAborted(signal)
+  if (!Number.isFinite(targetId)) return 0
+  const targetRes = await db.query(
+    `SELECT id, project_id FROM entries WHERE id = $1 AND is_root = 1`,
+    [targetId],
+  )
+  throwIfAborted(signal)
+  const target = targetRes.rows[0]
+  if (!target) return 0
+  let moved = 0
+  for (const src of sourceIds) {
+    throwIfAborted(signal)
+    const sid = Number(src)
+    if (!Number.isFinite(sid) || sid === targetId) continue
+    const srcRes = await db.query(
+      `SELECT id, project_id, status FROM entries WHERE id = $1 AND is_root = 1`,
+      [sid],
+    )
+    throwIfAborted(signal)
+    const srcRow = srcRes.rows[0]
+    if (!srcRow) continue
+    if (target.project_id !== srcRow.project_id) {
+      process.stderr.write(
+        `[cycle2] merge rejected: cross-pool (target=${targetId} project_id=${target.project_id ?? 'COMMON'} src=${sid} project_id=${srcRow.project_id ?? 'COMMON'})\n`,
+      )
+      continue
+    }
+    try {
+      // One source merge is the mutation unit: DB reassignment/archive plus
+      // embedding cleanup. The next abort checkpoint is before the next source.
+      await db.transaction(async (tx) => {
+        await tx.query(
+          `UPDATE entries SET chunk_root = $1, project_id = $2 WHERE chunk_root = $3 AND id != $4 AND is_root = 0`,
+          [targetId, target.project_id, sid, sid],
+        )
+        await tx.query(
+          `UPDATE entries SET status = 'archived' WHERE id = $1 AND is_root = 1`,
+          [sid],
+        )
+      })
+      await deleteRootEmbedding(db, sid)
+      moved += 1
+    } catch (err) {
+      process.stderr.write(`[cycle2] merge failed (target=${targetId} src=${sid}): ${err.message}\n`)
+    }
+  }
+  return moved
+}
+
+// ─── phase_merge: cosine-similarity dedup pass ───────────────────────────────
+
+function _pickKeeper(a, b) {
+  if ((a.score ?? 0) !== (b.score ?? 0)) return (a.score ?? 0) > (b.score ?? 0) ? a : b
+  if ((a.last_seen_at ?? 0) !== (b.last_seen_at ?? 0)) return (a.last_seen_at ?? 0) > (b.last_seen_at ?? 0) ? a : b
+  return a.id < b.id ? a : b
+}
+
+async function _llmJudgePair(summaryA, summaryB, siblingContext = [], options = {}) {
+  const signal = options?.signal
+  throwIfAborted(signal)
+  const llmCall = typeof options?.callLlm === 'function' ? options.callLlm : callBridgeLlm
+  const siblings = Array.isArray(siblingContext) && siblingContext.length > 0
+    ? `\n\nSibling near-matches (recall context only — do not absorb these into the verdict):\n${siblingContext.slice(0, 5).map((p, i) => `${i + 1}. ${String(p.a?.summary ?? '')} ↔ ${String(p.b?.summary ?? '')}`).join('\n')}`
+    : ''
+  const prompt =
+    `Two memory entries below. Are they restating the same principle? Reply ONE WORD: merge or distinct.\n\nA: ${summaryA}\nB: ${summaryB}${siblings}`
+  try {
+    const raw = await llmCall({
+      role: 'cycle2-agent',
+      taskType: 'maintenance',
+      mode: 'cycle2-phase_merge_judge',
+      preset: 'HAIKU',
+      timeout: 30000,
+      cwd: null,
+    }, prompt)
+    throwIfAborted(signal)
+    return String(raw ?? '').trim().toLowerCase().startsWith('merge')
+  } catch (err) {
+    if (signal?.aborted) throw signal.reason ?? err
+    process.stderr.write(`[cycle2] phase_merge llm-judge error: ${err.message}\n`)
+    return false
+  }
+}
+
+export async function runPhaseMerge(db, options = {}) {
+  const signal = options?.signal
+  throwIfAborted(signal)
+  // PG-side lateral nearest-neighbor via HNSW index — replaces JS O(n²) double loop.
+  const pairRes = await db.query(
+    `WITH active AS (
+       SELECT id, category, summary, score, last_seen_at, status, embedding, project_id
+       FROM entries
+       WHERE is_root = 1 AND status = 'active' AND embedding IS NOT NULL
+     )
+     SELECT a.id AS a_id, a.category AS a_category, a.summary AS a_summary, a.score AS a_score, a.last_seen_at AS a_last_seen_at, a.status AS a_status,
+            b.id AS b_id, b.category AS b_category, b.summary AS b_summary, b.score AS b_score, b.last_seen_at AS b_last_seen_at, b.status AS b_status,
+            1 - (a.embedding <=> b.embedding)::float8 AS sim
+     FROM active a
+     CROSS JOIN LATERAL (
+       SELECT id, category, summary, score, last_seen_at, status, embedding
+       FROM active inner_b
+       WHERE inner_b.id != a.id AND inner_b.category = a.category
+         AND inner_b.project_id IS NOT DISTINCT FROM a.project_id
+       ORDER BY inner_b.embedding <=> a.embedding
+       LIMIT 8
+     ) b
+     WHERE a.id < b.id
+       AND 1 - (a.embedding <=> b.embedding) >= $1
+     ORDER BY sim DESC`,
+    [TIER2_LOW],
+  )
+  throwIfAborted(signal)
+
+  const tier1Pairs = []
+  const tier2Pairs = []
+  for (const row of pairRes.rows) {
+    throwIfAborted(signal)
+    const a = { id: row.a_id, category: row.a_category, summary: row.a_summary, score: row.a_score, last_seen_at: row.a_last_seen_at, status: row.a_status }
+    const b = { id: row.b_id, category: row.b_category, summary: row.b_summary, score: row.b_score, last_seen_at: row.b_last_seen_at, status: row.b_status }
+    if (row.sim >= TIER1_THRESHOLD) tier1Pairs.push({ a, b, sim: row.sim })
+    else tier2Pairs.push({ a, b, sim: row.sim })
+  }
+
+  // No active/active similarity pairs is NOT a reason to skip the
+  // core_entries overlap sweep below — that pass archives active entries
+  // that restate a user-curated core row and is independent of intra-
+  // entry pairing. Falling through with merged=0 keeps the cross-table
+  // sweep running and the per-phase log shape intact.
+  let merged = 0
+  let llmCalls = 0
+  const mergedIds = new Set()
+
+  const doMerge = async (a, b, sim) => {
+    throwIfAborted(signal)
+    if (mergedIds.has(a.id) || mergedIds.has(b.id)) return
+    const keeper = _pickKeeper(a, b)
+    const loser = keeper.id === a.id ? b : a
+    const moved = await applyMerge(db, keeper.id, [loser.id], { signal })
+    if (moved > 0) {
+      merged += moved
+      mergedIds.add(loser.id)
+      process.stderr.write(
+        `[cycle2] phase_merge merged id=${loser.id} -> keeper=${keeper.id} category=${keeper.category} sim=${typeof sim === 'number' ? sim.toFixed(3) : '?'}\n`,
+      )
+    }
+  }
+
+  // Only tier1 pairs enter the LLM judge. Tier2 pairs (0.65 ≤ sim < 0.78)
+  // are recall context only — passed as sibling examples to the judge, never
+  // as judge input themselves, and never archived here.
+  for (const pair of tier1Pairs) {
+    throwIfAborted(signal)
+    if (llmCalls >= LLM_JUDGE_CAP) break
+    if (mergedIds.has(pair.a.id) || mergedIds.has(pair.b.id)) continue
+    llmCalls++
+    const shouldMerge = await _llmJudgePair(
+      String(pair.a.summary ?? ''),
+      String(pair.b.summary ?? ''),
+      tier2Pairs,
+      { signal },
+    )
+    throwIfAborted(signal)
+    if (shouldMerge) await doMerge(pair.a, pair.b, pair.sim)
+  }
+
+  // Cross-table sweep: surface every active entry whose embedding sits near
+  // a user-curated core_entries row (sim ≥ TIER2_LOW for broad recall) and
+  // ask the LLM whether the entry is a restatement of that user rule. Only
+  // the LLM verdict moves the entry to archived — embedding sim alone is
+  // never authoritative. Project-scoped core only matches the same pool;
+  // COMMON core is global and may absorb duplicate generated project memory.
+  throwIfAborted(signal)
+  const coreOverlapRes = await db.query(
+    `WITH active_e AS (
+       SELECT id, project_id, summary, embedding
+       FROM entries
+       WHERE is_root = 1 AND status = 'active' AND embedding IS NOT NULL
+     )
+     SELECT e.id AS entry_id, e.summary AS entry_summary, c.core_id, c.core_summary, c.sim
+     FROM active_e e
+     CROSS JOIN LATERAL (
+       SELECT inner_c.id AS core_id, inner_c.summary AS core_summary,
+              1 - (e.embedding <=> inner_c.embedding)::float8 AS sim
+       FROM core_entries inner_c
+       WHERE inner_c.embedding IS NOT NULL
+         AND (inner_c.project_id IS NULL OR inner_c.project_id IS NOT DISTINCT FROM e.project_id)
+       ORDER BY
+         CASE WHEN inner_c.project_id IS NOT DISTINCT FROM e.project_id THEN 0 ELSE 1 END,
+         inner_c.embedding <=> e.embedding
+       LIMIT 1
+     ) c
+     WHERE c.sim >= $1`,
+    [TIER1_THRESHOLD],
+  )
+  throwIfAborted(signal)
+  let coreOverlap = 0
+  for (const row of coreOverlapRes.rows) {
+    throwIfAborted(signal)
+    if (llmCalls >= LLM_JUDGE_CAP) break
+    llmCalls++
+    const verdictMerge = await _llmJudgePair(
+      String(row.entry_summary ?? ''),
+      String(row.core_summary ?? ''),
+      [],
+      { signal },
+    )
+    throwIfAborted(signal)
+    if (!verdictMerge) continue
+    // Archiving one overlap and deleting its embedding is one mutation unit;
+    // cancellation resumes at the next row boundary.
+    const r = await db.query(
+      `UPDATE entries SET status = 'archived' WHERE id = $1 AND is_root = 1 AND status = 'active'`,
+      [Number(row.entry_id)],
+    )
+    if (Number(r.rowCount ?? r.affectedRows ?? 0) > 0) {
+      coreOverlap++
+      await deleteRootEmbedding(db, Number(row.entry_id))
+    }
+  }
+  throwIfAborted(signal)
+  if (coreOverlap > 0) {
+    process.stderr.write(
+      `[cycle2] phase_merge core_overlap archived=${coreOverlap} (LLM-judged restatements of user-curated core_entries)\n`,
+    )
+  }
+
+  process.stderr.write(
+    `[cycle2] phase_merge tier1_pairs=${tier1Pairs.length} tier2_pairs=${tier2Pairs.length}` +
+    ` llm_calls=${llmCalls} merged=${merged} core_overlap=${coreOverlap}\n`,
+  )
+
+  return { merged, llm_calls: llmCalls, tier1_pairs: tier1Pairs.length, tier2_pairs: tier2Pairs.length, core_overlap: coreOverlap }
+}
+
+// ─── Current rules digest cache ──────────────────────────────────────────────
+
+let _currentRulesDigest = null
+let _currentRulesDigestTs = 0
+export function loadCurrentRulesDigest() {
+  const now = Date.now()
+  if (_currentRulesDigest && now - _currentRulesDigestTs < 60_000) return _currentRulesDigest
+  const sources = [
+    join(_homedir(), '.claude', 'CLAUDE.md'),
+    join(resourceDir(), 'rules', 'shared', '00-language.md'),
+    join(resourceDir(), 'rules', 'shared', '01-general.md'),
+    join(resourceDir(), 'rules', 'shared', '01-tool.md'),
+    join(resourceDir(), 'rules', 'shared', '04-memory.md'),
+    join(resourceDir(), 'rules', 'shared', '06-team.md'),
+    join(resourceDir(), 'rules', 'shared', '07-workflow.md'),
+  ]
+  const parts = []
+  for (const p of sources) {
+    try {
+      if (!existsSync(p)) continue
+      const txt = readFileSync(p, 'utf8').trim()
+      if (txt) parts.push(`# Source: ${p}\n${txt}`)
+    } catch {}
+  }
+  const joined = parts.join('\n\n---\n\n')
+  const CAP = 40_000
+  _currentRulesDigest = joined.length > CAP ? joined.slice(0, CAP) + '\n…[truncated]' : joined
+  _currentRulesDigestTs = now
+  return _currentRulesDigest
+}
+
+function uniqueIds(values) {
+  return [...new Set(values
+    .map(id => Number(id))
+    .filter(id => Number.isFinite(id)))]
+}
+
+function validateUnifiedGate(parsed, statusById) {
+  const actions = Array.isArray(parsed?.actions) ? parsed.actions : []
+  const primary = actions.filter(a => a?.action !== 'core')
+  const verdictCounts = new Map()
+  for (const action of primary) {
+    const id = Number(action?.entry_id)
+    if (!Number.isFinite(id)) continue
+    verdictCounts.set(id, (verdictCounts.get(id) || 0) + 1)
+  }
+  const expectedIds = [...statusById.keys()]
+  const missingVerdictIds = expectedIds.filter(id => !verdictCounts.has(id))
+  const duplicateVerdictIds = [...verdictCounts.entries()]
+    .filter(([, count]) => count > 1)
+    .map(([id]) => id)
+  const support = parsed?.support instanceof Map ? parsed.support : new Map()
+  const coreIds = new Set(actions
+    .filter(a => a?.action === 'core')
+    .map(a => Number(a.entry_id))
+    .filter(id => Number.isFinite(id)))
+  const missingSupportIds = []
+  const missingCoreIds = []
+  for (const action of primary) {
+    if (!NON_ARCHIVE_VERBS.has(action?.action)) continue
+    const id = Number(action.entry_id)
+    if (!Number.isFinite(id)) continue
+    const coreId = action.action === 'merge' && Number.isFinite(Number(action.target_id))
+      ? Number(action.target_id)
+      : id
+    const hasSupport = support.has(id) || (action.action === 'merge' && support.has(coreId))
+    if (!hasSupport) missingSupportIds.push(id)
+    if (!coreIds.has(coreId)) missingCoreIds.push(id)
+  }
+  return {
+    missingVerdictIds: uniqueIds(missingVerdictIds),
+    duplicateVerdictIds: uniqueIds(duplicateVerdictIds),
+    missingSupportIds: uniqueIds(missingSupportIds),
+    missingCoreIds: uniqueIds(missingCoreIds),
+  }
+}
+
+function gateQualitySummary(quality) {
+  const parts = []
+  if (quality?.missingVerdictIds?.length) parts.push(`missing verdict ids=${quality.missingVerdictIds.join(',')}`)
+  if (quality?.duplicateVerdictIds?.length) parts.push(`duplicate verdict ids=${quality.duplicateVerdictIds.join(',')}`)
+  if (quality?.missingSupportIds?.length) parts.push(`missing why ids=${quality.missingSupportIds.join(',')}`)
+  if (quality?.missingCoreIds?.length) parts.push(`missing core ids=${quality.missingCoreIds.join(',')}`)
+  return parts.join('; ')
+}
+
+function stripUnsupportedPromotions(parsed, unsupportedIds) {
+  const ids = new Set(uniqueIds(unsupportedIds))
+  if (ids.size === 0) return parsed
+  const rejected = new Set(parsed?.rejected || [])
+  for (const id of ids) rejected.add(id)
+  const actions = (parsed?.actions || []).filter(a => {
+    if (a?.action === 'core') return true
+    return !ids.has(Number(a?.entry_id))
+  })
+  return { ...parsed, actions, rejected }
+}
+
+function requiredCoreIdForAction(action) {
+  if (action?.action === 'merge' && Number.isFinite(Number(action.target_id))) {
+    return Number(action.target_id)
+  }
+  return Number(action?.entry_id)
+}
+
+// ─── Unified gate ────────────────────────────────────────────────────────────
+
+// Single LLM pass over rows whose status is in {pending, active}.
+// Returns { actions, rejected, parseOk } following parseUnifiedFormat shape.
+export async function runUnifiedGate(db, rows, activeContext, config = {}, options = {}) {
+  const signal = options?.signal
+  throwIfAborted(signal)
+  if (!rows || rows.length === 0) return { actions: [], rejected: new Set(), parseOk: true }
+  const promptPath = join(resourceDir(), 'defaults', 'memory-promote-prompt.md')
+  if (!existsSync(promptPath)) {
+    throw new Error(`runCycle2: prompt file missing at ${promptPath}`)
+  }
+  const template = readFileSync(promptPath, 'utf8')
+  const userCoreRows = options.dataDir ? await listCore(options.dataDir, '*').catch(() => []) : []
+  throwIfAborted(signal)
+  const sharedPidMap = buildPidMap([activeContext ?? [], rows ?? [], userCoreRows ?? []])
+  const rulesDigest = loadCurrentRulesDigest() || '(no current rules digest available)'
+  const activeCount = activeContext?.length ?? 0
+  const activeCap = options.activeCap ?? CYCLE2_ACTIVE_TARGET_CAP
+
+  const prompt = template
+    .replace('{{CURRENT_RULES}}', rulesDigest)
+    .replace('{{USER_CORE}}', formatUserCoreForPrompt(userCoreRows, sharedPidMap))
+    .replace('{{CORE_MEMORY}}', formatEntriesForPromotePrompt(activeContext, sharedPidMap))
+    .replace('{{ITEMS}}', formatEntriesForPromotePrompt(rows, sharedPidMap, { numbered: true }))
+    .replace('{{ACTIVE_COUNT}}', String(activeCount))
+    .replace('{{ACTIVE_CAP}}', String(activeCap))
+
+  const preset = options.preset || resolveMaintenancePreset('cycle2')
+  const timeout = Number(config?.cycle2?.timeout ?? 600000)
+  const mode = 'cycle2-unified'
+
+  const previewRaw = (raw) => String(raw ?? '').replace(/\s+/g, ' ').slice(0, 200)
+  const callOnce = async (extraTag) => {
+    throwIfAborted(signal)
+    const p = extraTag ? `${prompt}\n\n[retry:${extraTag}]` : prompt
+    const raw = await invokeLlm(p, mode, preset, timeout, options.callLlm)
+    throwIfAborted(signal)
+    return raw
+  }
+
+  const statusById = new Map(rows.map(r => [Number(r.id), String(r.status)]))
+  // Ordinal → batch-id map, keyed by 1-based prompt order (the same order the
+  // numbered Entries block uses). The gate may echo either the real 5-digit
+  // batch id or the row ordinal 1..N; the parser resolves both. The two domains
+  // MUST be disjoint or an ordinal could shadow a real id (ids are 5-digit,
+  // ordinals are <= rows.length, so disjointness always holds in practice).
+  // On a violation a row-number line is indistinguishable from an exact-id
+  // line, so no safe interpretation exists — skip this batch (gate failure)
+  // rather than risk applying a verdict to the wrong entry. The cycle itself
+  // proceeds; the batch re-queues for a later run.
+  const ordinalToId = new Map(rows.map((r, i) => [i + 1, Number(r.id)]))
+  const minBatchId = Math.min(...[...statusById.keys()])
+  if (Number.isFinite(minBatchId) && minBatchId <= rows.length) {
+    process.stderr.write(`[cycle2] batch id ${minBatchId} collides with ordinal range 1..${rows.length} — skipping batch (no safe id resolution)\n`)
+    return { actions: null, rejected: new Set(), parseOk: false }
+  }
+
+  process.stderr.write(`[cycle2-diag] unified prompt=${prompt.length} bytes; rows=${rows.length}\n`)
+
+  let raw
+  try {
+    raw = await callOnce(null)
+  } catch (err) {
+    if (signal?.aborted) throw signal.reason ?? err
+    process.stderr.write(`[cycle2] unified LLM error: ${err.message}\n`)
+    return { actions: null, rejected: new Set(), parseOk: false }
+  }
+  throwIfAborted(signal)
+  process.stderr.write(`[cycle2-diag] unified raw (first 1500): ${String(raw ?? '').replace(/\n/g, '⏎').slice(0, 1500)}\n`)
+
+  let parsed = parseUnifiedFormat(raw, statusById, ordinalToId)
+  let quality = parsed ? validateUnifiedGate(parsed, statusById) : null
+  const qualityIssue = () => gateQualitySummary(quality)
+  if (!parsed || qualityIssue()) {
+    throwIfAborted(signal)
+    const issue = parsed ? qualityIssue() : `unparseable (${previewRaw(raw)})`
+    process.stderr.write(`[cycle2] unified quality retry: ${issue}\n`)
+    // Preserve the first pass before retrying. A retry fired for a mere quality
+    // issue (e.g. a few missing verdicts) must not throw away an otherwise-valid
+    // first-pass parse if the retry comes back unparseable.
+    const firstParsed = parsed
+    const firstQuality = quality
+    try {
+      const retryTag = parsed
+        ? 'complete-verdicts-with-why-and-core-lines'
+        : 'first-field-must-be-the-listed-row-number'
+      const raw2 = await callOnce(retryTag)
+      const retryParsed = parseUnifiedFormat(raw2, statusById, ordinalToId)
+      if (retryParsed) {
+        parsed = retryParsed
+        quality = validateUnifiedGate(retryParsed, statusById)
+      } else if (firstParsed) {
+        process.stderr.write(`[cycle2] unparseable after retry — falling back to first-pass parse (${previewRaw(raw2)})\n`)
+        parsed = firstParsed
+        quality = firstQuality
+      } else {
+        process.stderr.write(`[cycle2] unparseable after retry — skipping batch (${previewRaw(raw2)})\n`)
+        return { actions: null, rejected: new Set(), parseOk: false }
+      }
+    } catch (err) {
+      if (signal?.aborted) throw signal.reason ?? err
+      if (firstParsed) {
+        process.stderr.write(`[cycle2] retry LLM error: ${err.message} — falling back to first-pass parse\n`)
+        parsed = firstParsed
+        quality = firstQuality
+      } else {
+        process.stderr.write(`[cycle2] retry LLM error: ${err.message}\n`)
+        return { actions: null, rejected: new Set(), parseOk: false }
+      }
+    }
+  }
+  const finalIssue = gateQualitySummary(quality)
+  // duplicateVerdictIds are genuinely ambiguous (the same row got two conflicting
+  // verbs) — keep the full-skip. missingVerdictIds, by contrast, used to skip the
+  // WHOLE batch, so a handful of persistently-missing poison rows could livelock
+  // the gate. Partial-apply instead: keep the valid verdicts we did receive, just
+  // log the missing ids and leave those rows for a later run.
+  if (quality?.duplicateVerdictIds?.length) {
+    process.stderr.write(`[cycle2] duplicate verdict coverage after retry — skipping batch (${finalIssue})\n`)
+    return { actions: null, rejected: new Set(), parseOk: false }
+  }
+  if (quality?.missingVerdictIds?.length) {
+    process.stderr.write(`[cycle2] missing verdicts after retry — partial apply, leaving ids=${quality.missingVerdictIds.join(',')} for a later run (${finalIssue})\n`)
+  }
+  // A response made up solely of why/core lines parses "ok" yet carries zero
+  // primary (status-verb) verdicts. Without this guard parseOk stays true and
+  // the caller treats the batch as a clean no-op, masking the coverage failure
+  // and marking the rows reviewed. Fail the parse so the rows are re-queued.
+  const primaryCount = (parsed.actions || []).filter(a => a?.action !== 'core').length
+  if (rows.length > 0 && primaryCount === 0) {
+    process.stderr.write(`[cycle2] gate produced zero primary verdicts for ${rows.length} rows — failing parse\n`)
+    return { actions: null, rejected: new Set(), parseOk: false, missingIds: [...statusById.keys()] }
+  }
+  const incompletePromotionIds = uniqueIds([
+    ...(quality?.missingSupportIds || []),
+    ...(quality?.missingCoreIds || []),
+  ])
+  if (incompletePromotionIds.length > 0) {
+    process.stderr.write(`[cycle2] incomplete non-archive verdicts rejected after retry ids=${incompletePromotionIds.join(',')} (${finalIssue})\n`)
+    parsed = stripUnsupportedPromotions(parsed, incompletePromotionIds)
+  }
+  return {
+    actions: parsed.actions,
+    rejected: parsed.rejected,
+    parseOk: true,
+    missingIds: quality?.missingVerdictIds || [],
+  }
+}
+
+// ─── Sonnet cascade ──────────────────────────────────────────────────────────
+
+// Sonnet re-judge over first-pass keep verdicts. Sonnet sees rules + summary
+// and returns binary keep/drop. Failures fail-open (preserve first-pass).
+async function sonnetCascade(candidates, rulesDigest, options = {}) {
+  const signal = options?.signal
+  throwIfAborted(signal)
+  if (!candidates || candidates.length === 0) return new Map()
+  const lines = candidates.map(c =>
+    `id:${c.id} status:${c.status} verb:${c.verb} cat:${c.category} el:${c.element} sm:${String(c.summary || '').slice(0, 200)}${c.core ? ` core:${String(c.core).slice(0, 200)}` : ''}`,
+  ).join('\n')
+  const prompt = [
+    `Final gate over first-pass keep verdicts.`,
+    `Keep a candidate ONLY if it lands in one of three layers: L1 relationship/communication`,
+    `(user identity, address form, reply-style preferences, disliked patterns); L2 behavior rules`,
+    `(principles the user corrected/insisted on, hard safety boundaries, quality bars); or L3 current`,
+    `map (one-line project-landscape summaries, live long-running goals, environment anchors documented`,
+    `nowhere else). For a past decision/failure, keep only the one-line lesson that still constrains`,
+    `behavior, else drop. DROP anything whose source of truth is code, rules files, or skill docs, plus`,
+    `implementation specs, code-internal constants, measurements, resolved-bug stories, status snapshots,`,
+    `and duplicates of source-of-truth rules.`,
+    `When a candidate has a core: field, judge THAT extracted one-line lesson (the entry will live as`,
+    `that line), not the raw narrative in el:/sm:.`,
+    ``,
+    `Source-of-truth rules (excerpt — DO NOT duplicate in memory):`,
+    String(rulesDigest || '').slice(0, 4000),
+    ``,
+    `Candidates:`,
+    lines,
+    ``,
+    `Reply one line per id: "<id>|keep" to retain, "<id>|drop" to reject.`,
+    `NO prose, NO preamble, NO meta-commentary. First character must be a digit.`,
+  ].join('\n')
+
+  // Hardcoded — resolveMaintenancePreset falls back to first preset (HAIKU)
+  // when no binding exists, which would defeat the cascade. SONNET HIGH
+  // matches the worker pool's default preset id from agent-config.
+  const preset = options.cascadePreset || 'SONNET HIGH'
+  const llmCall = typeof options?.callLlm === 'function' ? options.callLlm : callBridgeLlm
+  let raw
+  try {
+    raw = await llmCall({
+      role: 'cycle2-agent',
+      taskType: 'maintenance',
+      mode: 'cycle2-cascade',
+      preset,
+      timeout: 600000,
+      cwd: null,
+    }, prompt)
+  } catch (err) {
+    if (signal?.aborted) throw signal.reason ?? err
+    process.stderr.write(`[cycle2] cascade error: ${err.message} — fail-open\n`)
+    return new Map()
+  }
+  throwIfAborted(signal)
+
+  const verdicts = new Map()
+  for (const line of String(raw ?? '').split('\n')) {
+    throwIfAborted(signal)
+    const trimmed = line.trim()
+    if (!trimmed) continue
+    if (trimmed.startsWith('//') || trimmed.startsWith('#') || trimmed.startsWith('```')) continue
+    const parts = trimmed.split('|')
+    if (parts.length < 2) continue
+    const id = Number(parts[0].trim())
+    const v = parts[1].trim().toLowerCase()
+    if (Number.isFinite(id) && (v === 'keep' || v === 'drop')) verdicts.set(id, v)
+  }
+  process.stderr.write(`[cycle2] cascade evaluated=${candidates.length} drops=${[...verdicts.values()].filter(v => v === 'drop').length}\n`)
+  return verdicts
+}
+
+// ─── runCycle2 ───────────────────────────────────────────────────────────────
+
+const _runCycle2InFlight = new WeakMap()
+
+export async function runCycle2(db, config = {}, options = {}, dataDir = null) {
+  const signal = options?.signal
+  throwIfAborted(signal)
+  const partial = {
+    promoted: 0, archived: 0, merged: 0, updated: 0, kept: 0, rejected_verb: 0,
+    merge_rejected: 0,
+    missing_core_summary: 0,
+    core_embedding_backfill: 0,
+    rescore: { updated: 0 },
+    phase_merge: { merged: 0, llm_calls: 0, tier1_pairs: 0, tier2_pairs: 0, core_overlap: 0 },
+    cascade: { evaluated: 0, dropped: 0 },
+  }
+  if (_runCycle2InFlight.has(db)) {
+    process.stderr.write('[cycle2] skipped: already in flight for this db\n')
+    return { ok: true, ...partial, skippedInFlight: true }
+  }
+  const client = await db._pool.connect()
+  let gotLock = false
+  try {
+    throwIfAborted(signal)
+    const r = await client.query(`SELECT pg_try_advisory_lock(hashtext($1)) AS got`, ['mixdog.cycle2'])
+    gotLock = r.rows[0]?.got === true
+  } catch (err) {
+    client.release()
+    if (signal?.aborted) throw signal.reason ?? err
+    process.stderr.write(`[cycle2] advisory lock query failed: ${err.message}\n`)
+    return { ok: true, ...partial, skippedInFlight: true }
+  }
+  if (!gotLock) {
+    client.release()
+    process.stderr.write('[cycle2] skipped: advisory lock held by another worker\n')
+    return { ok: true, ...partial, skippedInFlight: true }
+  }
+  const _p = (async () => {
+    try {
+      const result = await _runCycle2Impl(db, config, options, dataDir)
+      return { ok: true, ...result }
+    } catch (e) {
+      if (signal?.aborted) throw signal.reason ?? e
+      return { ok: false, error: e.message, ...partial }
+    } finally {
+      try { await client.query(`SELECT pg_advisory_unlock(hashtext($1))`, ['mixdog.cycle2']) } catch {}
+      client.release()
+    }
+  })()
+  _runCycle2InFlight.set(db, _p)
+  try { return await _p }
+  finally { _runCycle2InFlight.delete(db) }
+}
+
+async function _runCycle2Impl(db, config = {}, options = {}, dataDir = null) {
+  const signal = options?.signal
+  throwIfAborted(signal)
+  const batchSize = Math.max(1, Number(config.batch_size ?? 50))
+  const activeTargetCap = Number.isFinite(Number(config.active_target_cap))
+    ? Math.max(1, Number(config.active_target_cap))
+    : CYCLE2_ACTIVE_TARGET_CAP
+  const nowMs = Date.now()
+
+  const stats = {
+    promoted: 0, archived: 0, merged: 0,
+    updated: 0, kept: 0, rejected_verb: 0,
+    merge_rejected: 0,
+    missing_core_summary: 0,
+    core_embedding_backfill: 0,
+    rescore: { updated: 0 },
+    phase_merge: { merged: 0, llm_calls: 0, tier1_pairs: 0, tier2_pairs: 0, core_overlap: 0 },
+    cascade: { evaluated: 0, dropped: 0 },
+  }
+
+  if (dataDir) {
+    try {
+      stats.core_embedding_backfill = await backfillCoreEmbeddings(dataDir, { signal })
+      throwIfAborted(signal)
+    } catch (err) {
+      if (signal?.aborted) throw signal.reason ?? err
+      process.stderr.write(`[cycle2] core embedding backfill failed: ${err.message}\n`)
+    }
+  }
+
+  const activeCountRes = await db.query(
+    `SELECT COUNT(*) AS c FROM entries WHERE is_root = 1 AND status = 'active'`,
+    [],
+  )
+  throwIfAborted(signal)
+  const activeCount = Number(activeCountRes.rows[0]?.c ?? 0)
+  const reviewActiveRows = activeCount > activeTargetCap
+
+  // Rolling active re-review quota. Under cap, the unified selection below
+  // pulls only pending rows, so an already-promoted entry that later drifts
+  // stale or turns out to restate a rule file never gets re-judged — the
+  // over-cap path was historically the ONLY one that re-examined active.
+  // Reserve a bounded slice of batch slots for the stalest active rows so
+  // rule-duplicate / drifted promotions are archived continuously instead of
+  // sitting forever un-rechecked. Bounded count + reviewed_at rotation
+  // prevents eroding the set to zero (the original over-cap-only concern):
+  // only the oldest few are re-judged per cycle, and the gate — shown
+  // {{CURRENT_RULES}} — keeps genuine A/B entries and archives only
+  // restatements. Embedding dedup is skipped on purpose: rule restatements
+  // are often cross-language paraphrases whose cosine never clears the merge
+  // threshold, but the LLM gate catches the semantic overlap.
+  const activeRecheckQuota = reviewActiveRows
+    ? 0
+    : Math.max(0, Math.min(Number(config.active_recheck_quota ?? 8), batchSize - 1))
+  const pendingLimit = batchSize - activeRecheckQuota
+  // Score direction depends on the phase. Under cap we are SEEDING the active
+  // set: evaluate the highest-value pending first so promotion-worthy rows
+  // reach the gate instead of starving behind low-score cycle1 churn. Over cap
+  // we are CONTRACTING: evaluate the lowest-score rows first to shed the
+  // weakest. (The active-recheck slice below stays ASC — demote weakest active
+  // first.)
+  const scoreDir = reviewActiveRows ? 'ASC' : 'DESC'
+
+  // Unified candidate selection. Pending rows (and, when over cap, active
+  // rows) reach the gate here; the reserved active-recheck slice is appended
+  // below. Cleanup of duplicates/stale user-core overlap also runs via
+  // phase_merge / cycle3.
+  const rowsRes = await db.query(`
+    SELECT id, element, category, summary, score, last_seen_at, project_id, status
+    FROM entries
+    WHERE is_root = 1
+      AND (status = 'pending' OR ($2::boolean AND status = 'active'))
+    ORDER BY
+      CASE status WHEN 'pending' THEN 0 WHEN 'active' THEN 1 END ASC,
+      reviewed_at ASC NULLS FIRST,
+      error_count ASC,
+      score ${scoreDir},
+      id ASC
+    LIMIT $1
+  `, [pendingLimit, reviewActiveRows])
+  throwIfAborted(signal)
+  const rows = rowsRes.rows
+
+  // Append the reserved rolling slice of stalest active rows (under-cap only;
+  // the over-cap branch already pulls active broadly). De-duped against the
+  // primary selection so an id never gets two verdicts in one batch.
+  if (activeRecheckQuota > 0 && activeCount > 0) {
+    const seen = new Set(rows.map(r => Number(r.id)))
+    const recheckRes = await db.query(`
+      SELECT id, element, category, summary, score, last_seen_at, project_id, status
+      FROM entries
+      WHERE is_root = 1 AND status = 'active'
+      ORDER BY reviewed_at ASC NULLS FIRST, score ASC, id ASC
+      LIMIT $1
+    `, [activeRecheckQuota])
+    throwIfAborted(signal)
+    for (const r of recheckRes.rows) {
+      if (!seen.has(Number(r.id))) rows.push(r)
+    }
+  }
+
+  // Active snapshot for prompt context (do-not-duplicate reference).
+  const activeContextRes = await db.query(`
+    SELECT id, element, category, summary, score, last_seen_at, project_id, status
+    FROM entries
+    WHERE is_root = 1 AND status = 'active'
+    ORDER BY score DESC, last_seen_at DESC, id ASC
+    LIMIT 100
+  `, [])
+  throwIfAborted(signal)
+  const activeContext = activeContextRes.rows
+
+  const gateResult = rows.length > 0
+    ? await runUnifiedGate(db, rows, activeContext, config, { activeCap: activeTargetCap, preset: options.preset, dataDir, signal, callLlm: options.callLlm })
+    : { actions: [], rejected: new Set(), parseOk: true }
+  throwIfAborted(signal)
+  // Surface a gate parse/coverage failure so the caller can distinguish a
+  // clean no-op run from one where the LLM gate produced nothing usable.
+  if (gateResult.parseOk === false) stats.gate_failed = true
+
+  const sweepCursor = nowMs
+
+  const rowsById = new Map(rows.map(r => [Number(r.id), r]))
+
+  // Cascade pre-pass: pull first-pass keeps (verb 'active') into Sonnet for
+  // re-judge. update/merge/archived skip.
+  const cascadeCandidates = []
+  if (gateResult.actions) {
+    // First-pass proposed core lines: under the pending-row transform the L2
+    // lesson lives only in the core line, so thread it into the cascade.
+    const proposedCoreById = new Map()
+    for (const a of gateResult.actions) {
+      if (a.action !== 'core') continue
+      const id = Number(a.entry_id)
+      const core = String(a.core_summary ?? '').replace(/\s+/g, ' ').trim()
+      if (Number.isFinite(id) && core) proposedCoreById.set(id, core)
+    }
+    for (const a of gateResult.actions) {
+      throwIfAborted(signal)
+      if (a.action !== 'active') continue
+      const row = rowsById.get(Number(a.entry_id))
+      if (!row) continue
+      cascadeCandidates.push({
+        id: row.id, status: row.status, verb: a.action,
+        category: row.category, element: row.element, summary: row.summary,
+        core: proposedCoreById.get(Number(a.entry_id)) || '',
+      })
+    }
+  }
+
+  const rulesDigest = loadCurrentRulesDigest() || ''
+  let cascadeVerdicts = new Map()
+  if (cascadeCandidates.length > 0) {
+    cascadeVerdicts = await sonnetCascade(cascadeCandidates, rulesDigest, { ...options, signal })
+    throwIfAborted(signal)
+    stats.cascade.evaluated = cascadeCandidates.length
+  }
+
+  // Apply actions.
+  if (gateResult.actions) {
+    const reviewedIds = []
+    const rejectedActionIds = []
+    const cascadeDropArchiveIds = []
+    const statusBatch = []
+    const coreSummaryById = new Map()
+    const primaryActions = []
+
+    for (const a of gateResult.actions) {
+      throwIfAborted(signal)
+      if (a.action === 'core') {
+        const id = Number(a.entry_id)
+        const core = String(a.core_summary ?? '').replace(/\s+/g, ' ').trim().slice(0, CORE_SUMMARY_MAX)
+        if (Number.isFinite(id) && core) coreSummaryById.set(id, core)
+      } else {
+        primaryActions.push(a)
+      }
+    }
+
+    const setCoreSummary = async (entryId, explicitSummary) => {
+      const id = Number(entryId)
+      if (!Number.isFinite(id)) return false
+      let core = String(explicitSummary ?? '').replace(/\s+/g, ' ').trim().slice(0, CORE_SUMMARY_MAX)
+      if (!core) return false
+      await db.query(`UPDATE entries SET core_summary = $1 WHERE id = $2 AND is_root = 1`, [core, id])
+      return true
+    }
+
+    for (const a of primaryActions) {
+      throwIfAborted(signal)
+      const id = Number(a.entry_id)
+      if (!Number.isFinite(id)) continue
+      const row = rowsById.get(id)
+      if (!row) continue
+      let accepted = false
+
+      try {
+        const requiresCore = NON_ARCHIVE_VERBS.has(a.action)
+        const coreId = requiredCoreIdForAction(a)
+        const explicitCore = coreSummaryById.get(coreId) || coreSummaryById.get(id)
+        if (requiresCore && !explicitCore) {
+          stats.missing_core_summary += 1
+          rejectedActionIds.push(id)
+          process.stderr.write(`[cycle2] non-archive action rejected: missing explicit core line id=${id} action=${a.action}\n`)
+          continue
+        }
+
+        // Cascade override: drop a tentatively-kept entry → archive.
+        if (a.action === 'active' && cascadeVerdicts.get(id) === 'drop') {
+          cascadeDropArchiveIds.push(id)
+          accepted = true
+          reviewedIds.push(id)
+          continue
+        }
+
+        if (a.action === 'active') {
+          if (row.status === 'pending') {
+            statusBatch.push({ entry_id: id, new_status: 'active', was_pending: true })
+          } else if (row.status === 'active') {
+            stats.kept += 1
+          }
+          await setCoreSummary(id, explicitCore)
+          accepted = true
+        } else if (a.action === 'archived') {
+          statusBatch.push({ entry_id: id, new_status: 'archived', was_pending: row.status === 'pending' })
+          accepted = true
+        } else if (a.action === 'update') {
+          if (await applyUpdate(db, id, a.element, a.summary, { signal })) stats.updated += 1
+          await setCoreSummary(id, explicitCore)
+          accepted = true
+        } else if (a.action === 'merge') {
+          const sourceIds = Array.isArray(a.source_ids) ? a.source_ids : []
+          const targetId = Number(a.target_id)
+          if (!Number.isFinite(targetId) || sourceIds.length === 0) {
+            stats.merge_rejected += 1
+            rejectedActionIds.push(id)
+            continue
+          }
+          if (targetId !== id && !sourceIds.map(Number).includes(id)) {
+            stats.merge_rejected += 1
+            rejectedActionIds.push(id)
+            process.stderr.write(
+              `[cycle2] merge rejected during apply: id=${id} target=${targetId} sources=${sourceIds.join(',')}\n`,
+            )
+            continue
+          }
+          // Bounded-erosion invariant: a merge may only consolidate entries
+          // that are themselves candidates in this batch. Otherwise a single
+          // rechecked active row could list source_ids pointing at active
+          // entries outside the batch (e.g. ids drawn from the activeContext
+          // reference list), and applyMerge would archive those too —
+          // un-judged and beyond the rolling-recheck quota. Out-of-batch
+          // target/source ids are rejected; a true duplicate of an existing
+          // active entry is handled by the `archived` verdict instead.
+          if (![targetId, ...sourceIds.map(Number)].every(mid => rowsById.has(mid))) {
+            stats.merge_rejected += 1
+            rejectedActionIds.push(id)
+            process.stderr.write(
+              `[cycle2] merge rejected: out-of-batch target/source (target=${targetId} sources=${sourceIds.join(',')})\n`,
+            )
+            continue
+          }
+          const moved = await applyMerge(db, targetId, sourceIds, { signal })
+          throwIfAborted(signal)
+          if (moved > 0) {
+            stats.merged += moved
+            if (typeof a.element === 'string' || typeof a.summary === 'string') {
+              try { if (await applyUpdate(db, targetId, a.element, a.summary, { signal })) stats.updated += 1 }
+              catch (err) {
+                if (signal?.aborted) throw signal.reason ?? err
+                process.stderr.write(`[cycle2] merge target update failed (target=${targetId}): ${err.message}\n`)
+              }
+            }
+            await setCoreSummary(targetId, explicitCore)
+            accepted = true
+          } else {
+            stats.merge_rejected += 1
+            rejectedActionIds.push(id)
+          }
+        }
+        if (accepted) reviewedIds.push(id)
+      } catch (err) {
+        if (signal?.aborted) throw signal.reason ?? err
+        process.stderr.write(`[cycle2] action error (id=${id}): ${err.message}\n`)
+      }
+    }
+
+    if (statusBatch.length > 0) {
+      // Status verdicts are applied as one SQL batch; checkpoint before the
+      // batch and then again at the next cycle2 unit boundary.
+      throwIfAborted(signal)
+      const batchRes = await applyBatchStatusVerdicts(db, statusBatch, nowMs)
+      stats.promoted += batchRes.promoted
+      stats.archived += batchRes.archived
+    }
+
+    if (cascadeDropArchiveIds.length > 0) {
+      throwIfAborted(signal)
+      const r = await db.query(`UPDATE entries SET status = 'archived' WHERE id = ANY($1::bigint[]) AND is_root = 1`, [cascadeDropArchiveIds])
+      stats.cascade.dropped += Number(r.rowCount ?? r.affectedRows ?? 0)
+      stats.archived += Number(r.rowCount ?? r.affectedRows ?? 0)
+    }
+    if (reviewedIds.length > 0) {
+      throwIfAborted(signal)
+      await db.query(`UPDATE entries SET reviewed_at = $1 WHERE id = ANY($2::bigint[])`, [sweepCursor, reviewedIds])
+    }
+    if (rejectedActionIds.length > 0) {
+      throwIfAborted(signal)
+      await db.query(
+        `UPDATE entries SET error_count = COALESCE(error_count, 0) + 1 WHERE id = ANY($1::bigint[])`,
+        [[...new Set(rejectedActionIds)]],
+      )
+    }
+  } else if (rows.length > 0) {
+    // Parse failure — bump error_count, do not advance reviewed_at.
+    for (const r of rows) {
+      throwIfAborted(signal)
+      try {
+        await db.query(
+          `UPDATE entries SET error_count = COALESCE(error_count, 0) + 1 WHERE id = $1`,
+          [r.id],
+        )
+      } catch {}
+    }
+  }
+
+  // Rejected verb rows: advance reviewed_at + bump error_count so an all-reject
+  // batch does not loop forever. error_count ASC sort pushes them to the back.
+  if (gateResult.rejected && gateResult.rejected.size > 0) {
+    stats.rejected_verb = gateResult.rejected.size
+    for (const id of gateResult.rejected) {
+      throwIfAborted(signal)
+      try {
+        await db.query(`UPDATE entries SET reviewed_at = $1 WHERE id = $2`, [sweepCursor, id])
+        await db.query(
+          `UPDATE entries SET error_count = COALESCE(error_count, 0) + 1 WHERE id = $1`,
+          [id],
+        )
+      } catch {}
+    }
+  }
+
+  // Flush embeddings BEFORE phase_merge: newly promoted/dirty roots have
+  // NULL embeddings until the dirty queue drains, and runPhaseMerge filters
+  // on `embedding IS NOT NULL` for both the cosine dedup and the core-overlap
+  // pass. Running the flush after the merge would skip those rows for an
+  // entire cycle. Reordering ensures same-cycle dedup/core-overlap sees them.
+  try {
+    throwIfAborted(signal)
+    const d = await flushEmbeddingDirty(db, { signal })
+    throwIfAborted(signal)
+    if (d.attempted > 0) {
+      process.stderr.write(
+        `[cycle2] embedding flush attempted=${d.attempted} ok=${d.succeeded} failed=${d.failed.length}\n`,
+      )
+    }
+  } catch (err) {
+    if (signal?.aborted) throw signal.reason ?? err
+    process.stderr.write(`[cycle2] embedding flush failed: ${err.message}\n`)
+  }
+
+  // phase_merge: cosine dedup over active entries.
+  const phaseMergeStats = await runPhaseMerge(db, { ...options, signal })
+  throwIfAborted(signal)
+  stats.phase_merge = phaseMergeStats
+
+  // Active-cap enforcement is delegated to the gate (phases 1-3): the prompt
+  // exposes Active/cap counts and instructs aggressive `archived` verdicts on
+  // overflow. No deterministic safety net here — if the gate ever fails to
+  // contain growth, fix the prompt, not bolt a fallback back on.
+
+  process.stderr.write(
+    `[cycle2] rescore=${stats.rescore.updated}` +
+    ` core_backfill=${stats.core_embedding_backfill}` +
+    ` active=${activeCount}/${activeTargetCap} review_active=${reviewActiveRows ? 1 : 0}` +
+    ` | gate promoted=${stats.promoted} archived=${stats.archived}` +
+    ` updated=${stats.updated} kept=${stats.kept}` +
+    ` rejected_verb=${stats.rejected_verb} merge_rejected=${stats.merge_rejected}` +
+    ` missing_core=${stats.missing_core_summary}` +
+    ` | cascade eval=${stats.cascade.evaluated} drop=${stats.cascade.dropped}` +
+    ` | phase_merge merged=${stats.phase_merge.merged} core_overlap=${stats.phase_merge.core_overlap || 0}` +
+    ` llm=${stats.phase_merge.llm_calls}\n`,
+  )
+
+  return stats
+}
+
+export function parseInterval(s) {
+  if (String(s).toLowerCase() === 'immediate') return 0
+  const match = String(s).match(/^(\d+)(s|m|h)$/)
+  if (!match) throw new Error(`[memory-cycle2] invalid interval config: ${s}`)
+  const [, num, unit] = match
+  const multiplier = { s: 1000, m: 60000, h: 3600000 }
+  return Number(num) * multiplier[unit]
+}