docs(search): record FTS removal benchmarks

docs/cn/standard/chatlab-import.md

@@ -13,7 +13,7 @@ outline: deep
 - **Push 模式**（本文档）：外部系统主动将数据推送到 ChatLab 的导入接口。适用于脚本集成、一次性文件导入等场景。
 - **[Pull 模式](./chatlab-pull.md)**：第三方暴露标准 HTTP 端点，ChatLab 主动拉取数据。**推荐的第三方集成方式。**
 
-两种模式底层共用同一套导入逻辑（去重、meta/members 更新、FTS 索引），数据格式统一为 [ChatLab Format](./chatlab-format.md)。
+两种模式底层共用同一套导入逻辑（去重、meta/members 更新、会话索引维护），数据格式统一为 [ChatLab Format](./chatlab-format.md)。
 
 :::
 

docs/cn/standard/chatlab-pull.md

@@ -239,7 +239,7 @@ ChatLab 按用户配置的间隔，定期对已订阅的对话执行增量拉取
 GET {baseUrl}/sessions/{sessionId}/messages?format=chatlab&since={lastPullAt}
 ```
 
-远程数据源返回 `since` 之后的增量消息。ChatLab 通过内部导入管道处理（去重、meta/members 更新、FTS 索引等全部复用 Push 模式逻辑）。
+远程数据源返回 `since` 之后的增量消息。ChatLab 通过内部导入管道处理（去重、meta/members 更新、会话索引维护等全部复用 Push 模式逻辑）。
 
 ---
 

docs/en/standard/chatlab-import.md

@@ -13,7 +13,7 @@ This document defines the standard import protocol for external data sources to
 - **Push mode** (this document): The external system actively pushes data to ChatLab's import endpoint. Suitable for script integrations and one-time file imports.
 - **[Pull mode](./chatlab-pull.md)**: A third-party exposes standard HTTP endpoints and ChatLab pulls data on demand. **The recommended integration approach for third-party tools.**
 
-Both modes share the same underlying import pipeline (deduplication, meta/members update, FTS indexing). Data format is unified as [ChatLab Format](./chatlab-format.md).
+Both modes share the same underlying import pipeline (deduplication, meta/members updates, and session-index maintenance). Data format is unified as [ChatLab Format](./chatlab-format.md).
 
 :::
 

docs/en/standard/chatlab-pull.md

@@ -239,7 +239,7 @@ ChatLab periodically performs incremental pulls on subscribed conversations at t
 GET {baseUrl}/sessions/{sessionId}/messages?format=chatlab&since={lastPullAt}
 ```
 
-The remote data source returns incremental messages since `lastPullAt`. ChatLab processes them through the internal import pipeline (deduplication, meta/members update, FTS indexing — all the same as Push mode).
+The remote data source returns incremental messages since `lastPullAt`. ChatLab processes them through the internal import pipeline (deduplication, meta/members updates, and session-index maintenance — all the same as Push mode).
 
 ---
 

packages/node-runtime/src/semantic-index/retrieval/evidence.ts

@@ -2,7 +2,7 @@
  * 证据块组装
  *
  * chunking-decision-final.md 第 14 节：
- * - RRF 后取 topK 命中 chunk，每个命中在 parent 内前后各扩展若干条消息，不跨 parent。
+ * - 语义检索后取 topK 命中 chunk，每个命中在 parent 内前后各扩展若干条消息，不跨 parent。
  * - 相邻命中可合并，但合并后超过单块 soft cap 则拆开注入。
  * - 总证据预算固定 1500 token：不够时优先减少前后扩展，再减少 chunk 数。
  *

packages/tools/src/tool-metadata.ts

@@ -25,7 +25,7 @@ const ENGLISH_TOOL_METADATA: Record<string, EnglishToolMetadata> = {
   },
   deep_search_messages: {
     description:
-      'Search messages using slower substring matching. Use this for exact full-text matches or when regular search misses results.',
+      'Search messages using slower substring matching. Use this for exact phrase matches or when regular search misses results.',
     properties: {
       keywords: 'Keywords to search for.',
       sender_id: 'Filter by sender ID obtained from get_members.',

scripts/bench-streaming-import.mts

@@ -0,0 +1,579 @@
+/**
+ * End-to-end streaming import benchmark.
+ *
+ * Each measured run executes in a fresh process and writes fresh databases.
+ * Fixture generation and result verification are excluded from the timed range.
+ *
+ * Usage:
+ *   pnpm exec tsx scripts/bench-streaming-import.mts single 100000 [runs=3]
+ *   pnpm exec tsx scripts/bench-streaming-import.mts batch 100 10000 [runs=3] [concurrency=1]
+ *
+ * Single-file runs also append 1,000 messages incrementally and measure a
+ * representative LIKE query against the final database.
+ */
+
+import { createHash } from 'node:crypto'
+import { once } from 'node:events'
+import { copyFileSync, createWriteStream, linkSync, mkdirSync, rmSync, statSync, unlinkSync } from 'node:fs'
+import path from 'node:path'
+import { fileURLToPath } from 'node:url'
+import { spawnSync } from 'node:child_process'
+import Database from 'better-sqlite3'
+import { CHAT_DB_TABLES } from '@openchatlab/core'
+import { getNativeParserStatus, isNativeFormatAvailable, type NativeParserStatus } from '@openchatlab/parser'
+import { BetterSqliteAdapter } from '../packages/node-runtime/src/better-sqlite3-adapter'
+import { computeAndSetOverviewCache } from '../packages/node-runtime/src/cache/session-cache'
+import { runKeyedBatch } from '../packages/node-runtime/src/import/batch-coordinator'
+import { incrementalImport, type IncrementalImportDeps } from '../packages/node-runtime/src/import/incremental-importer'
+import {
+  streamingImport,
+  type ImportLogger,
+  type ImportStageTimings,
+  type StreamImportDeps,
+} from '../packages/node-runtime/src/import/streaming-importer'
+import { createChatLabTempDir } from './chatlab-temp.mjs'
+
+const scriptPath = fileURLToPath(import.meta.url)
+const nativeBinding = path.resolve('apps/cli/native/better_sqlite3.node')
+const memberCount = 200
+const sampleTexts = [
+  '今天天气不错，我们出去玩吧！',
+  '哈哈哈哈哈哈这也太好笑了',
+  '[图片]',
+  '好的，收到，明天见。',
+  'This is a mixed language message with some English words 和中文混排。',
+  '周末有人一起打球吗？地点老地方，时间下午三点，人齐就开打。',
+]
+
+interface BenchmarkParser {
+  formatId: 'weflow'
+  implementation: 'rust-native'
+  nativeModuleAvailable: true
+}
+
+interface BenchmarkResult {
+  scenario: 'single' | 'batch'
+  parser: BenchmarkParser
+  fileCount: number
+  concurrency: number
+  inputMessages: number
+  inputBytes: number
+  durationMs: number
+  incrementalMessages: number
+  incrementalDurationMs: number
+  sampledPeakRssMb: number
+  sampledRssDeltaMb: number
+  databaseBytes: number
+  messagesWritten: number
+  membersWritten: number
+  outputSignature: string
+  likeSearch: {
+    coldMs: number
+    warmMedianMs: number
+    resultCount: number
+    resultSignature: string
+  }
+  timings: ImportStageTimings
+}
+
+export interface BenchmarkParserMonitor {
+  logger: ImportLogger
+  assertRustNativeCompleted(importIndex: number): void
+}
+
+export function resolveBenchmarkParser(status: NativeParserStatus, nativeFormatAvailable: boolean): BenchmarkParser {
+  if (!status.available || !nativeFormatAvailable) {
+    const reason = status.disabled
+      ? 'CHATLAB_DISABLE_NATIVE_PERF=1 disables the native parser'
+      : status.error
+        ? `native parser failed to load: ${status.error}`
+        : 'the native parser does not provide the WeFlow kernel'
+    throw new Error(
+      `Streaming import benchmark requires the Rust Native WeFlow parser; ${reason}. ` +
+        'Run pnpm build:native before collecting a baseline.'
+    )
+  }
+
+  return {
+    formatId: 'weflow',
+    implementation: 'rust-native',
+    nativeModuleAvailable: true,
+  }
+}
+
+export function createBenchmarkParserMonitor(): BenchmarkParserMonitor {
+  let nativeStarted = false
+  let fallbackDetected = false
+  const ignoreNonParserLog = () => undefined
+
+  const inspectMessage = (message: string) => {
+    if (message.includes('[NativeParser] Parsing WeFlow export with Rust kernel')) {
+      nativeStarted = true
+    }
+    if (message.includes('[NativeParser]') && message.includes('falling back to TS parser')) {
+      fallbackDetected = true
+    }
+  }
+
+  return {
+    logger: {
+      info: inspectMessage,
+      error: inspectMessage,
+      perf: ignoreNonParserLog,
+      perfDetail: ignoreNonParserLog,
+      summary: ignoreNonParserLog,
+      reset() {
+        nativeStarted = false
+        fallbackDetected = false
+      },
+      init: ignoreNonParserLog,
+      getCurrentLogFile() {
+        return null
+      },
+    },
+    assertRustNativeCompleted(importIndex: number) {
+      if (fallbackDetected) {
+        throw new Error(`Import ${importIndex} fell back to the TypeScript WeFlow parser; benchmark result rejected`)
+      }
+      if (!nativeStarted) {
+        throw new Error(`Import ${importIndex} did not start the Rust Native WeFlow parser; benchmark result rejected`)
+      }
+    },
+  }
+}
+
+const zeroTimings = (): ImportStageTimings => ({
+  detectionMs: 0,
+  preprocessingMs: 0,
+  databaseSetupMs: 0,
+  parserMs: 0,
+  metaWriteMs: 0,
+  memberWriteMs: 0,
+  messageWriteMs: 0,
+  nicknameHistoryMs: 0,
+  indexCreationMs: 0,
+  checkpointMs: 0,
+  sessionIndexMs: 0,
+  postImportHookMs: 0,
+  totalMs: 0,
+})
+
+function addTimings(target: ImportStageTimings, source: ImportStageTimings): void {
+  for (const key of Object.keys(target) as Array<keyof ImportStageTimings>) target[key] += source[key]
+}
+
+async function generateWeflowFixture(filePath: string, count: number, startIndex = 0): Promise<void> {
+  const stream = createWriteStream(filePath, { encoding: 'utf-8' })
+  const write = async (chunk: string) => {
+    if (!stream.write(chunk)) await once(stream, 'drain')
+  }
+
+  await write(
+    `{"weflow":{"version":"1.0.0","exportedAt":1704164645},` +
+      `"session":{"wxid":"bench@chatroom","nickname":"性能测试群","displayName":"性能测试群","type":"群聊"},` +
+      '"avatars":{},"messages":['
+  )
+
+  const batch: string[] = []
+  for (let index = 0; index < count; index++) {
+    const absoluteIndex = startIndex + index
+    const member = absoluteIndex % memberCount
+    batch.push(
+      JSON.stringify({
+        localId: absoluteIndex + 1,
+        createTime: 1_704_164_645 + absoluteIndex * 3,
+        formattedTime: '2024-01-02 03:04:05',
+        type: '文本消息',
+        localType: 1,
+        content: `${sampleTexts[absoluteIndex % sampleTexts.length]} #${absoluteIndex}`,
+        isSend: member === 0 ? 1 : 0,
+        senderUsername: `wxid_member_${member}`,
+        senderDisplayName: `成员${member}号`,
+        senderAvatarKey: `wxid_member_${member}`,
+        source: '',
+      })
+    )
+    if (batch.length >= 5000) {
+      await write((index + 1 > batch.length ? ',' : '') + batch.join(','))
+      batch.length = 0
+    }
+  }
+  if (batch.length > 0) await write((count > batch.length ? ',' : '') + batch.join(','))
+  await write(']}')
+  stream.end()
+  await once(stream, 'finish')
+}
+
+function createIncrementalDeps(dbPath: string): IncrementalImportDeps {
+  return {
+    openDatabase() {
+      const raw = new Database(dbPath, { nativeBinding })
+      raw.pragma('journal_mode = WAL')
+      raw.pragma('synchronous = NORMAL')
+      return new BetterSqliteAdapter(raw)
+    },
+    onProgress() {
+      /* benchmark excludes UI transport */
+    },
+  }
+}
+
+function createImportDeps(dbPath: string, cacheDir: string, logger: ImportLogger): StreamImportDeps {
+  return {
+    openDatabase() {
+      const raw = new Database(dbPath, { nativeBinding })
+      raw.pragma('journal_mode = WAL')
+      raw.pragma('synchronous = NORMAL')
+      raw.exec(CHAT_DB_TABLES)
+      return new BetterSqliteAdapter(raw)
+    },
+    deleteDatabase() {
+      for (const suffix of ['', '-wal', '-shm']) {
+        try {
+          unlinkSync(dbPath + suffix)
+        } catch {
+          /* ignore cleanup failures in a disposable benchmark workspace */
+        }
+      }
+    },
+    onProgress() {
+      /* benchmark excludes UI transport */
+    },
+    logger,
+    postImportHook(db, sessionId) {
+      computeAndSetOverviewCache(db, sessionId, cacheDir)
+    },
+  }
+}
+
+export function inspectDatabase(dbPath: string): {
+  messages: number
+  members: number
+  signature: string
+} {
+  const db = new Database(dbPath, { readonly: true, nativeBinding })
+  const hash = createHash('sha256')
+  const hashRows = (label: string, query: string): number => {
+    hash.update(`table:${label}\n`)
+    let count = 0
+    let chunk = ''
+    for (const row of db.prepare(query).raw().iterate() as Iterable<unknown[]>) {
+      chunk += `${JSON.stringify(row)}\n`
+      count++
+      if (chunk.length >= 1024 * 1024) {
+        hash.update(chunk)
+        chunk = ''
+      }
+    }
+    if (chunk) hash.update(chunk)
+    return count
+  }
+
+  try {
+    hashRows(
+      'meta',
+      `SELECT name, platform, type, group_id, group_avatar, owner_id, schema_version, session_gap_threshold
+       FROM meta ORDER BY rowid`
+    )
+    const members = hashRows(
+      'member',
+      `SELECT id, platform_id, account_name, group_nickname, aliases, avatar, roles
+       FROM member ORDER BY id`
+    )
+    hashRows(
+      'member_name_history',
+      `SELECT id, member_id, name_type, name, start_ts, end_ts
+       FROM member_name_history ORDER BY id`
+    )
+    const messages = hashRows(
+      'message',
+      `SELECT id, sender_id, sender_account_name, sender_group_nickname, ts, type, content,
+              reply_to_message_id, platform_message_id
+       FROM message ORDER BY id`
+    )
+    hashRows(
+      'segment',
+      `SELECT id, start_ts, end_ts, message_count, is_manual, summary
+       FROM segment ORDER BY id`
+    )
+    hashRows(
+      'message_context',
+      `SELECT message_id, segment_id, topic_id
+       FROM message_context ORDER BY message_id`
+    )
+    const legacyFts = db.prepare("SELECT name FROM sqlite_master WHERE type = 'table' AND name = 'message_fts'").get()
+    if (legacyFts) throw new Error('Benchmark database unexpectedly contains message_fts')
+    return { messages, members, signature: hash.digest('hex') }
+  } finally {
+    db.close()
+  }
+}
+
+function measureLikeSearch(dbPath: string): BenchmarkResult['likeSearch'] {
+  const db = new Database(dbPath, { readonly: true, nativeBinding })
+  const statement = db.prepare(
+    `SELECT id, content
+     FROM message
+     WHERE content LIKE ? ESCAPE '\\'
+     ORDER BY ts DESC, id DESC
+     LIMIT 50`
+  )
+  // Match only the oldest synthetic row so ORDER BY cannot stop after finding
+  // a few recent matches. This measures the existing LIKE path's full-scan cost.
+  const params = ['%#0']
+  const run = () => {
+    const startedAt = performance.now()
+    const rows = statement.all(...params) as Array<{ id: number; content: string }>
+    return { durationMs: performance.now() - startedAt, rows }
+  }
+
+  try {
+    const cold = run()
+    const warmRuns = Array.from({ length: 5 }, run).sort((left, right) => left.durationMs - right.durationMs)
+    const warmMedian = warmRuns[Math.floor(warmRuns.length / 2)]
+    return {
+      coldMs: cold.durationMs,
+      warmMedianMs: warmMedian.durationMs,
+      resultCount: cold.rows.length,
+      resultSignature: createHash('sha256').update(JSON.stringify(cold.rows)).digest('hex').slice(0, 16),
+    }
+  } finally {
+    db.close()
+  }
+}
+
+async function runWorker(fileCount: number, messagesPerFile: number, concurrency: number): Promise<BenchmarkResult> {
+  if (typeof global.gc !== 'function') {
+    throw new Error('Benchmark worker requires Node.js --expose-gc')
+  }
+
+  const parser = resolveBenchmarkParser(getNativeParserStatus(), isNativeFormatAvailable('weflow'))
+  const root = createChatLabTempDir('bench', 'streaming-import-')
+  const fixturePath = path.join(root, 'fixture.json')
+  const incrementalFixturePath = path.join(root, 'incremental.json')
+  const cacheDir = path.join(root, 'cache')
+  mkdirSync(cacheDir, { recursive: true })
+
+  try {
+    await generateWeflowFixture(fixturePath, messagesPerFile)
+    const incrementalMessages = fileCount === 1 ? 1000 : 0
+    if (incrementalMessages > 0) {
+      await generateWeflowFixture(incrementalFixturePath, incrementalMessages, messagesPerFile)
+    }
+    const inputPaths = [fixturePath]
+    for (let index = 1; index < fileCount; index++) {
+      const linkedPath = path.join(root, `fixture-${index}.json`)
+      try {
+        linkSync(fixturePath, linkedPath)
+      } catch {
+        copyFileSync(fixturePath, linkedPath)
+      }
+      inputPaths.push(linkedPath)
+    }
+
+    global.gc()
+    const rssStartMb = process.memoryUsage().rss / 1024 / 1024
+    const timings = zeroTimings()
+    let sampledPeakRssMb = rssStartMb
+    const dbPaths: string[] = []
+    const startedAt = performance.now()
+
+    await runKeyedBatch(
+      inputPaths.map((inputPath, index) => ({
+        value: { inputPath, index },
+        key: `session:${index}`,
+      })),
+      {
+        concurrency,
+        async run({ inputPath, index }) {
+          const dbPath = path.join(root, `session-${index}.db`)
+          dbPaths[index] = dbPath
+          const parserMonitor = createBenchmarkParserMonitor()
+          const result = await streamingImport(
+            inputPath,
+            createImportDeps(dbPath, cacheDir, parserMonitor.logger),
+            { formatId: 'weflow' },
+            `bench-${index}`
+          )
+          if (!result.success || !result.diagnostics?.performance) {
+            throw new Error(`Import ${index} failed: ${result.error ?? 'missing performance diagnostics'}`)
+          }
+          parserMonitor.assertRustNativeCompleted(index)
+          addTimings(timings, result.diagnostics.performance.timings)
+          sampledPeakRssMb = Math.max(sampledPeakRssMb, result.diagnostics.performance.rssSampledPeakMb)
+        },
+      }
+    )
+    const durationMs = performance.now() - startedAt
+
+    let incrementalDurationMs = 0
+    if (incrementalMessages > 0) {
+      const incrementalStartedAt = performance.now()
+      const result = await incrementalImport('bench-0', incrementalFixturePath, createIncrementalDeps(dbPaths[0]), {
+        formatId: 'weflow',
+      })
+      incrementalDurationMs = performance.now() - incrementalStartedAt
+      if (!result.success || result.newMessageCount !== incrementalMessages) {
+        throw new Error(
+          `Incremental import failed: expected ${incrementalMessages}, got ${result.newMessageCount} (${result.error ?? 'unknown error'})`
+        )
+      }
+    }
+
+    let messagesWritten = 0
+    let membersWritten = 0
+    let databaseBytes = 0
+    const signatures: string[] = []
+    let likeSearch: BenchmarkResult['likeSearch'] | undefined
+    for (const dbPath of dbPaths) {
+      const inspected = inspectDatabase(dbPath)
+      messagesWritten += inspected.messages
+      membersWritten += inspected.members
+      signatures.push(inspected.signature)
+      databaseBytes += statSync(dbPath).size
+      likeSearch ??= measureLikeSearch(dbPath)
+    }
+
+    return {
+      scenario: fileCount === 1 ? 'single' : 'batch',
+      parser,
+      fileCount,
+      concurrency,
+      inputMessages: fileCount * messagesPerFile,
+      inputBytes: statSync(fixturePath).size * fileCount,
+      durationMs,
+      incrementalMessages,
+      incrementalDurationMs,
+      sampledPeakRssMb,
+      sampledRssDeltaMb: Math.max(0, sampledPeakRssMb - rssStartMb),
+      databaseBytes,
+      messagesWritten,
+      membersWritten,
+      outputSignature: createHash('sha256').update(signatures.sort().join(':')).digest('hex'),
+      likeSearch: likeSearch!,
+      timings,
+    }
+  } finally {
+    rmSync(root, { recursive: true, force: true })
+  }
+}
+
+function runIsolated(fileCount: number, messagesPerFile: number, concurrency: number): BenchmarkResult {
+  const child = spawnSync(
+    process.execPath,
+    [
+      '--expose-gc',
+      ...process.execArgv,
+      scriptPath,
+      '--worker',
+      String(fileCount),
+      String(messagesPerFile),
+      String(concurrency),
+    ],
+    {
+      cwd: process.cwd(),
+      encoding: 'utf-8',
+      maxBuffer: 10 * 1024 * 1024,
+      env: { ...process.env },
+    }
+  )
+  if (child.status !== 0) {
+    throw new Error(`Benchmark worker failed (${child.status}):\n${child.stdout}\n${child.stderr}`)
+  }
+  const resultLine = child.stdout.split('\n').find((line) => line.startsWith('BENCH_RESULT '))
+  if (!resultLine) throw new Error(`Benchmark worker returned no result:\n${child.stdout}\n${child.stderr}`)
+  return JSON.parse(resultLine.slice('BENCH_RESULT '.length)) as BenchmarkResult
+}
+
+function formatBytes(bytes: number): string {
+  return `${(bytes / 1024 / 1024).toFixed(1)} MB`
+}
+
+function printResult(result: BenchmarkResult, label: string): void {
+  console.log(
+    `${label}: parser ${result.parser.implementation} | ${result.durationMs.toFixed(0)} ms | ` +
+      `incremental ${result.incrementalDurationMs.toFixed(0)} ms/${result.incrementalMessages} rows | ` +
+      `sampled peak RSS ${result.sampledPeakRssMb.toFixed(0)} MB ` +
+      `(+${result.sampledRssDeltaMb.toFixed(0)} MB) | DB ${formatBytes(result.databaseBytes)} | ` +
+      `LIKE ${result.likeSearch.coldMs.toFixed(1)} ms cold/${result.likeSearch.warmMedianMs.toFixed(1)} ms warm | ` +
+      `signature ${result.outputSignature}`
+  )
+}
+
+async function main(): Promise<void> {
+  const [mode, first, second, third, fourth] = process.argv.slice(2)
+  if (mode === '--worker') {
+    const result = await runWorker(Number(first), Number(second), Number(third))
+    console.log(`BENCH_RESULT ${JSON.stringify(result)}`)
+    return
+  }
+
+  let fileCount: number
+  let messagesPerFile: number
+  let runs: number
+  let concurrency: number
+  if (mode === 'single') {
+    fileCount = 1
+    messagesPerFile = Number(first ?? 100_000)
+    runs = Number(second ?? 3)
+    concurrency = 1
+  } else if (mode === 'batch') {
+    fileCount = Number(first ?? 100)
+    messagesPerFile = Number(second ?? 10_000)
+    runs = Number(third ?? 3)
+    concurrency = Number(fourth ?? 1)
+  } else {
+    throw new Error('Usage: single <messages> [runs=3] | batch <files> <messages-per-file> [runs=3] [concurrency=1]')
+  }
+  if (![fileCount, messagesPerFile, runs, concurrency].every((value) => Number.isInteger(value) && value > 0)) {
+    throw new Error('All benchmark counts must be positive integers')
+  }
+
+  console.log(
+    `Benchmarking ${fileCount} file(s), ${messagesPerFile.toLocaleString()} messages/file, concurrency ${concurrency}, ${runs} isolated run(s)`
+  )
+  const results: BenchmarkResult[] = []
+  for (let index = 0; index < runs; index++) {
+    const result = runIsolated(fileCount, messagesPerFile, concurrency)
+    results.push(result)
+    printResult(result, `run ${index + 1}`)
+  }
+
+  const sorted = [...results].sort((left, right) => left.durationMs - right.durationMs)
+  const median = sorted[Math.floor(sorted.length / 2)]
+  const parserSignatures = new Set(results.map((result) => JSON.stringify(result.parser)))
+  if (parserSignatures.size !== 1) {
+    throw new Error('Parser implementations differ between isolated runs')
+  }
+  if (results.some((result) => result.outputSignature !== median.outputSignature)) {
+    throw new Error('Output signatures differ between isolated runs')
+  }
+  if (
+    results.some(
+      (result) =>
+        result.likeSearch.resultCount !== 1 || result.likeSearch.resultSignature !== median.likeSearch.resultSignature
+    )
+  ) {
+    throw new Error('LIKE search results differ between isolated runs or do not match the expected fixture row')
+  }
+  const expectedMessages = fileCount * messagesPerFile + median.incrementalMessages
+  if (median.messagesWritten !== expectedMessages) {
+    throw new Error(`Expected ${expectedMessages} messages, got ${median.messagesWritten}`)
+  }
+
+  printResult(median, 'median')
+  const stageRows = Object.entries(median.timings)
+    .filter(([stage]) => stage !== 'totalMs')
+    .sort(([, left], [, right]) => right - left)
+  console.log('Median phase totals:')
+  for (const [stage, durationMs] of stageRows) {
+    console.log(`  ${stage.padEnd(22)} ${durationMs.toFixed(1).padStart(10)} ms`)
+  }
+}
+
+if (path.resolve(process.argv[1] ?? '') === scriptPath) {
+  void main().catch((error: unknown) => {
+    console.error(error)
+    process.exitCode = 1
+  })
+}