agentworkshop 0.8.3 → 0.8.5

package/scripts/e2e-pi-simulator-paper.mjs

@@ -0,0 +1,2317 @@
+#!/usr/bin/env node
+/**
+ * Bounded integration evidence for AgentWorkShop's registered `pi` harness.
+ * Pi receives a deterministic OpenAI-compatible model response from a local
+ * loopback fixture; the actual Pi CLI, Pi extension, AgentWorkShop bridge,
+ * protected approval API, simulator, device readback, and run-tagged sampling
+ * remain in the execution path. No external model or provider is used.
+ */
+import { spawn, spawnSync } from 'node:child_process'
+import { createHash, randomBytes } from 'node:crypto'
+import { chmodSync, closeSync, copyFileSync, cpSync, existsSync, mkdirSync, mkdtempSync, openSync, readFileSync, readdirSync, symlinkSync, writeFileSync } from 'node:fs'
+import { createServer } from 'node:net'
+import { createServer as createHttpServer } from 'node:http'
+import { dirname, isAbsolute, join, relative, resolve, sep } from 'node:path'
+import { fileURLToPath } from 'node:url'
+import { setTimeout as delay } from 'node:timers/promises'
+import { resolveRunMode } from '../shared/config/home.mjs'
+
+const REPO = resolve(dirname(fileURLToPath(import.meta.url)), '..')
+const AW_PORT = 3997
+const SIM_PORT = 4020
+const MODEL_API_PORT = 4021
+const MQTT_PORT = 18840
+const OFFLINE_TSDB_PORT = 45321
+const OFFLINE_OBJECTSTORE_PORT = 45322
+const OFFLINE_APP_MQTT_QUEUE_PORT = 45323
+const FORBIDDEN_PORT = 4010
+const APP_BASE = `http://127.0.0.1:${AW_PORT}`
+const SIM_BASE = `http://127.0.0.1:${SIM_PORT}`
+const SIM_SOURCE = resolve(REPO, 'plc-node-simulator')
+const VERSION = JSON.parse(readFileSync(join(REPO, 'package.json'), 'utf8')).version
+const TAG = `pi${Date.now().toString(36)}`
+mkdirSync(join(REPO, '.omc'), { recursive: true })
+const ROOT = mkdtempSync(join(REPO, '.omc', 'registered-pi-paper-'))
+const HOME = join(ROOT, 'aw-home')
+const EVIDENCE_PATH = join(ROOT, 'evidence.json')
+const PI_AGENT_DIR = join(HOME, 'pi-agent')
+const OUTPUT_BASE = join(REPO, 'docs', 'audit', 'evidence', '2026-10-25-current-cross-scene')
+const OUTPUT_DIR = join(OUTPUT_BASE, `r23-pi-registered-harness-${new Date().toISOString().replace(/[-:.]/g, '').replace('T', '-').slice(0, 15)}-${randomBytes(3).toString('hex')}`)
+const PI_MODELS_PATH = join(PI_AGENT_DIR, 'models.json')
+const SIM_ROOT = join(ROOT, 'simulator')
+const SIM_LOG = join(ROOT, 'simulator.log')
+const APP_LOG = join(ROOT, 'agentworkshop.log')
+const EMPTY_DOTENV = join(HOME, 'empty.env')
+const ISOLATED_PROMPTS_DIR = join(HOME, 'prompts')
+const NUXT_CLI = join(REPO, 'node_modules', 'nuxt', 'bin', 'nuxt.mjs')
+const RAG_BRIDGE_DISABLED_BASE_URL = 'http://paper-harness-outbound-disabled.invalid'
+const SIM_LOG_FD = openSync(SIM_LOG, 'a')
+const APP_LOG_FD = openSync(APP_LOG, 'a')
+const HARNESS_MODE = 'pi-local-fixture'
+const PI_PROVIDER = 'agentworkshop-local'
+const PI_MODEL = 'aw-deterministic-paper-fixture'
+const PI_COMMAND = join(dirname(process.execPath), 'pi.cmd')
+const PI_CLI_PACKAGE_PATH = join(dirname(process.execPath), 'node_modules', '@mariozechner', 'pi-coding-agent', 'package.json')
+const PI_APPROVAL_TIMEOUT_MS = 180_000
+const PI_TASK_COMPLETION_TIMEOUT_MS = 180_000
+const PI_PROMPT_TIMEOUT_MS = 180_000
+// Keep the shared evidence helpers usable while this runner selects only the
+// local Pi fixture mode; no OMP CLI or external model is invoked.
+const REAL_OMP_PROVIDER = PI_PROVIDER
+const REAL_OMP_MODEL = PI_MODEL
+const REAL_OMP_TRIALS_PER_SCENE = 1
+const REAL_OMP_APPROVAL_TIMEOUT_MS = PI_APPROVAL_TIMEOUT_MS
+const REAL_OMP_TASK_COMPLETION_TIMEOUT_MS = PI_TASK_COMPLETION_TIMEOUT_MS
+const PI_API_TRACE_PATH = join(ROOT, 'pi-model-api-trace.json')
+const piApiTrace = []
+let piFixtureTarget = null
+let piApiServer
+
+const evidence = {
+  schemaVersion: 1,
+  experiment: {
+    id: TAG,
+    projectVersion: VERSION,
+    campaign: 'Pi registered-harness local-fixture cross-scene E2E',
+    harnessMode: HARNESS_MODE,
+    provider: PI_PROVIDER,
+    model: PI_MODEL,
+    trialsPerScene: 1,
+    runOutcomes: [],
+    modelFixture: 'deterministic local OpenAI-compatible streaming API; fixture scripts one recipe_trial call and one complete_task call per scene; no LLM or external provider',
+    startedAt: new Date().toISOString(),
+    postTrialSnapshotDelaySeconds: 20,
+    scenes: [],
+  },
+  isolation: {
+    result: 'pending',
+    awHome: 'fresh isolated AW_HOME',
+    appBase: APP_BASE,
+    simulatorBase: SIM_BASE,
+    simulatorMqttLoopbackPort: MQTT_PORT,
+    appMqttQueueOfflineProbePort: OFFLINE_APP_MQTT_QUEUE_PORT,
+    explicitLoopbackOnly: true,
+    externalLlmDisabled: true,
+    modelApiBase: `http://127.0.0.1:${MODEL_API_PORT}/v1`,
+    forbiddenServicePortExcluded: true,
+    checks: [],
+  },
+  checks: [],
+  status: 'running',
+}
+
+let appProcess
+let simProcess
+let currentStage = 'preflight'
+let currentScene
+let authToken
+let syntheticAdminId
+let isolatedPluginNames = []
+let expectedPluginNames = []
+const allowedEnabledPluginNames = ['rag-bridge']
+let projectPluginTreeAtPrepare
+let projectPromptTreeAtPrepare
+let promptTreeAtPrepare
+const generatedSessionPassword = randomBytes(32).toString('hex')
+const children = new Set()
+const allocatedDevicePorts = []
+const CHILD_OS_ENV_KEYS = process.platform === 'win32'
+  ? ['PATH', 'PATHEXT', 'SystemRoot', 'WINDIR', 'TEMP', 'TMP', 'USERPROFILE', 'APPDATA', 'LOCALAPPDATA', 'HOMEDRIVE', 'HOMEPATH', 'ProgramData']
+  : ['PATH', 'HOME', 'TMPDIR', 'TMP', 'TEMP', 'LANG', 'LC_ALL']
+const CHILD_EXTRA_ENV_KEYS = new Set(['SIM_PORT', 'MQTT_BROKER_PORT'])
+const RELEASE_CHECK_SUFFIXES = ['-process-stopped', '-port-released']
+
+function check(id, pass, detail = {}) {
+  evidence.checks.push({ id, status: pass ? 'pass' : 'fail', detail })
+  if (!pass) throw new Error(`check failed: ${id}`)
+}
+
+function noteIsolation(id, pass, detail = {}) {
+  evidence.isolation.checks.push({ id, status: pass ? 'pass' : 'fail', detail })
+  if (!pass) throw new Error(`isolation preflight failed: ${id}`)
+}
+
+function safeText(value, limit = 500) {
+  let text = String(value ?? '')
+  if (authToken) text = text.split(authToken).join('[REDACTED]')
+  text = text.replace(/\but-[a-f\d]{32}\b/gi, '[REDACTED_TOKEN]')
+  text = text.replace(/\bBearer\s+[^\s,;]+/gi, 'Bearer [REDACTED]')
+  text = text.replace(/\b(?:api[_-]?key|access[_-]?token|client[_-]?secret|password|secret)\s*[:=]\s*["']?[^\s,"'}]+/gi, '[REDACTED_CREDENTIAL]')
+  text = text.replace(/\bsk-[A-Za-z0-9_-]{16,}\b/g, '[REDACTED_CREDENTIAL]')
+  return text.slice(0, limit)
+}
+
+function compactError(error) {
+  const message = error instanceof Error ? error.message : String(error)
+  return { name: error instanceof Error ? error.name : 'Error', message: safeText(message) }
+}
+
+async function readRequestJson(req) {
+  const chunks = []
+  let bytes = 0
+  for await (const chunk of req) {
+    bytes += chunk.length
+    if (bytes > 2_000_000) throw new Error('model fixture rejected request larger than 2 MB')
+    chunks.push(chunk)
+  }
+  return JSON.parse(Buffer.concat(chunks).toString('utf8'))
+}
+
+function sendSse(res, chunks) {
+  res.writeHead(200, {
+    'content-type': 'text/event-stream; charset=utf-8',
+    'cache-control': 'no-cache, no-transform',
+    'connection': 'keep-alive',
+  })
+  for (const chunk of chunks) res.write(`data: ${JSON.stringify(chunk)}\n\n`)
+  res.write('data: [DONE]\n\n')
+  res.end()
+}
+
+function chatChunk(id, delta, finishReason = null, usage) {
+  return {
+    id,
+    object: 'chat.completion.chunk',
+    created: Math.floor(Date.now() / 1000),
+    model: PI_MODEL,
+    choices: [{ index: 0, delta, finish_reason: finishReason }],
+    ...(usage ? { usage } : {}),
+  }
+}
+
+async function startPiModelFixture() {
+  piApiServer = createHttpServer(async (req, res) => {
+    if (req.method !== 'POST' || req.url !== '/v1/chat/completions') {
+      res.writeHead(404, { 'content-type': 'application/json' })
+      res.end(JSON.stringify({ error: 'fixture route not found' }))
+      return
+    }
+    try {
+      const body = await readRequestJson(req)
+      const target = piFixtureTarget
+      const messages = Array.isArray(body.messages) ? body.messages : []
+      const toolNames = (Array.isArray(body.tools) ? body.tools : [])
+        .map(tool => String(tool?.function?.name ?? ''))
+        .filter(Boolean)
+      const roles = messages.map(message => String(message?.role ?? ''))
+      const toolMessages = messages.filter(message => String(message?.role ?? '') === 'tool')
+      const latestToolMessage = toolMessages.at(-1)
+      const returnedToolContent = latestToolMessage?.content
+      const returnedToolText = typeof returnedToolContent === 'string'
+        ? returnedToolContent
+        : JSON.stringify(returnedToolContent ?? '')
+      const callIndex = target ? piApiTrace.filter(item => item.scene === target.scene).length : 0
+      const requestId = `fixture-${target?.scene ?? 'unassigned'}-${callIndex + 1}`
+      const requestEvidence = {
+        requestId,
+        scene: target?.scene ?? null,
+        requestModel: safeText(body.model, 120),
+        apiPath: req.url,
+        loopbackPeer: req.socket.remoteAddress === '127.0.0.1' || req.socket.remoteAddress === '::ffff:127.0.0.1',
+        apiKeyMatchesLocalFixture: req.headers.authorization === 'Bearer local-fixture-key-not-a-secret',
+        streamRequested: body.stream === true,
+        toolNames,
+        messageRoles: roles,
+        toolResultMessageCount: toolMessages.length,
+        toolResultCallId: safeText(latestToolMessage?.tool_call_id ?? '', 120),
+        toolResultText: toolMessages.length ? safeText(returnedToolText, 1200) : null,
+        response: null,
+      }
+      if (!target || body.model !== PI_MODEL || !requestEvidence.loopbackPeer || !requestEvidence.apiKeyMatchesLocalFixture || body.stream !== true) {
+        throw new Error('model fixture rejected a request outside the expected loopback Pi custom-provider configuration')
+      }
+      const id = `chatcmpl-${requestId}`
+      if (callIndex === 0) {
+        const promptText = JSON.stringify(messages)
+        if (!toolNames.includes('recipe_trial')
+          || roles.includes('tool')
+          || !promptText.includes(target.recipeId)
+          || !promptText.includes(target.nodeId)
+          || !promptText.includes(String(target.value))) {
+          throw new Error(`${target.scene} initial Pi request did not include the registered recipe_trial declaration and task identifiers`)
+        }
+        const args = JSON.stringify({
+          recipe_id: target.recipeId,
+          params: [{ node_id: target.nodeId, value: target.value }],
+          hypothesis: 'Deterministic local fixture selection for a bounded simulator integration check.',
+        })
+        requestEvidence.response = { type: 'tool_call', toolName: 'recipe_trial', candidate: target.value }
+        piApiTrace.push(requestEvidence)
+        sendSse(res, [
+          chatChunk(id, { role: 'assistant' }),
+          chatChunk(id, { tool_calls: [{ index: 0, id: `call-${target.scene}`, type: 'function', function: { name: 'recipe_trial', arguments: args } }] }),
+          chatChunk(id, {}, 'tool_calls', { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 }),
+        ])
+        return
+      }
+      if (callIndex === 1) {
+        if (!roles.includes('tool') || !toolNames.includes('recipe_trial')) {
+          throw new Error(`${target.scene} Pi follow-up request did not contain the platform tool result`)
+        }
+        if (toolMessages.length !== 1 || !requestEvidence.toolResultText || requestEvidence.toolResultCallId !== `call-${target.scene}`) {
+          throw new Error(`${target.scene} Pi follow-up did not preserve exactly one non-empty platform tool result`)
+        }
+        requestEvidence.response = { type: 'tool_call', toolName: 'complete_task', taskId: target.taskId }
+        piApiTrace.push(requestEvidence)
+        sendSse(res, [
+          chatChunk(id, { role: 'assistant' }),
+          chatChunk(id, { tool_calls: [{ index: 0, id: `complete-${target.scene}`, type: 'function', function: { name: 'complete_task', arguments: JSON.stringify({
+            task_id: target.taskId,
+            summary: 'Completed the bounded local simulator integration check after the approved recipe trial returned its verified command result.',
+            deliverable: `Task ${target.taskId}: the registered Pi host-tool bridge returned the recipe_trial result for candidate ${target.value}; this is integration evidence, not a process-improvement claim.`,
+          }) } }] }),
+          chatChunk(id, {}, 'tool_calls', { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 }),
+        ])
+        return
+      }
+      if (callIndex === 2) {
+        if (!roles.includes('tool') || toolMessages.length !== 2 || !requestEvidence.toolResultText
+          || requestEvidence.toolResultCallId !== `complete-${target.scene}`
+          || !/COMPLETED/.test(requestEvidence.toolResultText)) {
+          throw new Error(`${target.scene} Pi follow-up did not preserve the successful complete_task result`)
+        }
+        requestEvidence.response = { type: 'final_text_after_task_completion' }
+        piApiTrace.push(requestEvidence)
+        sendSse(res, [
+          chatChunk(id, { role: 'assistant' }),
+          chatChunk(id, { content: 'The bounded simulator integration task is complete. The approved recipe trial returned a verified command result; no process improvement is claimed.' }),
+          chatChunk(id, {}, 'stop', { prompt_tokens: 1, completion_tokens: 1, total_tokens: 2 }),
+        ])
+        return
+      }
+      throw new Error(`${target.scene} made an unexpected extra local model request`)
+    }
+    catch (error) {
+      const message = error instanceof Error ? error.message : String(error)
+      res.writeHead(400, { 'content-type': 'application/json' })
+      res.end(JSON.stringify({ error: { message: safeText(message, 300) } }))
+    }
+  })
+  await new Promise((resolvePromise, reject) => {
+    piApiServer.once('error', reject)
+    piApiServer.listen(MODEL_API_PORT, '127.0.0.1', resolvePromise)
+  })
+  const address = piApiServer.address()
+  if (!address || typeof address === 'string' || address.address !== '127.0.0.1' || address.port !== MODEL_API_PORT) {
+    throw new Error('local Pi model fixture did not bind exactly to the approved IPv4 loopback endpoint')
+  }
+  noteIsolation('local-pi-model-api-bound-loopback', true, { host: address.address, port: address.port, route: '/v1/chat/completions' })
+}
+
+function assertLoopbackUrl(url, expectedPort) {
+  const parsed = new URL(url)
+  if (parsed.protocol !== 'http:' || parsed.hostname !== '127.0.0.1' || Number(parsed.port) !== expectedPort || Number(parsed.port) === FORBIDDEN_PORT) {
+    throw new Error('network boundary rejected a non-approved loopback URL')
+  }
+}
+
+function fixtureTracePath(sceneId) {
+  if (!['cast-film', 'annealing'].includes(sceneId)) throw new Error('fixture trace scene is not approved')
+  const root = resolve(ROOT)
+  const tracePath = resolve(root, `fixture-${sceneId}.json`)
+  const relativePath = relative(root, tracePath)
+  if (isAbsolute(relativePath) || relativePath === '..' || relativePath.startsWith(`..${sep}`)) {
+    throw new Error('fixture trace path escaped the experiment root')
+  }
+  return tracePath
+}
+
+async function assertFreeLoopback(port, label) {
+  if (!Number.isInteger(port) || port < 1024 || port > 65535 || port === FORBIDDEN_PORT) {
+    throw new Error(`port allocation rejected for ${label}`)
+  }
+  const server = createServer()
+  await new Promise((resolvePromise, reject) => {
+    server.once('error', reject)
+    server.listen(port, '127.0.0.1', resolvePromise)
+  })
+  await new Promise(resolvePromise => server.close(resolvePromise))
+}
+
+async function preflight() {
+  if (VERSION !== '0.8.0') throw new Error(`expected current package version 0.8.0; found ${safeText(VERSION, 80)}`)
+  if (!existsSync(join(SIM_SOURCE, 'src', 'server', 'index.ts')) || !existsSync(join(SIM_SOURCE, 'node_modules')) || !existsSync(NUXT_CLI)
+    || !existsSync(PI_COMMAND) || !existsSync(PI_CLI_PACKAGE_PATH)) {
+    throw new Error('required registered Pi CLI, AgentWorkShop app, or simulator runtime is missing')
+  }
+  if (existsSync(HOME)) throw new Error('fresh AW_HOME path unexpectedly exists')
+  const mode = resolveRunMode({ cwd: REPO, packageRoot: REPO, env: { AW_HOME: HOME, AW_MODE: 'home' } })
+  if (mode.mode !== 'home' || resolve(mode.home) !== resolve(HOME) || resolve(mode.dataDir) !== resolve(join(HOME, 'data'))) {
+    throw new Error('AW_HOME did not resolve to a new isolated home-mode data directory')
+  }
+  noteIsolation('fresh-home-mode-resolution', true, { mode: mode.mode, dataDir: mode.dataDir })
+  for (const [label, port] of [['app', AW_PORT], ['simulator-http', SIM_PORT], ['local-model-api', MODEL_API_PORT], ['simulator-mqtt', MQTT_PORT], ['offline-app-mqtt-queue-probe', OFFLINE_APP_MQTT_QUEUE_PORT], ['offline-timescale-probe', OFFLINE_TSDB_PORT], ['offline-objectstore-probe', OFFLINE_OBJECTSTORE_PORT]]) {
+    await assertFreeLoopback(port, label)
+    noteIsolation(`${label}-free-loopback-port`, true, { port, host: '127.0.0.1' })
+  }
+  if ([AW_PORT, SIM_PORT, MODEL_API_PORT, MQTT_PORT].includes(FORBIDDEN_PORT)) {
+    throw new Error('forbidden service port appeared in explicit service configuration')
+  }
+  noteIsolation('project-version-and-local-inputs', true, {
+    version: VERSION,
+    simulatorSourcePresent: true,
+    piCommandPresent: true,
+    piVersion: JSON.parse(readFileSync(PI_CLI_PACKAGE_PATH, 'utf8')).version,
+    modelApiBoundToLoopback: true,
+    noExternalModelCredentialsRequired: true,
+    nuxtCliPresent: true,
+    noUserOrBusinessArchiveInput: true,
+  })
+  noteIsolation('strict-child-environment-allowlist', true, {
+    osKeys: CHILD_OS_ENV_KEYS,
+    harnessMode: HARNESS_MODE,
+    piAgentDirectoryIsolated: true,
+    piOfflineMode: true,
+    noStorageServiceEnvironment: true,
+  })
+}
+
+function cleanChildEnv(extra = {}) {
+  const env = {}
+  for (const key of CHILD_OS_ENV_KEYS) {
+    if (process.env[key] !== undefined) env[key] = process.env[key]
+  }
+  env.AW_HOME = HOME
+  env.AW_MODE = 'home'
+  env.AW_PROMPTS_DIR = ISOLATED_PROMPTS_DIR
+  env.HOME = HOME
+  env.PI_CODING_AGENT_DIR = PI_AGENT_DIR
+  env.PI_CODING_AGENT_SESSION_DIR = join(PI_AGENT_DIR, 'sessions')
+  env.PI_OFFLINE = '1'
+  env.NUXT_SESSION_PASSWORD = generatedSessionPassword
+  env.NO_PROXY = '127.0.0.1,localhost'
+  env.no_proxy = '127.0.0.1,localhost'
+  for (const [key, value] of Object.entries(extra)) {
+    if (!CHILD_EXTRA_ENV_KEYS.has(key)) throw new Error(`child environment rejected unsupported override ${key}`)
+    env[key] = String(value)
+  }
+  return env
+}
+
+function gitText(cwd, args) {
+  const result = spawnSync('git', args, {
+    cwd,
+    windowsHide: true,
+    encoding: 'utf8',
+    stdio: ['ignore', 'pipe', 'pipe'],
+    env: cleanChildEnv(),
+  })
+  if (result.error || result.status !== 0) {
+    throw new Error(`could not capture source identity: ${safeText(result.stderr || result.error?.message || result.status)}`)
+  }
+  return String(result.stdout ?? '').trim()
+}
+
+function hashGitFiles(cwd, roots) {
+  const names = gitText(cwd, ['ls-files', '-co', '--exclude-standard', '--', ...roots])
+    .split(/\r?\n/).filter(Boolean).sort()
+  if (names.length === 0) throw new Error(`source identity has no files under ${roots.join(', ')}`)
+  return hashFileList(cwd, names)
+}
+
+function hashFileList(cwd, names) {
+  const hash = createHash('sha256')
+  for (const name of names) {
+    hash.update(name.replaceAll('\\', '/') + '\0')
+    hash.update(readFileSync(join(cwd, name)))
+    hash.update('\0')
+  }
+  return { sha256: hash.digest('hex'), fileCount: names.length }
+}
+
+function hashDirectoryTree(root) {
+  if (!existsSync(root)) return hashFileList(root, [])
+  const names = []
+  const visit = (directory) => {
+    for (const entry of readdirSync(directory, { withFileTypes: true })) {
+      const absolute = join(directory, entry.name)
+      const name = relative(root, absolute).replaceAll('\\', '/')
+      if (entry.isSymbolicLink()) throw new Error(`source identity rejected symlink in plugin tree: ${name}`)
+      if (entry.isDirectory()) visit(absolute)
+      else if (entry.isFile()) names.push(name)
+    }
+  }
+  visit(root)
+  names.sort()
+  return hashFileList(root, names)
+}
+
+function pluginNamesUnder(root) {
+  if (!existsSync(root)) return []
+  return readdirSync(root, { withFileTypes: true })
+    .filter(entry => entry.isDirectory() && existsSync(join(root, entry.name, 'index.mjs')))
+    .map(entry => entry.name)
+    .sort()
+}
+
+function captureSourceIdentity() {
+  const appRoots = ['app', 'server', 'shared', 'mcp', 'cli', 'sdk', 'i18n', 'public', 'tsconfig.json', 'package.json', 'pnpm-lock.yaml', 'nuxt.config.ts', 'config.yml']
+  const simulatorRoots = ['src', 'package.json', 'package-lock.json']
+  const simulatorSourceFiles = gitText(SIM_SOURCE, ['ls-files', '-co', '--exclude-standard', '--', ...simulatorRoots])
+    .split(/\r?\n/).filter(Boolean).sort()
+  const executedSimulatorFiles = simulatorSourceFiles.filter(name => existsSync(join(SIM_ROOT, name)))
+  if (!executedSimulatorFiles.includes('src/server/index.ts') || !executedSimulatorFiles.includes('package.json')) {
+    throw new Error('executed simulator copy is missing required source files for provenance')
+  }
+  const projectPluginTree = hashDirectoryTree(join(REPO, '.AgentWorkShop', 'plugins'))
+  if (projectPluginTree.sha256 !== projectPluginTreeAtPrepare?.sha256) {
+    throw new Error('project plugin source changed after isolation was prepared')
+  }
+  const promptTree = hashDirectoryTree(ISOLATED_PROMPTS_DIR)
+  if (promptTree.sha256 !== promptTreeAtPrepare?.sha256) {
+    throw new Error('isolated prompt assets changed after isolation was prepared')
+  }
+  const piCliVersion = JSON.parse(readFileSync(PI_CLI_PACKAGE_PATH, 'utf8')).version
+  return {
+    recordedAt: new Date().toISOString(),
+    nodeVersion: process.version,
+    platform: process.platform,
+    architecture: process.arch,
+    app: {
+      version: VERSION,
+      gitHead: gitText(REPO, ['rev-parse', 'HEAD']),
+      sourceTree: hashGitFiles(REPO, appRoots),
+      workingTreeStatus: gitText(REPO, ['status', '--porcelain', '--', ...appRoots]),
+      packageManifestSha256: createHash('sha256').update(readFileSync(join(REPO, 'package.json'))).digest('hex'),
+      lockfileSha256: createHash('sha256').update(readFileSync(join(REPO, 'pnpm-lock.yaml'))).digest('hex'),
+      projectPluginTree,
+      projectPluginNames: pluginNamesUnder(join(REPO, '.AgentWorkShop', 'plugins')),
+    },
+    pluginIsolation: {
+      disabledNames: isolatedPluginNames,
+      enabledAllowlist: allowedEnabledPluginNames,
+      setupExpectedForOtherPlugins: false,
+      disabledStatePath: 'AW_HOME/plugins-state.json',
+      projectPluginNames: pluginNamesUnder(join(REPO, '.AgentWorkShop', 'plugins')),
+      projectPluginTree,
+    },
+    promptIsolation: {
+      sourceDirectory: 'repository/.AgentWorkShop/prompts',
+      isolatedDirectory: 'AW_HOME/prompts',
+      sourceTree: projectPromptTreeAtPrepare,
+      runtimeTree: promptTree,
+    },
+    experimentScriptSha256: createHash('sha256').update(readFileSync(fileURLToPath(import.meta.url))).digest('hex'),
+    registeredHarness: {
+      mode: HARNESS_MODE,
+      harness: 'pi',
+      command: PI_COMMAND,
+      piCliVersion,
+      provider: PI_PROVIDER,
+      model: PI_MODEL,
+      isolatedModelConfigSha256: createHash('sha256').update(readFileSync(PI_MODELS_PATH)).digest('hex'),
+      localModelFixtureSha256: createHash('sha256').update(readFileSync(fileURLToPath(import.meta.url))).digest('hex'),
+      externalCredentialsUsed: false,
+    },
+    simulator: {
+      gitlinkHead: gitText(REPO, ['rev-parse', 'HEAD:plc-node-simulator']),
+      gitHead: gitText(SIM_SOURCE, ['rev-parse', 'HEAD']),
+      sourceTree: hashGitFiles(SIM_SOURCE, simulatorRoots),
+      workingTreeStatus: gitText(SIM_SOURCE, ['status', '--porcelain', '--', ...simulatorRoots]),
+      executedTree: hashFileList(SIM_ROOT, executedSimulatorFiles),
+      lockfileSha256: createHash('sha256').update(readFileSync(join(SIM_SOURCE, 'package-lock.json'))).digest('hex'),
+      copyPatch: 'bind copied simulator HTTP and MQTT listeners to 127.0.0.1 and remap only their ports; source script hash records the transformation',
+    },
+  }
+}
+
+function prepareIsolatedHome() {
+  mkdirSync(join(HOME, 'data'), { recursive: true })
+  mkdirSync(join(PI_AGENT_DIR, 'sessions'), { recursive: true })
+  const piModels = {
+    providers: {
+      [PI_PROVIDER]: {
+        name: 'AgentWorkShop loopback deterministic fixture',
+        baseUrl: `http://127.0.0.1:${MODEL_API_PORT}/v1`,
+        apiKey: 'local-fixture-key-not-a-secret',
+        api: 'openai-completions',
+        models: [{
+          id: PI_MODEL,
+          name: 'AgentWorkShop deterministic paper fixture',
+          reasoning: false,
+          input: ['text'],
+          cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+          contextWindow: 64000,
+          maxTokens: 2048,
+        }],
+      },
+    },
+  }
+  writeFileSync(PI_MODELS_PATH, `${JSON.stringify(piModels, null, 2)}\n`, 'utf8')
+  assertLoopbackUrl(piModels.providers[PI_PROVIDER].baseUrl, MODEL_API_PORT)
+  noteIsolation('pi-custom-provider-isolated-to-local-fixture', true, {
+    configPath: 'AW_HOME/pi-agent/models.json',
+    provider: PI_PROVIDER,
+    model: PI_MODEL,
+    api: 'openai-completions',
+    baseUrl: piModels.providers[PI_PROVIDER].baseUrl,
+    apiKeyValueRecorded: false,
+    externalCredentialsUsed: false,
+  })
+  copyFileSync(join(REPO, 'config.yml'), join(HOME, 'config.yml'))
+  const overrides = {
+    'daq.startInfrastructure': 'never',
+    'daq.mqtt.host': '127.0.0.1',
+    'daq.mqtt.port': OFFLINE_APP_MQTT_QUEUE_PORT,
+    'daq.timescale.host': '127.0.0.1',
+    'daq.timescale.port': OFFLINE_TSDB_PORT,
+    'daq.objectstore.host': '127.0.0.1',
+    'daq.objectstore.port': OFFLINE_OBJECTSTORE_PORT,
+    'plugins.rag-bridge.base_url': RAG_BRIDGE_DISABLED_BASE_URL,
+  }
+  const runtimeSettingsPath = join(HOME, 'runtime-settings.json')
+  writeFileSync(runtimeSettingsPath, `${JSON.stringify({ version: 1, updatedAt: new Date().toISOString(), overrides }, null, 2)}\n`, 'utf8')
+  const savedOverrides = JSON.parse(readFileSync(runtimeSettingsPath, 'utf8')).overrides
+  if (savedOverrides?.['plugins.rag-bridge.base_url'] !== overrides['plugins.rag-bridge.base_url']) {
+    throw new Error('isolated runtime settings did not preserve the rag-bridge outbound-disable host guard setting')
+  }
+  const ragBridgeKvDir = join(HOME, 'data', 'plugins', 'rag-bridge')
+  mkdirSync(ragBridgeKvDir, { recursive: true })
+  const ragBridgeKvPath = join(ragBridgeKvDir, 'kv.json')
+  writeFileSync(ragBridgeKvPath, `${JSON.stringify({ 'kb.base_url': RAG_BRIDGE_DISABLED_BASE_URL }, null, 2)}\n`, 'utf8')
+  const savedRagBridgeKv = JSON.parse(readFileSync(ragBridgeKvPath, 'utf8'))
+  if (savedRagBridgeKv?.['kb.base_url'] !== RAG_BRIDGE_DISABLED_BASE_URL) {
+    throw new Error('isolated rag-bridge KV fallback did not preserve the outbound-disable host guard setting')
+  }
+  const projectPromptsRoot = join(REPO, '.AgentWorkShop', 'prompts')
+  if (!existsSync(projectPromptsRoot)) throw new Error('project prompt source directory is missing')
+  cpSync(projectPromptsRoot, ISOLATED_PROMPTS_DIR, { recursive: true })
+  projectPromptTreeAtPrepare = hashDirectoryTree(projectPromptsRoot)
+  promptTreeAtPrepare = hashDirectoryTree(ISOLATED_PROMPTS_DIR)
+  if (projectPromptTreeAtPrepare.sha256 !== promptTreeAtPrepare.sha256) {
+    throw new Error('isolated prompt copy differs from the repository prompt source')
+  }
+  writeFileSync(EMPTY_DOTENV, '', 'utf8')
+  const builtinPlugins = pluginNamesUnder(join(REPO, 'server', 'plugins-builtin'))
+  const projectPluginsRoot = join(REPO, '.AgentWorkShop', 'plugins')
+  const projectPlugins = pluginNamesUnder(projectPluginsRoot)
+  const userPlugins = pluginNamesUnder(join(HOME, 'plugins'))
+  expectedPluginNames = [...new Set([...builtinPlugins, ...projectPlugins, ...userPlugins])].sort()
+  isolatedPluginNames = expectedPluginNames.filter(name => !allowedEnabledPluginNames.includes(name))
+  projectPluginTreeAtPrepare = hashDirectoryTree(projectPluginsRoot)
+  const pluginStatePath = join(HOME, 'plugins-state.json')
+  writeFileSync(pluginStatePath, `${JSON.stringify({ version: 1, updatedAt: new Date().toISOString(), disabled: isolatedPluginNames }, null, 2)}\n`, 'utf8')
+  const savedPluginState = JSON.parse(readFileSync(pluginStatePath, 'utf8'))
+  if (JSON.stringify([...savedPluginState.disabled].sort()) !== JSON.stringify(isolatedPluginNames)) {
+    throw new Error('isolated plugin state did not disable every discovered plugin')
+  }
+  noteIsolation('isolated-runtime-config', true, {
+    configRoot: 'fresh isolated AW_HOME',
+    startInfrastructure: overrides['daq.startInfrastructure'],
+    simulatorMqttBroker: `127.0.0.1:${MQTT_PORT}`,
+    appMqttQueueProbe: `127.0.0.1:${OFFLINE_APP_MQTT_QUEUE_PORT}`,
+    offlineTimescaleProbe: `127.0.0.1:${OFFLINE_TSDB_PORT}`,
+    offlineObjectstoreProbe: `127.0.0.1:${OFFLINE_OBJECTSTORE_PORT}`,
+    ragBridgeOutboundGuardSetting: 'plugins.rag-bridge.base_url',
+    ragBridgeOutboundGuardConfigPresent: true,
+    ragBridgeOutboundGuardKvFallbackPresent: true,
+    emptyDotenv: true,
+    repoDotenvBypassed: true,
+  })
+  noteIsolation('pre-existing-plugin-scopes-disabled', true, {
+    builtinNames: builtinPlugins,
+    projectNames: projectPlugins,
+    userNames: userPlugins,
+    disabledStatePath: 'AW_HOME/plugins-state.json',
+    projectPluginTreeSha256: projectPluginTreeAtPrepare.sha256,
+    enabledPluginAllowlist: allowedEnabledPluginNames,
+  })
+  noteIsolation('project-prompt-assets-isolated', true, {
+    sourceDirectory: 'repository/.AgentWorkShop/prompts',
+    isolatedDirectory: 'AW_HOME/prompts',
+    sourceTreeSha256: projectPromptTreeAtPrepare.sha256,
+    runtimeTreeSha256: promptTreeAtPrepare.sha256,
+    awPromptsDir: 'AW_HOME/prompts',
+  })
+}
+
+async function assertPluginIsolation() {
+  const expectedNames = expectedPluginNames
+  let plugins = []
+  for (let attempt = 0; attempt < 40; attempt++) {
+    const result = requireSuccess(await app('GET', '/api/workshop/plugins'), 'isolated plugin manifest')
+    plugins = Array.isArray(result.plugins) ? result.plugins : []
+    const actualNames = plugins.map(plugin => plugin.name).sort()
+    if (JSON.stringify(actualNames) === JSON.stringify(expectedNames)) {
+      const enabledNames = plugins.filter(plugin => plugin.enabled === true).map(plugin => plugin.name).sort()
+      const enabledMatchesAllowlist = JSON.stringify(enabledNames) === JSON.stringify([...allowedEnabledPluginNames].sort())
+      const statesMatchPolicy = plugins.every(plugin => allowedEnabledPluginNames.includes(plugin.name)
+        ? plugin.enabled === true
+        : plugin.enabled === false)
+      check('only-outbound-disabled-rag-bridge-enabled', enabledMatchesAllowlist && statesMatchPolicy, {
+        names: actualNames,
+        enabledNames,
+        disabledNames: plugins.filter(plugin => plugin.enabled === false).map(plugin => plugin.name).sort(),
+        invalidStateNames: plugins.filter(plugin => typeof plugin.enabled !== 'boolean').map(plugin => plugin.name),
+      })
+      return
+    }
+    await delay(250)
+  }
+  throw new Error(`isolated plugin manifest differed from the expected built-in/user set: ${plugins.map(plugin => plugin.name).sort().join(', ')}`)
+}
+
+function spawnLogged(command, args, { cwd = REPO, env = cleanChildEnv(), logFd, label }) {
+  const child = spawn(command, args, {
+    cwd,
+    env,
+    stdio: ['ignore', logFd, logFd],
+    windowsHide: true,
+    detached: process.platform !== 'win32',
+  })
+  child.label = label
+  children.add(child)
+  child.once('exit', () => children.delete(child))
+  child.once('error', () => children.delete(child))
+  return child
+}
+
+async function waitForExit(child, timeoutMs = 8_000) {
+  if (!child || child.exitCode !== null || child.signalCode !== null) return
+  await Promise.race([
+    new Promise(resolvePromise => child.once('exit', resolvePromise)),
+    delay(timeoutMs),
+  ])
+}
+
+function processGroupExists(pid) {
+  try {
+    process.kill(-pid, 0)
+    return true
+  }
+  catch (error) {
+    if (error?.code === 'ESRCH') return false
+    if (error?.code === 'EPERM') return true
+    throw error
+  }
+}
+
+function windowsProcessExists(pid) {
+  const result = spawnSync('tasklist.exe', ['/FI', `PID eq ${pid}`, '/FO', 'CSV', '/NH'], {
+    windowsHide: true,
+    encoding: 'utf8',
+    stdio: ['ignore', 'pipe', 'ignore'],
+    env: cleanChildEnv(),
+  })
+  if (result.error || result.status !== 0) return null
+  return (result.stdout ?? '').split(/\r?\n/).some((line) => {
+    const columns = line.match(/^"(?:[^"]|"")*","(\d+)"(?:,|$)/)
+    return columns?.[1] === String(pid)
+  })
+}
+
+async function stopProcessTree(child, label) {
+  if (!child?.pid) return false
+  const wasAlreadyExited = child.exitCode !== null || child.signalCode !== null
+  let terminationStatus = null
+  try {
+    if (process.platform === 'win32' && child.pid) {
+      const result = spawnSync('taskkill.exe', ['/PID', String(child.pid), '/T', '/F'], { windowsHide: true, stdio: 'ignore', env: cleanChildEnv() })
+      terminationStatus = result.error ? 'spawn-error' : result.status
+    }
+    else if (process.platform !== 'win32' && child.pid) {
+      process.kill(-child.pid, 'SIGTERM')
+    }
+  }
+  catch { /* a process may have already exited */ }
+  await waitForExit(child)
+  if (process.platform !== 'win32' && processGroupExists(child.pid)) {
+    try {
+      process.kill(-child.pid, 'SIGKILL')
+      terminationStatus = 'SIGKILL'
+    }
+    catch { /* already gone */ }
+  }
+  else if (process.platform === 'win32' && windowsProcessExists(child.pid) === true) {
+    const result = spawnSync('taskkill.exe', ['/PID', String(child.pid), '/T', '/F'], { windowsHide: true, stdio: 'ignore', env: cleanChildEnv() })
+    terminationStatus = result.error ? 'spawn-error' : result.status
+  }
+  await waitForExit(child, 2_000)
+  let treePresent
+  if (process.platform === 'win32') treePresent = windowsProcessExists(child.pid)
+  else {
+    const deadline = Date.now() + 3_000
+    do {
+      treePresent = processGroupExists(child.pid)
+      if (!treePresent) break
+      await delay(100)
+    } while (Date.now() < deadline)
+  }
+  const pass = !wasAlreadyExited && treePresent === false && (process.platform !== 'win32' || terminationStatus === 0)
+  evidence.checks.push({
+    id: `${label}-process-stopped`,
+    status: pass ? 'pass' : 'fail',
+    detail: { pid: child.pid, terminationStatus, processTreePresentAfterCleanup: treePresent, rootWasAlreadyExited: wasAlreadyExited },
+  })
+  return pass
+}
+
+function buildHarnessShim() {
+  let wrapperPath
+  if (process.platform === 'win32') {
+    const compiler = [
+      join(process.env.WINDIR || 'C:\\Windows', 'Microsoft.NET', 'Framework64', 'v4.0.30319', 'csc.exe'),
+      join(process.env.WINDIR || 'C:\\Windows', 'Microsoft.NET', 'Framework', 'v4.0.30319', 'csc.exe'),
+    ].find(existsSync)
+    if (!compiler) throw new Error('Windows C# compiler unavailable; cannot safely launch a registered fixture executable')
+    const source = join(ROOT, 'omp-fixture-shim.cs')
+    wrapperPath = join(ROOT, 'omp-fixture-shim.exe')
+    writeFileSync(source, String.raw`using System;
+using System.Diagnostics;
+using System.IO;
+using System.Linq;
+using System.Text;
+using System.Threading;
+class OmpFixtureShim {
+  static string Quote(string value) {
+    var result = new StringBuilder("\"");
+    int backslashes = 0;
+    foreach (char c in value) {
+      if (c == '\\') { backslashes++; continue; }
+      if (c == '"') {
+        result.Append('\\', backslashes * 2 + 1);
+        result.Append('"');
+      }
+      else {
+        result.Append('\\', backslashes);
+        result.Append(c);
+      }
+      backslashes = 0;
+    }
+    result.Append('\\', backslashes * 2);
+    result.Append('"');
+    return result.ToString();
+  }
+  static void Copy(Stream source, Stream destination) {
+    byte[] buffer = new byte[4096];
+    int count;
+    while ((count = source.Read(buffer, 0, buffer.Length)) > 0) {
+      destination.Write(buffer, 0, count);
+      destination.Flush();
+    }
+  }
+  static int Main(string[] args) {
+    string node = Environment.GetEnvironmentVariable("AW_PAPER_FIXTURE_NODE");
+    string script = Environment.GetEnvironmentVariable("AW_PAPER_FIXTURE_SCRIPT");
+    if (String.IsNullOrWhiteSpace(node) || String.IsNullOrWhiteSpace(script)) return 125;
+    var psi = new ProcessStartInfo(node);
+    psi.UseShellExecute = false;
+    psi.CreateNoWindow = true;
+    psi.RedirectStandardInput = true;
+    psi.RedirectStandardOutput = true;
+    psi.RedirectStandardError = true;
+    psi.Arguments = Quote(script) + " " + String.Join(" ", args.Skip(2).Select(Quote));
+    using (var child = Process.Start(psi)) {
+      if (child == null) return 126;
+      var stdin = new Thread(() => {
+        try {
+          Copy(Console.OpenStandardInput(), child.StandardInput.BaseStream);
+        }
+        finally {
+          child.StandardInput.Close();
+        }
+      });
+      var stdout = new Thread(() => Copy(child.StandardOutput.BaseStream, Console.OpenStandardOutput()));
+      var stderr = new Thread(() => Copy(child.StandardError.BaseStream, Console.OpenStandardError()));
+      stdin.IsBackground = true;
+      stdout.IsBackground = true;
+      stderr.IsBackground = true;
+      stdin.Start();
+      stdout.Start();
+      stderr.Start();
+      child.WaitForExit();
+      stdout.Join();
+      stderr.Join();
+      return child.ExitCode;
+    }
+  }
+}`.replace(/\n/g, '\n'), 'utf8')
+    const result = spawnSync(compiler, ['/nologo', '/target:exe', `/out:${wrapperPath}`, source], { cwd: ROOT, windowsHide: true, encoding: 'utf8', timeout: 30_000, env: cleanChildEnv() })
+    if (result.error || result.status !== 0 || !existsSync(wrapperPath)) {
+      throw new Error(`could not compile isolated OMP fixture shim: ${safeText(result.stderr || result.stdout || result.error?.message || result.status)}`)
+    }
+    return wrapperPath
+  }
+  if (process.platform === 'linux' || process.platform === 'darwin') {
+    wrapperPath = join(ROOT, 'omp-fixture-shim')
+    writeFileSync(wrapperPath, '#!/bin/sh\nshift 2\nexec "$AW_PAPER_FIXTURE_NODE" "$AW_PAPER_FIXTURE_SCRIPT" "$@"\n', 'utf8')
+    chmodSync(wrapperPath, 0o755)
+    return wrapperPath
+  }
+  throw new Error(`unsupported fixture shim platform ${process.platform}`)
+}
+
+function extractPayload(body) {
+  return body && typeof body === 'object' && Object.prototype.hasOwnProperty.call(body, 'data') ? body.data : body
+}
+
+async function request(base, expectedPort, method, path, body, { token, timeoutMs = 10_000 } = {}) {
+  assertLoopbackUrl(base, expectedPort)
+  if (!path.startsWith('/api/')) throw new Error('API path boundary rejected')
+  const response = await fetch(`${base}${path}`, {
+    method,
+    headers: {
+      ...(body === undefined ? {} : { 'content-type': 'application/json' }),
+      ...(token ? { authorization: `Bearer ${token}` } : {}),
+    },
+    ...(body === undefined ? {} : { body: JSON.stringify(body) }),
+    signal: AbortSignal.timeout(timeoutMs),
+  })
+  const parsed = await response.json().catch(() => null)
+  return { status: response.status, body: parsed, data: extractPayload(parsed) }
+}
+
+const app = (method, path, body, opts = {}) => request(APP_BASE, AW_PORT, method, path, body, { token: opts.token ?? authToken, timeoutMs: opts.timeoutMs })
+const sim = (method, path, body) => request(SIM_BASE, SIM_PORT, method, path, body, { timeoutMs: 8_000 })
+
+function requireSuccess(result, label, status = 200) {
+  if (result.status !== status) {
+    const payload = result.body && typeof result.body === 'object' ? result.body : {}
+    throw new Error(`${label} returned HTTP ${result.status}: ${safeText(payload.message || payload.statusMessage || payload.code || 'no response detail')}`)
+  }
+  return result.data
+}
+
+async function bootSimulator() {
+  mkdirSync(SIM_ROOT, { recursive: true })
+  cpSync(join(SIM_SOURCE, 'src'), join(SIM_ROOT, 'src'), { recursive: true })
+  cpSync(join(SIM_SOURCE, 'package.json'), join(SIM_ROOT, 'package.json'))
+  const simulatorEntry = join(SIM_ROOT, 'src', 'server', 'index.ts')
+  const originalSource = readFileSync(simulatorEntry, 'utf8')
+  const listenCall = 'server.listen(PORT, () => resolve())'
+  const brokerCall = 'await startBroker(Number(process.env.MQTT_BROKER_PORT ?? 18830))'
+  if (originalSource.split(listenCall).length !== 2) throw new Error('simulator copied source did not contain exactly one expected listener call')
+  if (originalSource.split(brokerCall).length !== 2) throw new Error('simulator copied source did not contain exactly one expected broker startup call')
+  const loopbackListenCall = `server.listen(PORT, '127.0.0.1', () => {
+    const boundAddress = server.address()
+    if (!boundAddress || typeof boundAddress === 'string' || boundAddress.address !== '127.0.0.1' || boundAddress.port !== PORT) {
+      console.error('[paper-isolation] simulator listener address verification failed')
+      server.close(() => process.exit(70))
+      return
+    }
+    console.log('[paper-isolation] listener=' + boundAddress.address + ':' + boundAddress.port)
+    resolve()
+  })`
+  const loopbackSource = originalSource
+    .replace(listenCall, loopbackListenCall)
+    .replace(brokerCall, `await startBroker(Number(process.env.MQTT_BROKER_PORT ?? 18830), '127.0.0.1')`)
+  writeFileSync(simulatorEntry, loopbackSource, 'utf8')
+  const modules = join(SIM_SOURCE, 'node_modules')
+  symlinkSync(modules, join(SIM_ROOT, 'node_modules'), process.platform === 'win32' ? 'junction' : 'dir')
+  simProcess = spawnLogged(process.execPath, [join(REPO, 'node_modules', 'tsx', 'dist', 'cli.mjs'), join(SIM_ROOT, 'src', 'server', 'index.ts')], {
+    cwd: SIM_ROOT,
+    logFd: SIM_LOG_FD,
+    label: 'simulator',
+    env: cleanChildEnv({ SIM_PORT: String(SIM_PORT), MQTT_BROKER_PORT: String(MQTT_PORT) }),
+  })
+  for (let i = 0; i < 50; i++) {
+    if (simProcess.exitCode !== null) throw new Error(`simulator exited during startup (code ${simProcess.exitCode})`)
+    try {
+      const result = await sim('GET', '/api/nodes')
+      if (result.status === 200) {
+        const data = result.data
+        const expectedListener = `[paper-isolation] listener=127.0.0.1:${SIM_PORT}`
+        const simulatorLog = readFileSync(SIM_LOG, 'utf8')
+        const expectedBroker = `[mqtt-broker] 内置 broker 就绪 mqtt://127.0.0.1:${MQTT_PORT}`
+        if (!simulatorLog.includes(expectedListener)) throw new Error('simulator did not report the verified 127.0.0.1 HTTP listener address')
+        if (!simulatorLog.includes(expectedBroker) || simulatorLog.includes('内置 broker 启动失败')) {
+          throw new Error('simulator did not confirm its embedded loopback MQTT broker before application startup')
+        }
+        noteIsolation('simulator-listener-address-verified', true, { address: '127.0.0.1', port: SIM_PORT, copiedSourceOnly: true })
+        noteIsolation('simulator-mqtt-broker-loopback-ready', true, { address: '127.0.0.1', port: MQTT_PORT, brokerStartupConfirmed: true })
+        noteIsolation('simulator-local-http-ready', true, { version: 'local repository source', port: SIM_PORT, host: '127.0.0.1' })
+        return data
+      }
+    }
+    catch { /* the local simulator is still starting */ }
+    await delay(400)
+  }
+  throw new Error('isolated simulator startup timed out')
+}
+
+function normalizeLocalNetworkConfig(value, portMap, key = '') {
+  if (Array.isArray(value)) return value.map(item => normalizeLocalNetworkConfig(item, portMap, key))
+  if (!value || typeof value !== 'object') {
+    if (typeof value !== 'string') return value
+    if (/^[a-z][a-z\d+.-]*:\/\//i.test(value)) {
+      const parsed = new URL(value)
+      if (!['127.0.0.1', 'localhost', '::1'].includes(parsed.hostname)) throw new Error('simulator preset contains a non-loopback endpoint')
+      if (parsed.port && portMap.has(Number(parsed.port))) parsed.port = String(portMap.get(Number(parsed.port)))
+      parsed.hostname = '127.0.0.1'
+      return parsed.toString()
+    }
+    if (/(?:host|hostname|ip|address)$/i.test(key)) return '127.0.0.1'
+    return value
+  }
+  const out = {}
+  for (const [childKey, childValue] of Object.entries(value)) {
+    if (/^port$/i.test(childKey) && typeof childValue === 'number' && portMap.has(childValue)) {
+      out[childKey] = portMap.get(childValue)
+    }
+    else if (/(?:host|hostname|ip|address)$/i.test(childKey) && typeof childValue === 'string') {
+      out[childKey] = '127.0.0.1'
+    }
+    else {
+      out[childKey] = normalizeLocalNetworkConfig(childValue, portMap, childKey)
+    }
+  }
+  return out
+}
+
+async function simulatorScene(scene) {
+  const presetResponse = await sim('GET', `/api/presets/${encodeURIComponent(scene.preset)}`)
+  const blueprint = requireSuccess(presetResponse, `${scene.id} simulator preset`)
+  if (!blueprint?.plantModel || !Array.isArray(blueprint.nodes)) throw new Error(`${scene.id} simulator preset shape is invalid`)
+
+  const portMapByDeviceId = new Map()
+  const toCreate = []
+  let nextPort = 44_000 + evidence.experiment.scenes.length * 100
+  for (const raw of blueprint.nodes) {
+    const node = structuredClone(raw)
+    if ((node.protocol && /serial|rtu/i.test(String(node.protocol))) || /serialPort|comPort|devicePath/i.test(JSON.stringify(node.config ?? {}))) continue
+    node.config ||= {}
+    const oldPort = Number(node.config.port)
+    const portMap = new Map()
+    if (Number.isInteger(oldPort) && oldPort > 0) {
+      while (allocatedDevicePorts.includes(nextPort) || nextPort === FORBIDDEN_PORT) nextPort += 1
+      await assertFreeLoopback(nextPort, `${scene.id}-simulator-device`)
+      portMap.set(oldPort, nextPort)
+      allocatedDevicePorts.push(nextPort)
+      noteIsolation(`${scene.id}-device-port-free-loopback`, true, { port: nextPort, host: '127.0.0.1', protocol: node.protocol ?? 'unspecified' })
+      node.config.port = nextPort
+      nextPort += 1
+    }
+    node.config = normalizeLocalNetworkConfig(node.config, portMap)
+    if (node.protocol === 'mqtt' || /mqtt/i.test(String(node.protocol))) node.config.brokerUrl = `mqtt://127.0.0.1:${MQTT_PORT}`
+    portMapByDeviceId.set(node.id, portMap)
+    toCreate.push(node)
+  }
+  if (toCreate.length === 0) throw new Error(`${scene.id} simulator preset has no loopback-safe device nodes`)
+
+  for (const node of toCreate) {
+    const created = await sim('POST', '/api/nodes', node)
+    requireSuccess(created, `${scene.id} simulator device ${safeText(node.id, 80)}`)
+  }
+  const plant = { ...blueprint.plantModel, id: `${TAG}-${scene.id}`, seed: 42, timeScale: 1, truthExport: false }
+  const loaded = await sim('PUT', '/api/plant/config', { plantModel: plant, warm: true, upsert: true })
+  requireSuccess(loaded, `${scene.id} simulator plant model`)
+
+  const exports = []
+  for (const node of toCreate) {
+    const result = await sim('GET', `/api/nodes/${encodeURIComponent(node.id)}/export`)
+    const data = requireSuccess(result, `${scene.id} simulator export ${safeText(node.id, 80)}`)
+    for (const item of data?.items ?? []) exports.push({ ...item, deviceId: node.id })
+  }
+  for (const item of exports) {
+    item.driverConfig = normalizeLocalNetworkConfig(item.driverConfig ?? {}, portMapByDeviceId.get(item.deviceId) ?? new Map())
+    for (const [key, value] of Object.entries(item.driverConfig)) {
+      if (/^(?:host|hostname|ip|address)$/i.test(key) && !['127.0.0.1', 'localhost', '::1'].includes(String(value))) {
+        throw new Error(`${scene.id} exported driver config escaped loopback boundary`)
+      }
+    }
+  }
+  const mv = exports.find(item => item.deviceId === scene.sourceDevice && item.signal === scene.mvSignal)
+  const cv = exports.find(item => item.signal === scene.cvSignal)
+  if (!mv || !cv) throw new Error(`${scene.id} simulator export lacks required MV/CV signals`)
+  if (!['opcua', 'modbus-tcp'].includes(String(mv.driver))) throw new Error(`${scene.id} MV driver is outside the approved simulator protocol set`)
+  const evidenceSignal = signal => Object.fromEntries(Object.entries(signal).filter(([key]) => key !== 'curl'))
+  return {
+    mv: evidenceSignal(mv),
+    cv: evidenceSignal(cv),
+    plant: { id: plant.id, kind: plant.kind ?? 'simulator', seed: 42, timeScale: 1 },
+    deviceCount: toCreate.length,
+  }
+}
+
+async function registerAdmin() {
+  const response = await app('POST', '/api/users/register', {
+    name: `Paper Test ${TAG}`,
+    email: `${TAG}@example.invalid`,
+    password: `Paper-${Date.now()}-safe9`,
+  }, { token: undefined })
+  const data = requireSuccess(response, 'isolated synthetic administrator registration')
+  if (!data?.token || data?.user?.role !== 'admin') throw new Error('fresh-home registration did not return its first administrator token')
+  authToken = String(data.token)
+  syntheticAdminId = String(data.user.id)
+  return { id: syntheticAdminId, role: 'admin' }
+}
+
+async function verifyRagBridgeOutboundDisabled() {
+  const ragBridgeHealth = requireSuccess(await app('GET', '/api/plugins/rag-bridge/health'), 'rag-bridge outbound guard health check')
+  const deadline = Date.now() + 3_000
+  let startupLogConfirmed = false
+  while (Date.now() < deadline) {
+    startupLogConfirmed = /\[rag-bridge\].*出站调用已禁用:/.test(readFileSync(APP_LOG, 'utf8'))
+    if (startupLogConfirmed) break
+    await delay(100)
+  }
+  noteIsolation('rag-bridge-outbound-disabled', ragBridgeHealth?.outbound === false && startupLogConfirmed, {
+    setting: 'plugins.rag-bridge.base_url',
+    pluginHealthOutbound: ragBridgeHealth?.outbound ?? null,
+    startupLogConfirmed,
+  })
+}
+
+function piFixtureSystemPrompt(scene, recipe, control, decisionStimulus) {
+  const selected = Number(scene.fixtureCandidate)
+  return [
+    'You are operating AgentWorkShop for a bounded integration test in a local process simulator.',
+    `Process scene: ${scene.processDescription}.`,
+    `Before task submission, the registered drivers measured the control at ${decisionStimulus.controlValue} ${control.unit} and the process variable at ${decisionStimulus.processValue} ${decisionStimulus.processUnit}.`,
+    `The task target band is ${decisionStimulus.targetMin} to ${decisionStimulus.targetMax} ${decisionStimulus.processUnit}.`,
+    `This local deterministic model fixture is configured to test one bounded candidate: ${selected} ${control.unit}. It is the only permitted value for this request.`,
+    `Call recipe_trial exactly once with recipe_id "${recipe.id}", one params entry for node_id "${control.id}" and value ${selected}. State only that this is a deterministic integration fixture in hypothesis.`,
+    'After receiving the recipe_trial result, call complete_task exactly once on the active root task. Include the selected value and returned command-verification outcome in a concise summary and deliverable.',
+    'Do not make any other tool calls. Do not retry. Do not create, update, apply, or roll back a recipe. Do not claim that the trial improves or settles the process.',
+    'After the complete_task result, give a brief factual report and make no further tool calls.',
+  ].join('\n')
+}
+
+async function setupScene(scene, simDetails) {
+  currentScene = scene.id
+  currentStage = `${scene.id}:application-setup`
+  const sceneEvidence = {
+    id: scene.id,
+    simulatorPreset: scene.preset,
+    simulator: simDetails,
+    status: 'running',
+    approvals: [],
+    runs: [],
+  }
+  evidence.experiment.scenes.push(sceneEvidence)
+
+  const line = requireSuccess(await app('POST', '/api/workshop/dcw/lines', {
+    name: `${TAG}-${scene.id}`,
+    description: 'Synthetic isolated paper integration run',
+    simulated: true,
+    simSceneId: scene.preset,
+    controlMode: 'manual',
+  }), `${scene.id} line creation`).line
+  const product = requireSuccess(await app('POST', '/api/workshop/dcw/products', {
+    name: `${TAG}-${scene.id}-product`, lineId: line.id,
+  }), `${scene.id} product creation`).product
+  const template = requireSuccess(await app('POST', '/api/workshop/dcw/templates', {
+    name: `${TAG}-${scene.id}-control-template`,
+    unit: simDetails.mv.unit,
+    min: simDetails.mv.min,
+    max: simDetails.mv.max,
+    decimals: simDetails.mv.decimals ?? 0,
+  }), `${scene.id} control template creation`).template
+  const control = requireSuccess(await app('POST', '/api/workshop/dcw', {
+    templateRef: template.key,
+    name: `${TAG}-${scene.id}-control`,
+    driver: simDetails.mv.driver,
+    driverConfig: simDetails.mv.driverConfig,
+    lineId: line.id,
+    min: simDetails.mv.min,
+    max: simDetails.mv.max,
+    unit: simDetails.mv.unit,
+    decimals: simDetails.mv.decimals ?? 0,
+    stepLimit: 30,
+    writeLockSeconds: 0,
+    holdIntervalMs: 0,
+    readIntervalMs: 500,
+    semantics: `Bounded simulated ${scene.mvSignal} setpoint for integration evidence`,
+  }), `${scene.id} control node creation`).node
+  const daq = requireSuccess(await app('POST', '/api/workshop/daq', {
+    templateRef: 'daq-temp-tc',
+    name: `${TAG}-${scene.id}-daq-cv`,
+    driver: simDetails.cv.driver,
+    driverConfig: simDetails.cv.driverConfig,
+    lineId: line.id,
+    min: simDetails.cv.min,
+    max: simDetails.cv.max,
+    unit: simDetails.cv.unit,
+    intervalMs: 1000,
+    publishIntervalMs: 0,
+    semantics: `Bounded simulated ${scene.cvSignal} controlled variable for run-tagged sampling`,
+  }), `${scene.id} DAQ node creation`).node
+  const recipe = requireSuccess(await app('POST', '/api/workshop/dcw/recipes', {
+    name: `${TAG}-${scene.id}-recipe`,
+    productId: product.id,
+    params: [{ nodeId: control.id, templateRef: template.key, value: scene.baseline, min: simDetails.mv.min, max: simDetails.mv.max, stepLimit: 30 }],
+    daqWindows: [{ nodeId: daq.id, min: simDetails.cv.min, max: simDetails.cv.max }],
+  }), `${scene.id} recipe creation`).recipe
+
+  const start = requireSuccess(await app('POST', `/api/workshop/dcw/lines/${encodeURIComponent(line.id)}/start`, { recipeId: recipe.id }), `${scene.id} line start`)
+  const runId = String(start.run?.id ?? '')
+  if (!runId) throw new Error(`${scene.id} line start returned no run identifier`)
+  sceneEvidence.activeRun = { id: runId, status: start.run?.status ?? null, startedAt: start.run?.startedAt ?? null, recipeId: start.run?.recipeId ?? recipe.id }
+
+  let decisionStimulus
+  if (HARNESS_MODE === 'pi-local-fixture') {
+    let controlRead
+    let controlReadAttempts = 0
+    do {
+      controlRead = await app('POST', `/api/workshop/dcw/${encodeURIComponent(control.id)}/read`)
+      controlReadAttempts += 1
+      if (controlRead.status === 409 && controlReadAttempts < 3) await delay(500)
+    } while (controlRead.status === 409 && controlReadAttempts < 3)
+    const controlReadData = requireSuccess(controlRead, `${scene.id} pre-task control read`)
+    const controlData = controlReadData?.outcome ?? controlReadData?.read ?? controlReadData
+    const controlValue = Number(controlData?.value ?? controlData?.data?.value)
+    let processProbeResponse
+    let processProbeAttempts = 0
+    do {
+      processProbeResponse = await app('POST', `/api/workshop/daq/${encodeURIComponent(daq.id)}/test`)
+      processProbeAttempts += 1
+      if (processProbeResponse.status === 409 && processProbeAttempts < 3) await delay(500)
+    } while (processProbeResponse.status === 409 && processProbeAttempts < 3)
+    const processProbe = requireSuccess(processProbeResponse, `${scene.id} pre-task process probe`).test
+    const processValue = Number(processProbe?.sampleValue)
+    if (!Number.isFinite(controlValue) || processProbe?.ok !== true || !Number.isFinite(processValue)) {
+      throw new Error(`${scene.id} pre-task protocol reads did not provide finite control and process values`)
+    }
+    decisionStimulus = {
+      controlValue,
+      processValue,
+      processUnit: simDetails.cv.unit,
+      targetMin: scene.targetMin,
+      targetMax: scene.targetMax,
+      allowedValues: [...scene.candidates],
+      probeDriver: simDetails.cv.driver,
+      probeLatencyMs: Number.isFinite(Number(processProbe.latencyMs)) ? Number(processProbe.latencyMs) : null,
+      controlReadAttempts,
+      processProbeAttempts,
+      measuredAt: new Date().toISOString(),
+    }
+    sceneEvidence.decisionStimulus = decisionStimulus
+    sceneEvidence.simulator.candidateChoices = [...scene.candidates]
+  }
+
+  const channel = requireSuccess(await app('POST', '/api/workshop/channels', {
+    name: `${TAG}-${scene.id}-harness-channel`,
+    description: 'Registered Pi CLI with a local deterministic model API fixture for simulator integration evidence',
+  }), `${scene.id} channel creation`)
+  const agentConfig = {
+    command: PI_COMMAND,
+    cwd: HOME,
+    baseUrl: APP_BASE,
+    provider: PI_PROVIDER,
+    model: PI_MODEL,
+    promptTimeoutMs: PI_PROMPT_TIMEOUT_MS,
+    systemPromptPrefix: piFixtureSystemPrompt(scene, recipe, control, decisionStimulus),
+  }
+  const agent = requireSuccess(await app('POST', `/api/workshop/channels/${encodeURIComponent(channel.channelId)}/agents`, {
+    name: `${TAG}-${scene.id}-pi-local-fixture`,
+    harness: 'pi',
+    role: 'lead',
+    config: agentConfig,
+  }), `${scene.id} registered OMP lead creation`)
+
+  const recipeBinding = requireSuccess(await app('POST', '/api/workshop/agent-tools/bindings', {
+    agentId: agent.id, nodeId: recipe.id, kind: 'recipe',
+  }), `${scene.id} recipe binding`).binding
+  const daqBinding = requireSuccess(await app('POST', '/api/workshop/agent-tools/bindings', {
+    agentId: agent.id, nodeId: daq.id, kind: 'daq', mode: 'auto',
+  }), `${scene.id} DAQ binding`).binding
+  check(`${scene.id}-registered-pi`, agent.harness === 'pi' && Boolean(agent.id), { harness: agent.harness, agentId: agent.id, provider: PI_PROVIDER, model: PI_MODEL })
+  check(`${scene.id}-manual-recipe-binding`, recipeBinding.kind === 'recipe' && recipeBinding.mode === 'manual', { kind: recipeBinding.kind, mode: recipeBinding.mode })
+  check(`${scene.id}-daq-observation-binding`, daqBinding.kind === 'daq', { kind: daqBinding.kind })
+
+  sceneEvidence.entities = { lineId: line.id, productId: product.id, recipeId: recipe.id, controlNodeId: control.id, daqNodeId: daq.id, channelId: channel.channelId, agentId: agent.id }
+  sceneEvidence.bindings = { recipe: { kind: recipeBinding.kind, mode: recipeBinding.mode }, daq: { kind: daqBinding.kind, mode: daqBinding.mode } }
+  sceneEvidence.routePolicy = { tool: 'recipe_trial', emergency: false, cadenceBypass: false, recipeBindingMode: recipeBinding.mode }
+  sceneEvidence.status = 'setup-complete'
+  return { line, product, template, control, daq, recipe, channel, agent, runId, evidence: sceneEvidence }
+}
+
+async function submitFixtureTask(context) {
+  currentStage = `${context.evidence.id}:submit-lead-task`
+  const sent = requireSuccess(await app('POST', `/api/workshop/channels/${encodeURIComponent(context.channel.channelId)}/tasks`, {
+    title: `Isolated ${context.evidence.id} recipe_trial verification`,
+    description: [
+      `Execute the preconfigured isolated ${context.evidence.id} recipe_trial candidate exactly once and wait for synthetic human approval.`,
+      'After the verified recipe_trial result returns, call complete_task exactly once with a concise factual deliverable. Do not claim process improvement.',
+      'Do not make any other host-tool calls, retry, or create/update/apply/roll back a recipe.',
+    ].join(' '),
+    fromLabel: 'Synthetic isolated test driver',
+  }), `${context.evidence.id} lead task submission`)
+  context.evidence.taskId = sent.id ?? sent.task?.id ?? null
+  if (piFixtureTarget) piFixtureTarget.taskId = context.evidence.taskId
+}
+
+async function submitRealOmpTask(context) {
+  currentStage = `${context.evidence.id}:submit-real-omp-task`
+  const sent = requireSuccess(await app('POST', `/api/workshop/channels/${encodeURIComponent(context.channel.channelId)}/tasks`, {
+    title: `R20 ${context.evidence.id} recipe_trial with explicit task closure`,
+    description: [
+      'Use the measured process context in your registered system prompt to choose one candidate from its prevalidated set.',
+      `Call recipe_trial exactly once with recipe_id ${context.recipe.id} and one params entry for node_id ${context.control.id} using your selected value.`,
+      'After receiving the recipe_trial result, call complete_task exactly once on the active root task shown in the supervision context. Include the selected value and returned command-verification outcome in a concise summary and deliverable.',
+      'Do not make any other tool calls. Do not retry, save, update, apply, or roll back a recipe. If no candidate is justified, abstain and explicitly close the task with that limitation.',
+    ].join(' '),
+    fromLabel: 'Synthetic isolated R20 test driver',
+  }), `${context.evidence.id} real OMP lead task submission`)
+  context.evidence.taskId = sent.id ?? sent.task?.id ?? null
+  if (!context.evidence.taskId) throw new Error(`${context.evidence.id} real OMP task submission returned no task identifier`)
+}
+
+function taskToolEvents(task) {
+  const events = []
+  for (const message of task?.history ?? []) {
+    for (const part of message?.parts ?? []) {
+      const text = typeof part?.text === 'string' ? part.text : ''
+      const match = text.match(/^🔧\s+([a-z][a-z\d_]*)/i)
+      if (!match) continue
+      events.push({
+        toolName: match[1],
+        phase: /(?:✅\s*完成|❌\s*失败)/.test(text) ? 'result' : 'call',
+        isError: /❌\s*失败/.test(text),
+        messageId: message.messageId ?? null,
+        text,
+      })
+    }
+  }
+  return events
+}
+
+async function readRealOmpTask(context) {
+  const result = await app('GET', `/api/workshop/tasks/${encodeURIComponent(context.evidence.taskId)}?channelId=${encodeURIComponent(context.channel.channelId)}`)
+  if (result.status !== 200) throw new Error(`${context.evidence.id} task detail returned HTTP ${result.status}`)
+  return result.data
+}
+
+async function waitForRealOmpTaskCompletion(context) {
+  const deadline = Date.now() + REAL_OMP_TASK_COMPLETION_TIMEOUT_MS
+  let latest
+  while (Date.now() < deadline) {
+    latest = await readRealOmpTask(context)
+    if (latest.state === 'COMPLETED') return latest
+    if (['FAILED', 'CANCELED'].includes(latest.state)) {
+      throw new Error(`${context.evidence.id} task reached ${latest.state} before completing the approved operation`)
+    }
+    await delay(500)
+  }
+  throw new Error(`${context.evidence.id} task did not reach COMPLETED within the bounded post-approval window (state=${latest?.state ?? 'unknown'})`)
+}
+
+function assertSingleRealOmpCall(context, task) {
+  const events = taskToolEvents(task)
+  const calls = events.filter(event => event.phase === 'call')
+  const recipeCalls = calls.filter(event => event.toolName === 'recipe_trial')
+  const recipeResults = events.filter(event => event.toolName === 'recipe_trial' && event.phase === 'result')
+  const completionCalls = calls.filter(event => event.toolName === 'complete_task')
+  // The active lead turn is journaled through both the task-event mapper and
+  // the supervisor listener. A single host invocation can therefore appear
+  // twice in history; identical start text identifies that mirror. Route
+  // ledgers below remain authoritative for the number of executed trials.
+  const uniqueRecipeCalls = [...new Set(recipeCalls.map(event => event.text))]
+  if (events.some(event => !['recipe_trial', 'complete_task'].includes(event.toolName))
+    || recipeCalls.length < 1 || recipeCalls.length > 2 || uniqueRecipeCalls.length !== 1
+    || recipeResults.length < 1 || recipeResults.length > 2 || recipeResults.some(event => event.isError)
+    || completionCalls.length !== 1) {
+    throw new Error(`${context.evidence.id} OMP task history must show one recipe_trial invocation (allowing its mirrored journal entry), a successful result, and one explicit task-completion call`)
+  }
+  return { events, calls: [recipeCalls[0]], results: [recipeResults[0]], completionCalls }
+}
+
+async function approveRealOmpTrial(context) {
+  currentStage = `${context.evidence.id}:await-real-omp-hitl-approval`
+  const deadline = Date.now() + REAL_OMP_APPROVAL_TIMEOUT_MS
+  let pending
+  let latestTask
+  while (Date.now() < deadline) {
+    latestTask = await readRealOmpTask(context)
+    if (['FAILED', 'CANCELED', 'COMPLETED'].includes(latestTask.state)) {
+      throw new Error(`${context.evidence.id} task reached ${latestTask.state} before synthetic approval`)
+    }
+    const result = await app('GET', `/api/workshop/agent-tools/approvals?agentId=${encodeURIComponent(context.agent.id)}&scope=pending`)
+    if (result.status === 200) {
+      const matches = (result.data?.approvals ?? []).filter(item => item.agentId === context.agent.id && item.nodeId === `recipe:${context.recipe.id}` && item.status === 'pending')
+      if (matches.length > 1) throw new Error(`${context.evidence.id} has multiple pending real OMP trial approvals`)
+      if (matches.length === 1) {
+        const payload = matches[0].payload
+        const params = Array.isArray(payload?.params) ? payload.params : []
+        const proposedValue = Number(params[0]?.to)
+        const decisionStimulus = context.evidence.decisionStimulus
+        const allowedValues = decisionStimulus?.allowedValues ?? []
+        const isExpectedTrial = payload?.kind === 'recipe-gate'
+          && payload?.op === 'trial'
+          && payload?.recipeId === context.recipe.id
+          && payload?.emergency !== true
+          && payload?.emergency !== 'true'
+          && params.length === 1
+          && params[0]?.nodeId === context.control.id
+          && Number.isFinite(proposedValue)
+          && allowedValues.includes(proposedValue)
+          && proposedValue >= Number(context.evidence.simulator.mv.min)
+          && proposedValue <= Number(context.evidence.simulator.mv.max)
+          && Math.abs(proposedValue - Number(decisionStimulus.controlValue)) <= 30
+        if (!isExpectedTrial) throw new Error(`${context.evidence.id} refused a pending approval outside the registered recipe_trial candidate set`)
+        context.evidence.simulator.candidate = proposedValue
+        context.evidence.modelDecision = {
+          source: 'registered OMP host-tool proposal, checked against current protocol reads and prevalidated set',
+          controlValue: decisionStimulus.controlValue,
+          processValue: decisionStimulus.processValue,
+          targetBand: [decisionStimulus.targetMin, decisionStimulus.targetMax],
+          allowedValues,
+          selectedValue: proposedValue,
+          selectionRationale: safeText(payload?.hypothesis ?? payload?.rationale ?? '', 400),
+          selectedValueWithinRangeAndStep: true,
+        }
+        pending = matches[0]
+        break
+      }
+    }
+    await delay(500)
+  }
+  if (!pending) throw new Error(`${context.evidence.id} real OMP recipe_trial approval did not appear within the bounded window`)
+  context.evidence.approvals.push({ id: pending.id, statusBeforeDecision: pending.status, nodeId: pending.nodeId, title: pending.title ?? null, candidateValue: context.evidence.simulator.candidate, decision: 'approve' })
+  requireSuccess(await app('POST', `/api/workshop/agent-tools/approvals/${encodeURIComponent(pending.id)}/decide`, {
+    approved: true,
+    comment: 'Approved for the bounded R20 local simulator integration check.',
+  }), `${context.evidence.id} synthetic real OMP HITL approval`)
+  const history = await app('GET', `/api/workshop/agent-tools/approvals?agentId=${encodeURIComponent(context.agent.id)}&scope=history`)
+  const finalApproval = (history.data?.approvals ?? []).find(item => item.id === pending.id)
+  if (history.status !== 200 || finalApproval?.status !== 'approved' || finalApproval?.decidedBy !== syntheticAdminId) {
+    throw new Error(`${context.evidence.id} real OMP HITL history did not confirm the synthetic administrator decision`)
+  }
+  context.evidence.approvals[0] = {
+    ...context.evidence.approvals[0],
+    statusAfterDecision: finalApproval.status,
+    decisionActor: 'synthetic isolated administrator via protected approval API',
+    commentRecorded: finalApproval.comment === 'Approved for the bounded R20 local simulator integration check.',
+  }
+}
+
+async function captureRealOmpUsage(context) {
+  const unavailable = {
+    source: 'not exposed by the channel queue endpoint',
+    context: null,
+    requestInputTokens: null,
+    requestOutputTokens: null,
+    costUsd: null,
+    costAvailable: false,
+  }
+  if (!context?.channel?.channelId || !context?.agent?.id) return unavailable
+  try {
+    const queueResult = await app('GET', `/api/workshop/channels/${encodeURIComponent(context.channel.channelId)}/queue`)
+    if (queueResult.status !== 200) return unavailable
+    const queue = extractPayload(queueResult.body)
+    const agentStatus = (Array.isArray(queue) ? queue : []).find(item => item.agentId === context.agent.id)
+    const raw = agentStatus?.context
+    if (!raw) return unavailable
+    return {
+      source: 'channel queue OMP context snapshot (not request-level billing)',
+      context: {
+        usedTokens: Number.isFinite(Number(raw.usedTokens)) ? Number(raw.usedTokens) : null,
+        contextWindow: Number.isFinite(Number(raw.contextWindow)) ? Number(raw.contextWindow) : null,
+        percent: Number.isFinite(Number(raw.percent)) ? Number(raw.percent) : null,
+        compacting: raw.compacting === true,
+      },
+      requestInputTokens: null,
+      requestOutputTokens: null,
+      costUsd: null,
+      costAvailable: false,
+    }
+  }
+  catch { return unavailable }
+}
+
+async function approvePendingTrial(context) {
+  currentStage = `${context.evidence.id}:manual-hitl-approval`
+  const deadline = Date.now() + 20_000
+  let pending
+  while (Date.now() < deadline) {
+    const result = await app('GET', `/api/workshop/agent-tools/approvals?agentId=${encodeURIComponent(context.agent.id)}&scope=pending`)
+    if (result.status === 200) {
+      const matches = (result.data?.approvals ?? []).filter(item => item.agentId === context.agent.id && item.nodeId === `recipe:${context.recipe.id}` && item.status === 'pending')
+      if (matches.length > 1) throw new Error(`${context.evidence.id} has multiple pending trial approvals`)
+      if (matches.length === 1) {
+        pending = matches[0]
+        break
+      }
+    }
+    await delay(400)
+  }
+  if (!pending) throw new Error(`${context.evidence.id} manual recipe_trial HITL approval did not appear within the bounded window`)
+  const payload = pending.payload ?? {}
+  const params = Array.isArray(payload.params) ? payload.params : []
+  const proposedValue = Number(params[0]?.to)
+  if (payload.kind !== 'recipe-gate'
+    || payload.op !== 'trial'
+    || payload.recipeId !== context.recipe.id
+    || params.length !== 1
+    || params[0]?.nodeId !== context.control.id
+    || proposedValue !== Number(piFixtureTarget?.value)
+    || payload.emergency === true
+    || payload.emergency === 'true') {
+    throw new Error(`${context.evidence.id} refused a Pi approval outside the scripted bounded recipe_trial request`)
+  }
+  context.evidence.simulator.candidate = proposedValue
+  context.evidence.modelDecision = {
+    source: 'registered Pi host-tool request generated by a deterministic loopback model fixture',
+    controlValue: context.evidence.decisionStimulus?.controlValue ?? null,
+    processValue: context.evidence.decisionStimulus?.processValue ?? null,
+    targetBand: [context.evidence.decisionStimulus?.targetMin, context.evidence.decisionStimulus?.targetMax],
+    allowedValues: context.evidence.decisionStimulus?.allowedValues ?? [],
+    selectedValue: proposedValue,
+    selectionRationale: safeText(payload.hypothesis ?? payload.rationale ?? '', 400),
+    selectedValueWithinRangeAndStep: true,
+  }
+  context.evidence.approvals.push({ id: pending.id, taskId: context.evidence.taskId, statusBeforeDecision: pending.status, nodeId: pending.nodeId, candidateValue: proposedValue, title: pending.title ?? null, decision: 'approve' })
+  const decided = await app('POST', `/api/workshop/agent-tools/approvals/${encodeURIComponent(pending.id)}/decide`, {
+    approved: true,
+    comment: 'Approved for the isolated deterministic simulator integration check.',
+  })
+  requireSuccess(decided, `${context.evidence.id} explicit HITL approval`)
+  const history = await app('GET', `/api/workshop/agent-tools/approvals?agentId=${encodeURIComponent(context.agent.id)}&scope=history`)
+  const finalApproval = (history.data?.approvals ?? []).find(item => item.id === pending.id)
+  if (history.status !== 200 || finalApproval?.status !== 'approved' || finalApproval?.decidedBy !== syntheticAdminId) {
+    throw new Error(`${context.evidence.id} HITL approval history failed to confirm the explicit administrator decision`)
+  }
+  context.evidence.approvals[0] = {
+    ...context.evidence.approvals[0],
+    statusAfterDecision: finalApproval.status,
+    decisionActor: 'synthetic isolated administrator via protected approval API',
+    commentRecorded: finalApproval.comment === 'Approved for the isolated deterministic simulator integration check.',
+  }
+}
+
+async function waitForPiTaskCompletion(context) {
+  const deadline = Date.now() + PI_TASK_COMPLETION_TIMEOUT_MS
+  let latest
+  while (Date.now() < deadline) {
+    latest = await readRealOmpTask(context)
+    if (latest.assigneeId !== context.agent.id) {
+      throw new Error(`${context.evidence.id} Pi task was not executed by its registered agent`)
+    }
+    if (latest.state === 'COMPLETED') return latest
+    if (['FAILED', 'CANCELED'].includes(latest.state)) {
+      throw new Error(`${context.evidence.id} Pi task reached ${latest.state} before returning the tool result`)
+    }
+    await delay(300)
+  }
+  throw new Error(`${context.evidence.id} Pi task did not complete within the bounded window (state=${latest?.state ?? 'unknown'})`)
+}
+
+async function collectSceneEvidence(context) {
+  currentStage = `${context.evidence.id}:collect-result-and-observations`
+  let trace
+  let toolName
+  let toolText
+  let toolIsError
+  let realTask
+  let realToolEvents
+  if (HARNESS_MODE === 'pi-local-fixture') {
+    const finalTask = await waitForPiTaskCompletion(context)
+    const traceDeadline = Date.now() + 10_000
+    let sceneRequests = piApiTrace.filter(item => item.scene === context.evidence.id)
+    while (sceneRequests.length < 3 && Date.now() < traceDeadline) {
+      await delay(100)
+      sceneRequests = piApiTrace.filter(item => item.scene === context.evidence.id)
+    }
+    const proposal = sceneRequests[0]
+    const recipeResult = sceneRequests[1]
+    const completionResult = sceneRequests[2]
+    const traceChecks = {
+      exactlyThreeModelRequests: sceneRequests.length === 3,
+      recipeTrialCall: proposal?.response?.type === 'tool_call' && proposal?.response?.toolName === 'recipe_trial',
+      approvedCandidateMatches: Number(proposal?.response?.candidate) === Number(context.evidence.simulator.candidate),
+      completionToolAvailable: proposal?.toolNames?.includes('complete_task') === true,
+      recipeResultFollowedByTaskClosure: recipeResult?.response?.type === 'tool_call' && recipeResult?.response?.toolName === 'complete_task',
+      recipeResultReturnedToPi: recipeResult?.toolResultCallId === `call-${context.evidence.id}` && recipeResult?.toolResultMessageCount === 1 && Boolean(recipeResult?.toolResultText),
+      recipeResultWasSuccessful: !/(?:isError\s*[:=]\s*true|❌\s*失败|tool error)/i.test(recipeResult?.toolResultText ?? ''),
+      completeTaskResultReturnedToPi: completionResult?.response?.type === 'final_text_after_task_completion'
+        && completionResult?.toolResultCallId === `complete-${context.evidence.id}`
+        && completionResult?.toolResultMessageCount === 2
+        && /COMPLETED/.test(completionResult?.toolResultText ?? ''),
+    }
+    if (Object.values(traceChecks).some(passed => !passed)) {
+      throw new Error(`${context.evidence.id} Pi model trace check failed: ${JSON.stringify(Object.fromEntries(Object.entries(traceChecks).filter(([, passed]) => !passed)))}`)
+    }
+    toolName = 'recipe_trial'
+    toolText = recipeResult.toolResultText
+    toolIsError = /(?:isError\s*[:=]\s*true|❌\s*失败|tool error)/i.test(toolText)
+    context.evidence.modelExecution = {
+      harness: 'registered Pi CLI adapter',
+      provider: PI_PROVIDER,
+      model: PI_MODEL,
+      modelMode: 'deterministic local OpenAI-compatible fixture; no LLM or external provider',
+      taskId: context.evidence.taskId,
+      taskState: finalTask.state,
+      taskCompleted: finalTask.state === 'COMPLETED',
+      taskAssigneeMatchesRegisteredPi: finalTask.assigneeId === context.agent.id,
+      modelApiRequests: sceneRequests.map(item => ({
+        requestId: item.requestId,
+        requestModel: item.requestModel,
+        loopbackPeer: item.loopbackPeer,
+        apiKeyMatchesLocalFixture: item.apiKeyMatchesLocalFixture,
+        streamRequested: item.streamRequested,
+        toolNames: item.toolNames,
+        messageRoles: item.messageRoles,
+        toolResultMessageCount: item.toolResultMessageCount,
+        response: item.response,
+      })),
+      recipeTrialResultReturnedToPi: recipeResult.toolResultMessageCount === 1,
+      completeTaskCallIssued: recipeResult.response.toolName === 'complete_task',
+      completeTaskResultReturnedToPi: completionResult.toolResultMessageCount === 2,
+      completeTaskResultConfirmed: /COMPLETED/.test(completionResult.toolResultText),
+    }
+  }
+  else if (HARNESS_MODE === 'real-omp') {
+    // The supervisor journals compact host-tool start/result events on the
+    // active root task. Approval and operation ledgers remain authoritative
+    // for the proposal and its execution.
+    realTask = await waitForRealOmpTaskCompletion(context)
+    if (realTask.assigneeId !== context.agent.id) throw new Error(`${context.evidence.id} real OMP task was not executed by its registered lead agent`)
+    realToolEvents = assertSingleRealOmpCall(context, realTask)
+    const channelMessages = requireSuccess(await app('GET', `/api/workshop/channels/${encodeURIComponent(context.channel.channelId)}/messages?limit=500`), `${context.evidence.id} channel task evidence`)
+    const taskMessages = (Array.isArray(channelMessages) ? channelMessages : []).filter(message => message.taskId === context.evidence.taskId)
+    const usage = await captureRealOmpUsage(context)
+    context.evidence.modelExecution = {
+      provider: REAL_OMP_PROVIDER,
+      model: REAL_OMP_MODEL,
+      taskId: context.evidence.taskId,
+      taskState: realTask.state,
+      taskArtifactCount: Array.isArray(realTask.artifacts) ? realTask.artifacts.length : 0,
+      evidenceSource: 'registered OMP task history, approval payload, task state, and route-specific recipe/ACK records',
+      hostToolCallEvents: realToolEvents.calls.length + realToolEvents.completionCalls.length,
+      hostToolResultEvents: realToolEvents.results.length,
+      taskCompletionToolCallEvents: realToolEvents.completionCalls.length,
+      taskToolEvents: realToolEvents.events.map(event => ({ toolName: event.toolName, phase: event.phase, isError: event.isError })),
+      hostToolEventHistoryAvailable: realToolEvents.events.length > 0,
+      approvedProposalValue: context.evidence.modelDecision?.selectedValue ?? null,
+      approvedProposalCount: context.evidence.approvals.length,
+      channelEvidence: {
+        source: 'GET /api/workshop/channels/:id/messages',
+        taskMessageCount: taskMessages.length,
+        agentMessageCount: taskMessages.filter(message => message.fromAgentId === context.agent.id).length,
+        roles: [...new Set(taskMessages.map(message => safeText(message.role, 40)))],
+      },
+      usage,
+    }
+  }
+  else if (HARNESS_MODE === 'fixture') {
+    const tracePath = fixtureTracePath(context.evidence.id)
+    const deadline = Date.now() + 20_000
+    while (Date.now() < deadline) {
+      if (existsSync(tracePath)) {
+        try {
+          trace = JSON.parse(readFileSync(tracePath, 'utf8'))
+        }
+        catch { /* fixture is rewriting a small JSON trace */ }
+      }
+      if (trace?.hostToolResults?.length) break
+      await delay(300)
+    }
+    if (!trace) throw new Error(`${context.evidence.id} deterministic fixture did not create its trace`)
+    if (trace.hostToolCalls?.length !== 1 || trace.hostToolCalls[0]?.toolName !== 'recipe_trial' || trace.hostToolResults?.length !== 1) {
+      throw new Error(`${context.evidence.id} fixture trace did not show exactly one recipe_trial host_tool_call/result`)
+    }
+    const toolResult = trace.hostToolResults[0].result
+    toolText = String(toolResult?.content?.[0]?.text ?? '')
+    toolName = trace.hostToolCalls[0].toolName
+    toolIsError = Boolean(toolResult?.isError)
+    if (toolIsError || !/全部设备证实|下发并全部设备证实/.test(toolText)) {
+      throw new Error(`${context.evidence.id} recipe_trial did not return a fully verified device ACK`)
+    }
+  }
+
+  const readback = await app('POST', `/api/workshop/dcw/${encodeURIComponent(context.control.id)}/read`)
+  requireSuccess(readback, `${context.evidence.id} DCW readback`)
+
+  const runData = await app('GET', `/api/workshop/dcw/runs/${encodeURIComponent(context.runId)}/data`)
+  requireSuccess(runData, `${context.evidence.id} active run data`)
+  const logs = await app('GET', `/api/workshop/ops-logs?scope=line&lineId=${encodeURIComponent(context.line.id)}&recipeId=${encodeURIComponent(context.recipe.id)}&limit=100`)
+  requireSuccess(logs, `${context.evidence.id} recipe operation log`)
+  const recipeRows = (logs.data?.logs ?? [])
+    .filter(row => String(row.action).startsWith('recipe.'))
+    .map((row) => {
+      let detail = row.detail
+      if (!detail || typeof detail !== 'object' || Array.isArray(detail)) {
+        try {
+          const parsed = JSON.parse(String(row.detailJson ?? '{}'))
+          detail = parsed && typeof parsed === 'object' && !Array.isArray(parsed) ? parsed : {}
+        }
+        catch { detail = {} }
+      }
+      return { ...row, detail }
+    })
+  const trialRows = recipeRows.filter(row => row.action === 'recipe.trial')
+  const trialAckRows = recipeRows.filter(row => row.action === 'recipe.dispatch.ack' && row.detail?.trial === true)
+  const trialRow = trialRows[0]
+  const ackRow = trialAckRows[0]
+  const trialRunId = String(ackRow?.detail?.runId ?? '')
+  if (trialRows.length !== 1 || trialAckRows.length !== 1 || !trialRunId || String(trialRow?.detail?.runId ?? '') !== trialRunId) {
+    throw new Error(`${context.evidence.id} recipe log did not contain exactly one matching recipe.trial and trial ACK pair`)
+  }
+  if (HARNESS_MODE === 'real-omp' && context.evidence.approvals.length !== 1) {
+    throw new Error(`${context.evidence.id} expected exactly one approved real OMP proposal`)
+  }
+  const trialRunData = await app('GET', `/api/workshop/dcw/runs/${encodeURIComponent(trialRunId)}/data`)
+  requireSuccess(trialRunData, `${context.evidence.id} trial run data`)
+  const ack = ackRow.detail?.ackSummary ?? null
+  if (!ack || ack.total !== 1 || ack.verified !== 1 || ack.failed !== 0 || ack.unverified !== 0) {
+    throw new Error(`${context.evidence.id} logged trial ACK is not one-of-one verified`)
+  }
+  const returnedTrialRun = trialRunData.data?.run
+  const trialResults = returnedTrialRun?.results
+  if (returnedTrialRun?.id !== trialRunId || !Array.isArray(trialResults) || trialResults.length !== 1 || trialResults[0]?.verify?.verdict !== 'verified' || trialResults[0]?.ok !== true) {
+    throw new Error(`${context.evidence.id} trial run data lacks an independently verified write result`)
+  }
+
+  const snapshotDelayStartedAt = new Date().toISOString()
+  await delay(20_000)
+  const probeStartedAt = new Date().toISOString()
+  const probeResponse = await app('POST', `/api/workshop/daq/${encodeURIComponent(context.daq.id)}/test`)
+  const probe = requireSuccess(probeResponse, `${context.evidence.id} DAQ driver probe`).test
+  const sampleValue = Number(probe?.sampleValue)
+  if (probe?.ok !== true || !Number.isFinite(sampleValue)) {
+    throw new Error(`${context.evidence.id} DAQ driver probe did not return a finite simulator value`)
+  }
+  const probeCompletedAt = new Date().toISOString()
+  const ingestAt = probeCompletedAt
+  const ingest = requireSuccess(await app('POST', `/api/workshop/daq/${encodeURIComponent(context.daq.id)}/ingest`, {
+    at: ingestAt,
+    value: sampleValue,
+    state: 'ok',
+  }), `${context.evidence.id} DAQ snapshot ingest`)
+  if (ingest?.accepted !== true || ingest.nodeId !== context.daq.id || ingest.lineId !== context.line.id) {
+    throw new Error(`${context.evidence.id} DAQ snapshot ingest was not accepted for the expected node and line`)
+  }
+
+  const persistedAt = String(ingest.at ?? ingestAt)
+  const persistedAtMs = Date.parse(persistedAt)
+  if (!Number.isFinite(persistedAtMs)) throw new Error(`${context.evidence.id} DAQ snapshot ingest returned an invalid sample timestamp`)
+  const fromMs = persistedAtMs
+  const toMs = persistedAtMs + 1
+  // Let the in-process queue and SQLite batch flush settle before a fixed-window query;
+  // retries outlast the short query-cache TTL if the first lookup is still empty.
+  await delay(800)
+  const queryDeadline = Date.now() + 12_000
+  let taggedQuery
+  let taggedSeries
+  let snapshotBucketAt
+  while (Date.now() < queryDeadline) {
+    const candidate = await app('GET', `/api/workshop/daq/query?line_id=${encodeURIComponent(context.line.id)}&node_ids=${encodeURIComponent(context.daq.id)}&run_id=${encodeURIComponent(context.runId)}&from=${fromMs}&to=${toMs}&bucket_ms=1000`)
+    if (candidate.status === 200) {
+      taggedQuery = candidate
+      taggedSeries = candidate.data?.series?.find(series => series.nodeId === context.daq.id)
+      const bucketMs = Number(candidate.data?.window?.bucketMs) || 1_000
+      snapshotBucketAt = Math.floor(persistedAtMs / bucketMs) * bucketMs
+      const matchingPoint = taggedSeries?.points?.find(point => Number(point.at) === snapshotBucketAt && Math.abs(Number(point.v) - sampleValue) <= 0.01)
+      if (matchingPoint && candidate.data?.activeRun?.runId === context.runId) break
+    }
+    await delay(300)
+  }
+  if (!taggedQuery || !taggedSeries?.points?.length || taggedQuery.data?.activeRun?.runId !== context.runId) {
+    throw new Error(`${context.evidence.id} platform DAQ query returned no snapshot tagged to the active run`)
+  }
+  const matchingProbePoint = taggedSeries.points.find(point => Number(point.at) === snapshotBucketAt && Math.abs(Number(point.v) - sampleValue) <= 0.01)
+  if (!matchingProbePoint) {
+    throw new Error(`${context.evidence.id} run-tagged DAQ data did not contain the driver-probe snapshot value`)
+  }
+
+  if (HARNESS_MODE === 'real-omp') {
+    const finalTask = await readRealOmpTask(context)
+    context.evidence.hostToolExecution = {
+      provenance: 'registered OMP model proposal → protected synthetic approval → one recipe.trial and one verified DCW receipt',
+      hostToolEventHistoryAvailable: realToolEvents.events.length > 0,
+      taskStateAfterObservation: finalTask.state,
+      taskCompleted: finalTask.state === 'COMPLETED',
+      approvedProposalCount: context.evidence.approvals.length,
+      recipeTrialCount: trialRows.length,
+      recipeTrialAckCount: trialAckRows.length,
+      toolName: 'recipe_trial',
+      resultReceived: ack?.verified === 1 && ack?.failed === 0,
+    }
+    context.evidence.modelExecution.taskStateAfterObservation = finalTask.state
+    context.evidence.modelExecution.taskCompleted = finalTask.state === 'COMPLETED'
+  }
+  else {
+    context.evidence.fixtureTrace = {
+      fixture: HARNESS_MODE === 'pi-local-fixture' ? 'registered Pi CLI with deterministic loopback model fixture' : trace.fixture,
+      decisionSource: HARNESS_MODE === 'pi-local-fixture' ? 'Pi CLI tool call through the AgentWorkShop host-tool bridge' : trace.decisionSource,
+      hostToolCallCount: HARNESS_MODE === 'pi-local-fixture' ? 1 : trace.hostToolCalls.length,
+      toolName,
+      resultReceived: HARNESS_MODE === 'pi-local-fixture' ? Boolean(context.evidence.modelExecution?.recipeTrialResultReturnedToPi) : trace.hostToolResults.length === 1,
+      toolResultText: safeText(toolText, 1200),
+      toolIsError,
+    }
+  }
+  context.evidence.trial = {
+    taskId: context.evidence.taskId,
+    runId: trialRunId,
+    toolName,
+    approvalId: context.evidence.approvals[0].id,
+    ackSummary: ack,
+    resultCount: trialResults.length,
+    resultRows: trialResults,
+  }
+  context.evidence.readback = {
+    expectedCandidate: context.evidence.simulator.candidate,
+    response: readback.data?.outcome ?? readback.data?.read ?? readback.data,
+    resultMatchesCandidate: (() => {
+      const details = readback.data?.outcome ?? readback.data?.read ?? readback.data ?? {}
+      const observed = Number(details.value ?? details.data?.value ?? details.result?.value)
+      return Number.isFinite(observed) && Math.abs(observed - context.evidence.simulator.candidate) <= 0.01
+    })(),
+  }
+  context.evidence.daq = {
+    taskId: context.evidence.taskId,
+    nodeId: context.daq.id,
+    method: 'single driver-probe/ingest snapshot',
+    storageBackend: 'isolated AW_HOME SQLite time-series fallback',
+    postTrialDelaySeconds: 20,
+    snapshotDelayStartedAt,
+    sampleCount: taggedSeries.points.length,
+    runTags: [context.runId],
+    expectedActiveRunId: context.runId,
+    probe: {
+      driver: context.daq.driver,
+      startedAt: probeStartedAt,
+      completedAt: probeCompletedAt,
+      sampleValue,
+      latencyMs: Number.isFinite(Number(probe.latencyMs)) ? Number(probe.latencyMs) : null,
+    },
+    ingest: {
+      accepted: ingest.accepted,
+      at: persistedAt,
+      value: sampleValue,
+      state: 'ok',
+    },
+    taggedQuery: {
+      requestedRunId: context.runId,
+      returnedRunId: taggedQuery.data?.activeRun?.runId ?? null,
+      seriesPointCount: taggedSeries.points.length,
+      probeValueMatched: true,
+      snapshotBucketAt: new Date(snapshotBucketAt).toISOString(),
+      matchingProbePoint,
+      series: taggedSeries.points.slice(-5),
+    },
+    activeRunDataKeys: Object.keys(runData.data ?? {}),
+  }
+  context.evidence.recipeLog = {
+    taskId: context.evidence.taskId,
+    query: 'ops-logs filtered by lineId and recipeId (recipe_log-equivalent audit view)',
+    rowCount: recipeRows.length,
+    trialAction: trialRow.action,
+    trialSummary: safeText(trialRow.summary, 600),
+    ackAction: ackRow.action,
+    ackSummary: safeText(ackRow.summary, 600),
+    trialRunDataKeys: Object.keys(trialRunData.data ?? {}),
+  }
+  context.evidence.status = 'verified'
+  context.evidence.traceability = {
+    taskId: context.evidence.taskId,
+    approvalId: context.evidence.approvals[0].id,
+    recipeId: context.recipe.id,
+    trialRunId,
+    activeRunId: context.runId,
+    daqNodeId: context.daq.id,
+    sampleAt: persistedAt,
+    taskState: context.evidence.modelExecution?.taskState ?? null,
+    taskCompleted: context.evidence.modelExecution?.taskCompleted === true,
+  }
+}
+
+async function runScene(scene) {
+  const simDetails = await simulatorScene(scene)
+  const context = await setupScene(scene, { ...simDetails, candidate: scene.candidate })
+  context.evidence.simulator.candidate = scene.candidate
+  context.evidence.simulator.baseline = scene.baseline
+  piFixtureTarget = {
+    scene: context.evidence.id,
+    recipeId: context.recipe.id,
+    nodeId: context.control.id,
+    value: scene.candidate,
+  }
+  await submitFixtureTask(context)
+  await approvePendingTrial(context)
+  await collectSceneEvidence(context)
+  check(`${scene.id}-pi-task-completed`, context.evidence.modelExecution?.taskCompleted === true && context.evidence.modelExecution?.taskAssigneeMatchesRegisteredPi === true, {
+    taskId: context.evidence.taskId,
+    state: context.evidence.modelExecution?.taskState ?? null,
+    assigneeMatchesRegisteredPi: context.evidence.modelExecution?.taskAssigneeMatchesRegisteredPi === true,
+  })
+  const linked = context.evidence.traceability?.taskId === context.evidence.taskId
+    && context.evidence.approvals[0]?.taskId === context.evidence.taskId
+    && context.evidence.trial?.taskId === context.evidence.taskId
+    && context.evidence.daq?.taskId === context.evidence.taskId
+    && context.evidence.traceability?.approvalId === context.evidence.approvals[0]?.id
+    && context.evidence.traceability?.trialRunId === context.evidence.trial?.runId
+    && context.evidence.traceability?.activeRunId === context.evidence.daq?.expectedActiveRunId
+    && context.evidence.daq?.taggedQuery?.returnedRunId === context.evidence.daq?.expectedActiveRunId
+    && Boolean(context.evidence.traceability?.sampleAt)
+  check(`${scene.id}-task-approval-receipt-sample-linked`, linked, context.evidence.traceability)
+  piFixtureTarget = null
+  const stopped = await app('POST', `/api/workshop/dcw/lines/${encodeURIComponent(context.line.id)}/stop`)
+  requireSuccess(stopped, `${scene.id} line stop`)
+  context.evidence.activeRun.endedAt = stopped.data?.run?.endedAt ?? null
+  check(`${scene.id}-host-dispatch-one-call`, context.evidence.fixtureTrace.hostToolCallCount === 1, { tool: context.evidence.fixtureTrace.toolName })
+  check(`${scene.id}-ack-verified`, context.evidence.trial.ackSummary.verified === 1 && context.evidence.trial.ackSummary.failed === 0, context.evidence.trial.ackSummary)
+  check(`${scene.id}-readback-matches`, context.evidence.readback.resultMatchesCandidate, context.evidence.readback)
+  check(`${scene.id}-run-tagged-daq-probe-ingest-snapshot`, context.evidence.daq.taggedQuery.seriesPointCount >= 1 && context.evidence.daq.taggedQuery.probeValueMatched, {
+    method: context.evidence.daq.method,
+    value: context.evidence.daq.probe.sampleValue,
+    runId: context.evidence.daq.expectedActiveRunId,
+    matchingPoint: context.evidence.daq.taggedQuery.matchingProbePoint,
+  })
+  check(`${scene.id}-recipe-log`, context.evidence.recipeLog.rowCount >= 2, { trialAction: context.evidence.recipeLog.trialAction, ackAction: context.evidence.recipeLog.ackAction })
+}
+
+async function recoverFailedRealOmpTrial(context, outcome) {
+  const recovery = []
+  if (context?.evidence?.taskId) {
+    try {
+      const task = await readRealOmpTask(context)
+      if (!['COMPLETED', 'FAILED', 'CANCELED'].includes(task.state)) {
+        const canceled = await app('POST', `/api/workshop/tasks/${encodeURIComponent(context.evidence.taskId)}/cancel`)
+        recovery.push({ action: 'cancel-nonterminal-task', status: canceled.status === 200 ? 'completed' : `http-${canceled.status}` })
+      }
+      else recovery.push({ action: 'cancel-nonterminal-task', status: 'not-needed', taskState: task.state })
+    }
+    catch (error) { recovery.push({ action: 'cancel-nonterminal-task', status: 'failed', error: compactError(error) }) }
+  }
+  if (context?.channel?.channelId && context?.agent?.id) {
+    try {
+      const stopped = await app('POST', `/api/workshop/channels/${encodeURIComponent(context.channel.channelId)}/agents/${encodeURIComponent(context.agent.id)}/stop`)
+      recovery.push({ action: 'stop-agent-runtime', status: stopped.status === 200 ? 'completed' : `http-${stopped.status}` })
+    }
+    catch (error) { recovery.push({ action: 'stop-agent-runtime', status: 'failed', error: compactError(error) }) }
+    try {
+      const pendingResult = await app('GET', `/api/workshop/agent-tools/approvals?agentId=${encodeURIComponent(context.agent.id)}&scope=pending`)
+      const pending = pendingResult.status === 200
+        ? (pendingResult.data?.approvals ?? []).filter(item => item.agentId === context.agent.id && item.status === 'pending')
+        : []
+      for (const approval of pending) {
+        const denied = await app('POST', `/api/workshop/agent-tools/approvals/${encodeURIComponent(approval.id)}/decide`, {
+          approved: false,
+          comment: 'Denied during bounded R20 failure cleanup.',
+        })
+        recovery.push({ action: 'deny-pending-agent-approval', approvalId: approval.id, status: denied.status === 200 ? 'completed' : `http-${denied.status}` })
+      }
+      if (pendingResult.status !== 200) recovery.push({ action: 'deny-pending-agent-approval', status: `list-http-${pendingResult.status}` })
+      else if (pending.length === 0) recovery.push({ action: 'deny-pending-agent-approval', status: 'not-needed' })
+    }
+    catch (error) { recovery.push({ action: 'deny-pending-agent-approval', status: 'failed', error: compactError(error) }) }
+  }
+  if (context?.line?.id) {
+    try {
+      const stopped = await app('POST', `/api/workshop/dcw/lines/${encodeURIComponent(context.line.id)}/stop`)
+      recovery.push({ action: 'stop-line', status: stopped.status === 200 ? 'completed' : `http-${stopped.status}` })
+    }
+    catch (error) { recovery.push({ action: 'stop-line', status: 'failed', error: compactError(error) }) }
+  }
+  outcome.recovery = recovery
+}
+
+async function runRealOmpScene(baseScene) {
+  currentScene = baseScene.id
+  let simDetails
+  try {
+    currentStage = `${baseScene.id}:simulator-setup-once`
+    simDetails = await simulatorScene(baseScene)
+    if (!Array.isArray(baseScene.candidates) || baseScene.candidates.length !== REAL_OMP_TRIALS_PER_SCENE
+      || baseScene.candidates.some(candidate => !Number.isFinite(candidate)
+        || candidate < Number(simDetails.mv.min)
+        || candidate > Number(simDetails.mv.max)
+        || Math.abs(candidate - baseScene.baseline) > 30)) {
+      throw new Error(`${baseScene.id} real OMP candidates violate the configured count, process range, or 30-unit recipe step limit`)
+    }
+  }
+  catch (error) {
+    for (let repetition = 1; repetition <= REAL_OMP_TRIALS_PER_SCENE; repetition++) {
+      const outcome = {
+        id: `${baseScene.id}-r${repetition}`,
+        scene: baseScene.id,
+        repetition,
+        provider: REAL_OMP_PROVIDER,
+        model: REAL_OMP_MODEL,
+        candidate: baseScene.candidates[repetition - 1],
+        usage: { source: 'not available because simulator setup failed', context: null, requestInputTokens: null, requestOutputTokens: null, costUsd: null, costAvailable: false },
+        costUsd: null,
+        status: 'failed',
+        failure: compactError(error),
+      }
+      evidence.experiment.runOutcomes.push(outcome)
+    }
+    return
+  }
+
+  for (let repetition = 1; repetition <= REAL_OMP_TRIALS_PER_SCENE; repetition++) {
+    const scene = { ...baseScene, id: `${baseScene.id}-r${repetition}` }
+    const outcome = {
+      id: scene.id,
+      scene: baseScene.id,
+      repetition,
+      provider: REAL_OMP_PROVIDER,
+      model: REAL_OMP_MODEL,
+      candidate: null,
+      candidateChoices: [...baseScene.candidates],
+      usage: { source: 'not yet available', context: null, requestInputTokens: null, requestOutputTokens: null, costUsd: null, costAvailable: false },
+      costUsd: null,
+      status: 'running',
+      startedAt: new Date().toISOString(),
+    }
+    evidence.experiment.runOutcomes.push(outcome)
+    let context
+    try {
+      currentScene = scene.id
+      const runSimDetails = { ...simDetails, candidateChoices: [...baseScene.candidates] }
+      context = await setupScene(scene, runSimDetails)
+      context.evidence.simulator.baseline = scene.baseline
+      outcome.entities = context.evidence.entities
+      outcome.runId = context.runId
+      await submitRealOmpTask(context)
+      outcome.taskId = context.evidence.taskId
+      await approveRealOmpTrial(context)
+      outcome.candidate = context.evidence.simulator.candidate
+      outcome.modelDecision = context.evidence.modelDecision
+      await collectSceneEvidence(context)
+      const stopped = await app('POST', `/api/workshop/dcw/lines/${encodeURIComponent(context.line.id)}/stop`)
+      requireSuccess(stopped, `${scene.id} line stop`)
+      context.evidence.activeRun.endedAt = stopped.data?.run?.endedAt ?? null
+      const finalTask = await readRealOmpTask(context)
+      context.evidence.hostToolExecution.taskStateAfterObservation = finalTask.state
+      context.evidence.hostToolExecution.taskCompleted = finalTask.state === 'COMPLETED'
+      context.evidence.modelExecution.taskStateAfterObservation = finalTask.state
+      context.evidence.modelExecution.taskCompleted = finalTask.state === 'COMPLETED'
+      check(`${scene.id}-real-omp-approved-model-proposal`, context.evidence.modelExecution.approvedProposalCount === 1 && context.evidence.modelExecution.approvedProposalValue === context.evidence.simulator.candidate, {
+        provider: REAL_OMP_PROVIDER,
+        model: REAL_OMP_MODEL,
+        selectedValue: context.evidence.modelExecution.approvedProposalValue,
+        toolEventHistoryAvailable: context.evidence.modelExecution.hostToolEventHistoryAvailable,
+      })
+      check(`${scene.id}-real-omp-one-recipe-trial-ledger`, context.evidence.hostToolExecution.recipeTrialCount === 1 && context.evidence.hostToolExecution.recipeTrialAckCount === 1, {
+        recipeTrialCount: context.evidence.hostToolExecution.recipeTrialCount,
+        recipeTrialAckCount: context.evidence.hostToolExecution.recipeTrialAckCount,
+      })
+      check(`${scene.id}-ack-verified`, context.evidence.trial.ackSummary.verified === 1 && context.evidence.trial.ackSummary.failed === 0, context.evidence.trial.ackSummary)
+      check(`${scene.id}-readback-matches`, context.evidence.readback.resultMatchesCandidate, context.evidence.readback)
+      check(`${scene.id}-run-tagged-daq-probe-ingest-snapshot`, context.evidence.daq.taggedQuery.seriesPointCount >= 1 && context.evidence.daq.taggedQuery.probeValueMatched, {
+        method: context.evidence.daq.method,
+        value: context.evidence.daq.probe.sampleValue,
+        runId: context.evidence.daq.expectedActiveRunId,
+        matchingPoint: context.evidence.daq.taggedQuery.matchingProbePoint,
+      })
+      check(`${scene.id}-recipe-log`, context.evidence.recipeLog.rowCount >= 2, { trialAction: context.evidence.recipeLog.trialAction, ackAction: context.evidence.recipeLog.ackAction })
+      check(`${scene.id}-real-omp-task-completed`, context.evidence.hostToolExecution.taskCompleted, {
+        taskState: context.evidence.hostToolExecution.taskStateAfterObservation,
+      })
+      context.evidence.status = 'verified'
+      outcome.status = context.evidence.hostToolExecution.taskCompleted ? 'verified' : 'verified-operation-task-incomplete'
+      outcome.finishedAt = new Date().toISOString()
+      outcome.trialRunId = context.evidence.trial.runId
+      outcome.taskState = context.evidence.hostToolExecution.taskStateAfterObservation
+      outcome.usage = context.evidence.modelExecution.usage
+      outcome.costUsd = null
+      outcome.evidenceSceneId = context.evidence.id
+    }
+    catch (error) {
+      context && (context.evidence.status = 'failed')
+      if (context) context.evidence.failure = compactError(error)
+      outcome.status = 'failed'
+      outcome.failure = compactError(error)
+      outcome.finishedAt = new Date().toISOString()
+      if (context) {
+        outcome.usage = await captureRealOmpUsage(context)
+        await recoverFailedRealOmpTrial(context, outcome)
+      }
+    }
+  }
+}
+
+async function startApplication() {
+  currentStage = 'application-startup'
+  await assertFreeLoopback(OFFLINE_APP_MQTT_QUEUE_PORT, 'offline-app-mqtt-queue-probe-before-startup')
+  noteIsolation('offline-app-mqtt-queue-port-free-immediately-before-startup', true, {
+    port: OFFLINE_APP_MQTT_QUEUE_PORT,
+    host: '127.0.0.1',
+  })
+  appProcess = spawnLogged(process.execPath, [NUXT_CLI, 'dev', '--host', '127.0.0.1', '--port', String(AW_PORT), '--dotenv', EMPTY_DOTENV], {
+    cwd: REPO,
+    env: cleanChildEnv(),
+    logFd: APP_LOG_FD,
+    label: 'agentworkshop',
+  })
+  for (let i = 0; i < 90; i++) {
+    if (appProcess.exitCode !== null) throw new Error(`AgentWorkShop startup exited (code ${appProcess.exitCode})`)
+    try {
+      const response = await app('GET', '/api/health', undefined, { token: undefined, timeoutMs: 2_000 })
+      if (response.status === 200 && response.data?.status === 'ok' && response.data?.workshop === 'ready') {
+        if (response.data?.version !== '0.8.0') throw new Error(`server reported unexpected version ${safeText(response.data?.version, 80)}`)
+        noteIsolation('application-loopback-ready', true, { version: response.data.version, workshop: response.data.workshop, port: AW_PORT, host: '127.0.0.1' })
+        return response.data
+      }
+    }
+    catch (error) {
+      if (error?.message?.startsWith('server reported unexpected version')) throw error
+    }
+    await delay(500)
+  }
+  throw new Error('isolated AgentWorkShop startup timed out or workshop manager did not become ready')
+}
+
+async function main() {
+  mkdirSync(join(REPO, '.omc'), { recursive: true })
+  if (VERSION !== '0.8.0') throw new Error(`expected current package version 0.8.0; found ${VERSION}`)
+  await preflight()
+  prepareIsolatedHome()
+  if (HARNESS_MODE === 'pi-local-fixture') {
+    await startPiModelFixture()
+    noteIsolation('registered-pi-fixture-configured', true, {
+      command: PI_COMMAND,
+      provider: PI_PROVIDER,
+      model: PI_MODEL,
+      localModelApi: `127.0.0.1:${MODEL_API_PORT}`,
+      noExternalProvider: true,
+    })
+  }
+  else if (HARNESS_MODE === 'fixture') {
+    currentStage = 'compile-local-omp-fixture-shim'
+    buildHarnessShim()
+    noteIsolation('registered-omp-fixture-launcher-ready', true, { platform: process.platform, wrapper: 'local process shim; fixture is deterministic' })
+  }
+  else {
+    noteIsolation('registered-omp-real-provider-model-declared', true, {
+      command: 'omp from allowlisted PATH',
+      provider: REAL_OMP_PROVIDER,
+      model: REAL_OMP_MODEL,
+      cwd: 'fresh isolated AW_HOME',
+      authSource: 'local OMP auth store; value is not inspected or added to evidence',
+      discoveryDisabled: ['extensions', 'skills', 'rules'],
+      credentialValuesExposedToRunnerOrEvidence: false,
+    })
+  }
+  currentStage = 'simulator-startup'
+  await bootSimulator()
+  const health = await startApplication()
+  evidence.experiment.server = { version: health.version, mode: health.mode, workshop: health.workshop, uptimeMsAtReady: health.uptimeMs }
+  evidence.experiment.sourceIdentity = captureSourceIdentity()
+  const admin = await registerAdmin()
+  evidence.experiment.syntheticAdmin = admin
+  await assertPluginIsolation()
+  await verifyRagBridgeOutboundDisabled()
+  const daqStatus = requireSuccess(await app('GET', '/api/workshop/daq', undefined, { timeoutMs: 45_000 }), 'isolated DAQ backend status')
+  const settings = requireSuccess(await app('GET', '/api/system/settings'), 'effective runtime settings snapshot')
+  const effectiveMqttHost = settings.effective?.['daq.mqtt.host']
+  const effectiveMqttPort = Number(settings.effective?.['daq.mqtt.port'])
+  const effectiveMqttPortSource = settings.sources?.['daq.mqtt.port']
+  const effectiveInfrastructurePolicy = settings.effective?.['daq.startInfrastructure']
+  const effectiveInfrastructurePolicySource = settings.sources?.['daq.startInfrastructure']
+  const runtimeTargetMatches = effectiveMqttHost === '127.0.0.1'
+    && effectiveMqttPort === OFFLINE_APP_MQTT_QUEUE_PORT
+    && effectiveMqttPortSource === 'runtime'
+    && effectiveInfrastructurePolicy === 'never'
+    && effectiveInfrastructurePolicySource === 'runtime'
+  check('runtime-settings-target-offline-loopback-queue', runtimeTargetMatches, {
+    mqttHost: effectiveMqttHost ?? null,
+    mqttPort: effectiveMqttPort,
+    mqttPortSource: effectiveMqttPortSource ?? null,
+    infrastructurePolicy: effectiveInfrastructurePolicy ?? null,
+    infrastructurePolicySource: effectiveInfrastructurePolicySource ?? null,
+  })
+  check('isolated-inproc-queue-and-sqlite-tsdb', runtimeTargetMatches && daqStatus.meta?.queue === 'inproc' && daqStatus.meta?.tsdb === 'sqlite-emulated', {
+    queue: daqStatus.meta?.queue ?? null,
+    tsdb: daqStatus.meta?.tsdb ?? null,
+    effectiveMqttHost,
+    effectiveMqttPort,
+    effectiveMqttPortSource,
+  })
+  noteIsolation('application-message-path-is-inproc', runtimeTargetMatches && daqStatus.meta?.queue === 'inproc', {
+    queue: daqStatus.meta?.queue ?? null,
+    effectiveMqttHost,
+    effectiveMqttPort,
+    effectiveMqttPortSource: effectiveMqttPortSource ?? null,
+    simulatorMqttBrokerRemainsLoopback: `127.0.0.1:${MQTT_PORT}`,
+  })
+  evidence.isolation.result = 'passed'
+  noteIsolation('application-env-and-dotenv-isolation', true, {
+    launch: 'direct Nuxt CLI; dev-guard .env preload bypassed',
+    dotenvFile: EMPTY_DOTENV,
+    childEnvironmentAllowlisted: true,
+    storageEndpointsRestrictedToFreeLoopbackPorts: true,
+  })
+
+  const scenes = [
+    {
+      id: 'cast-film', preset: 'cast-film-physics', sourceDevice: 'dev-extruder-opcua',
+      mvSignal: 'ScrewSpeedSP', cvSignal: 'thick', baseline: 120, fixtureCandidate: 125,
+      processDescription: 'cast-film extrusion; screw speed affects measured film thickness',
+      targetMin: 48, targetMax: 52,
+      candidates: [125, 130, 135],
+    },
+    {
+      id: 'annealing', preset: 'anneal-line', sourceDevice: 'anneal-heating-mbtcp',
+      mvSignal: '均热区3炉温SP', cvSignal: 'StripTemp', baseline: 730, fixtureCandidate: 735,
+      processDescription: 'continuous annealing; zone-3 furnace setpoint affects measured strip temperature',
+      targetMin: 700, targetMax: 760,
+      candidates: [735, 740, 745],
+    },
+  ]
+  if (HARNESS_MODE === 'real-omp') {
+    check('real-omp-six-unique-trials-planned', scenes.length * REAL_OMP_TRIALS_PER_SCENE === 6 && new Set(scenes.flatMap(scene => Array.from({ length: REAL_OMP_TRIALS_PER_SCENE }, (_, index) => `${scene.id}-r${index + 1}`))).size === 6, {
+      scenes: scenes.map(scene => scene.id),
+      repetitionsPerScene: REAL_OMP_TRIALS_PER_SCENE,
+    })
+    for (const scene of scenes) await runRealOmpScene(scene)
+    const failedRuns = evidence.experiment.runOutcomes.filter(outcome => outcome.status !== 'verified')
+    evidence.status = failedRuns.length === 0 ? 'passed' : 'failed'
+    if (failedRuns.length) {
+      evidence.failure = {
+        stage: 'real-omp-trial-set',
+        name: 'RealOmpTrialSetError',
+        message: `${failedRuns.length} of ${evidence.experiment.runOutcomes.length} bounded real OMP trials failed`,
+        failedRunIds: failedRuns.map(outcome => outcome.id),
+      }
+      evidence.isolation.result = 'incomplete'
+    }
+  }
+  else {
+    for (const scene of scenes) await runScene({ ...scene, candidate: scene.fixtureCandidate })
+    evidence.status = 'passed'
+  }
+}
+
+async function cleanup() {
+  currentStage = 'cleanup'
+  const processResults = []
+  const processByChild = new Map()
+  if (appProcess) processByChild.set(appProcess, 'agentworkshop')
+  if (simProcess) processByChild.set(simProcess, 'simulator')
+  for (const child of children) if (!processByChild.has(child)) processByChild.set(child, child.label ?? 'child')
+  for (const [child, label] of [...processByChild.entries()].reverse()) {
+    try {
+      const stopped = await stopProcessTree(child, label)
+      processResults.push(stopped)
+      if (!evidence.checks.some(item => item.id === `${label}-process-stopped`)) {
+        evidence.checks.push({ id: `${label}-process-stopped`, status: 'fail', detail: { pid: child.pid ?? null, reason: 'process pid unavailable for termination verification' } })
+      }
+    }
+    catch (error) {
+      evidence.checks.push({ id: `${label}-process-stopped`, status: 'fail', detail: { pid: child.pid ?? null, error: compactError(error) } })
+      processResults.push(false)
+    }
+  }
+  if (piApiServer) {
+    try {
+      await new Promise((resolvePromise, reject) => {
+        piApiServer.close(error => error ? reject(error) : resolvePromise())
+      })
+      piApiServer = null
+      evidence.checks.push({ id: 'local-model-api-process-stopped', status: 'pass', detail: { host: '127.0.0.1', port: MODEL_API_PORT } })
+    }
+    catch (error) {
+      evidence.checks.push({ id: 'local-model-api-process-stopped', status: 'fail', detail: { error: compactError(error) } })
+      processResults.push(false)
+    }
+  }
+  const portResults = []
+  for (const [label, port] of [
+    ['app', AW_PORT],
+    ['simulator-http', SIM_PORT],
+    ['local-model-api', MODEL_API_PORT],
+    ['simulator-mqtt', MQTT_PORT],
+    ['offline-app-mqtt-queue-probe', OFFLINE_APP_MQTT_QUEUE_PORT],
+    ['offline-timescale-probe', OFFLINE_TSDB_PORT],
+    ['offline-objectstore-probe', OFFLINE_OBJECTSTORE_PORT],
+    ...allocatedDevicePorts.map((port, index) => [`simulator-device-${index}`, port]),
+  ]) {
+    try {
+      await assertFreeLoopback(port, label)
+      evidence.checks.push({ id: `${label}-port-released`, status: 'pass', detail: { port, host: '127.0.0.1' } })
+      portResults.push(true)
+    }
+    catch (error) {
+      evidence.checks.push({ id: `${label}-port-released`, status: 'fail', detail: { port, host: '127.0.0.1', error: compactError(error) } })
+      portResults.push(false)
+    }
+  }
+  try {
+    closeSync(SIM_LOG_FD)
+  }
+  catch { /* already closed */ }
+  try {
+    closeSync(APP_LOG_FD)
+  }
+  catch { /* already closed */ }
+  const requiredProcessLabels = evidence.status === 'passed' ? ['agentworkshop', 'simulator'] : [...processByChild.values()]
+  const verifiedProcessLabels = new Set(evidence.checks.filter(item => item.id.endsWith('-process-stopped') && item.status === 'pass').map(item => item.id.replace(/-process-stopped$/, '')))
+  const allRequiredProcessesVerified = requiredProcessLabels.every(label => verifiedProcessLabels.has(label))
+  return processResults.every(Boolean) && portResults.every(Boolean) && allRequiredProcessesVerified
+}
+
+try {
+  await main()
+}
+catch (error) {
+  const preflightFailed = currentStage === 'preflight' || currentStage === 'compile-local-omp-fixture-shim' || currentStage === 'simulator-startup'
+  evidence.status = preflightFailed ? 'blocked' : 'failed'
+  evidence.failure = { stage: currentStage, scene: currentScene ?? null, ...compactError(error) }
+  if (evidence.isolation.result === 'pending') evidence.isolation.result = preflightFailed ? 'failed' : 'incomplete'
+}
+finally {
+  const cleanupPassed = await cleanup()
+  if (!cleanupPassed) {
+    evidence.status = 'failed'
+    evidence.isolation.result = 'failed'
+    evidence.failure ??= { stage: 'cleanup', name: 'CleanupVerificationError', message: 'process stop or loopback port release verification failed' }
+    evidence.failure.cleanupVerificationFailed = true
+  }
+  if (evidence.status === 'passed') {
+    const failedChecks = [...evidence.checks, ...evidence.isolation.checks].filter(item => item.status !== 'pass')
+    const cleanupChecks = evidence.checks.filter(item => RELEASE_CHECK_SUFFIXES.some(suffix => item.id.endsWith(suffix)))
+    if (failedChecks.length || !cleanupChecks.length || cleanupChecks.some(item => item.status !== 'pass')) {
+      evidence.status = 'failed'
+      evidence.isolation.result = 'failed'
+      evidence.failure ??= { stage: 'final-verification', name: 'FinalVerificationError', message: 'not all experiment and cleanup checks passed' }
+    }
+  }
+  evidence.cleanup = { status: cleanupPassed ? 'passed' : 'failed', requiredProcessAndPortChecks: true }
+  evidence.experiment.finishedAt = new Date().toISOString()
+  evidence.experiment.evidencePath = 'evidence.json'
+  evidence.experiment.runtimeLogPaths = { app: 'agentworkshop.log', simulator: 'simulator.log' }
+  evidence.experiment.modelApiTrace = piApiTrace
+  const sanitizeArchiveText = content => content
+    .replaceAll(ROOT, '<isolated-run>')
+    .replaceAll(HOME, '<isolated-home>')
+    .replaceAll(REPO, '<workspace>')
+    .replaceAll(PI_COMMAND, 'pi.cmd')
+    .replace(/[A-Za-z0-9._%+-]+@example\.invalid/gi, '<synthetic-admin>')
+    .replace(/[A-Z]:\\Users\\[^\\\s]+/gi, '<user-home>')
+  try {
+    mkdirSync(OUTPUT_DIR, { recursive: false })
+    writeFileSync(PI_API_TRACE_PATH, `${JSON.stringify(piApiTrace, null, 2)}\n`, 'utf8')
+    const archiveName = OUTPUT_DIR.slice(OUTPUT_DIR.lastIndexOf(sep) + 1)
+    evidence.experiment.archiveDirectory = archiveName
+    const archiveFiles = ['evidence.json', 'agentworkshop.log', 'simulator.log', 'pi-model-api-trace.json', 'e2e-pi-simulator-paper.mjs', 'README.md']
+    evidence.experiment.archiveFiles = archiveFiles
+    writeFileSync(EVIDENCE_PATH, `${JSON.stringify(evidence, null, 2)}\n`, 'utf8')
+    const archivedEvidence = `${sanitizeArchiveText(JSON.stringify(evidence, null, 2))}\n`
+    writeFileSync(join(OUTPUT_DIR, 'evidence.json'), archivedEvidence, 'utf8')
+    writeFileSync(join(OUTPUT_DIR, 'agentworkshop.log'), sanitizeArchiveText(readFileSync(APP_LOG, 'utf8')), 'utf8')
+    writeFileSync(join(OUTPUT_DIR, 'simulator.log'), sanitizeArchiveText(readFileSync(SIM_LOG, 'utf8')), 'utf8')
+    copyFileSync(PI_API_TRACE_PATH, join(OUTPUT_DIR, 'pi-model-api-trace.json'))
+    copyFileSync(fileURLToPath(import.meta.url), join(OUTPUT_DIR, 'e2e-pi-simulator-paper.mjs'))
+    const readme = [
+      '# R23 registered Pi CLI simulator evidence',
+      '',
+      `Run status: **${evidence.status}**.`,
+      '',
+      'This archive was produced by a bounded local integration run of the registered Pi CLI adapter, its AgentWorkShop host-tool extension, the protected synthetic approval endpoint, and the loopback process simulator. The model endpoint is a deterministic OpenAI-compatible fixture bound to IPv4 loopback; no external model, provider, or service was called.',
+      '',
+      'The fixture is scripted to issue one bounded `recipe_trial`, pass its returned host-tool result back into Pi, issue one explicit `complete_task`, and pass that result back into Pi. Device acknowledgement, direct readback, run-tagged DAQ sample persistence, task/approval/receipt/sample linkage, and process/port cleanup are recorded in `evidence.json` when observed.',
+      '',
+      'The archived application and simulator logs contain runtime diagnostics with workspace and temporary-home paths and the synthetic admin email scrubbed. The Pi API trace contains request metadata and bounded tool-result text only; it omits prompts, credentials, session files, and the temporary isolated home.',
+      '',
+      evidence.failure ? `Recorded blocker/failure: ${safeText(evidence.failure.message, 500)} (stage: ${safeText(evidence.failure.stage, 120)}).` : 'No run blocker was recorded.',
+      '',
+    ].join('\n')
+    writeFileSync(join(OUTPUT_DIR, 'README.md'), readme, 'utf8')
+    writeFileSync(EVIDENCE_PATH, `${JSON.stringify(evidence, null, 2)}\n`, 'utf8')
+    writeFileSync(join(OUTPUT_DIR, 'evidence.json'), `${sanitizeArchiveText(JSON.stringify(evidence, null, 2))}\n`, 'utf8')
+  }
+  catch (error) {
+    evidence.status = 'failed'
+    evidence.failure ??= { stage: 'evidence-archive', ...compactError(error) }
+    writeFileSync(EVIDENCE_PATH, `${JSON.stringify(evidence, null, 2)}\n`, 'utf8')
+  }
+  console.log(`REGISTERED_HARNESS_E2E=${evidence.status}`)
+  console.log(`EVIDENCE=${join(OUTPUT_DIR, 'evidence.json')}`)
+  console.log(`STAGE=${evidence.failure?.stage ?? 'complete'}`)
+}
+
+if (evidence.status !== 'passed') process.exitCode = evidence.status === 'blocked' ? 2 : 1