agentworkshop 0.8.1 → 0.8.3

package/scripts/live-aml-optloop.mjs

@@ -1,588 +1,588 @@
-/**
- * LIVE AML 闭环优化全链路验收(bench 方法论;真实平台 + 真实 PLC 模拟器 + 真实训练)。
- *
- * 链路:
- *   隔离平台(3005,含本日构建) + 专用模拟器(4011 影子实例,cast-film 预设)
- *   → admin 注册 → provisionTwinLine(6 DCW 执行器 + 5 DAQ 传感器,真实协议链路)
- *   → 3 批次 × 330s 真实采样(DAQ 1s 节拍;val 批 ≥300 窗满足 Twin Gate G4)
- *   → REST 构建数据集(输入=全部 6 个 DCW;输出目标=膜厚 goal;特征=熔温/熔压)
- *   → hybrid_twin Channel(模板实例化 + worker 节点绑定)
- *   → invoke 桥(POST /agent-tools/invoke): twin_scene_discover → twin_scene_compile
- *     → twin_scene_freeze(用户确认令牌) → twin_physics_spec_draft(骨架物理模型)
- *     → twin_physics_spec_validate
- *   → REST 提交 hybrid_residual 训练(**无 code → 平台参考训练器**:物理参数校准 →
- *     3 成员残差集成 → conformal 覆盖率),轮询状态/进度/日志尾(透明)
- *   → 模型注册表校验(label/元数据/aml/models/<id> 工件) → REST 晋升 shadow
- *   → invoke: twin_snapshot_create(auto_daq) → twin_trial_run×10 → twin_gate_evaluate
- *     → mpc_optimize(model_id=训练模型驱动 rollout) → 读推荐
- *   → 透明度文档 bench/results/<rid>/OPTIMIZATION-LOG.md
- *
- * 运行: SIM_BASE=http://127.0.0.1:4011 node scripts/live-aml-optloop.mjs(需先 aw build)
- */
-import { mkdirSync, writeFileSync, appendFileSync, existsSync, readdirSync, statSync, readFileSync } from 'node:fs'
-import { spawn } from 'node:child_process'
-import { join, resolve, dirname } from 'node:path'
-import { fileURLToPath } from 'node:url'
-import { ensureSimulator, applyPreset, simNodes } from '../bench/lib/sim.mjs'
-import { provisionTwinLine, startTwinBatch } from '../bench/lib/closedloop.mjs'
-
-const REPO = resolve(dirname(fileURLToPath(import.meta.url)), '..')
-const AW_BASE = process.env.AW_BASE ?? 'http://127.0.0.1:3005'
-const AW_PORT = Number(new URL(AW_BASE).port)
-const RUN_ID = new Date().toISOString().replace(/[:.]/g, '-').slice(0, 19)
-const OUT_DIR = join(REPO, 'bench', 'results', `${RUN_ID}-aml-live-optloop`)
-mkdirSync(OUT_DIR, { recursive: true })
-const LOG_PATH = join(OUT_DIR, 'live-run.log')
-const DOC_PATH = join(OUT_DIR, 'OPTIMIZATION-LOG.md')
-const AML_ROOT = process.env.AW_AML_DIR_LIVE ?? join(REPO, 'aml')
-
-process.env.NO_PROXY = '127.0.0.1,localhost'
-process.env.no_proxy = '127.0.0.1,localhost'
-
-const L = []
-function log(line) {
-  const text = typeof line === 'string' ? line : JSON.stringify(line)
-  const stamped = `[${new Date().toISOString()}] ${text}`
-  console.log(stamped)
-  L.push(text)
-  appendFileSync(LOG_PATH, stamped + '\n')
-}
-function section(title) {
-  log('')
-  log('════════════════════════════════════════════════════════')
-  log(`  ${title}`)
-  log('════════════════════════════════════════════════════════')
-}
-const sleep = ms => new Promise(r => setTimeout(r, ms))
-
-async function api(method, path, body, token) {
-  const r = await fetch(AW_BASE + path, {
-    method,
-    headers: { 'content-type': 'application/json', ...(token ? { authorization: `Bearer ${token}` } : {}) },
-    body: body === undefined ? undefined : JSON.stringify(body),
-    signal: AbortSignal.timeout(120_000),
-  })
-  const json = await r.json().catch(() => null)
-  return { status: r.status, json }
-}
-
-async function invoke(tool, args, workerId, token) {
-  const r = await api('POST', '/api/workshop/agent-tools/invoke', { agentId: workerId, tool, args }, token)
-  const text = r.json?.data?.result?.text ?? r.json?.result?.text ?? JSON.stringify(r.json ?? {}).slice(0, 600)
-  log(`  [invoke] ${tool} → ${r.status}`)
-  return { status: r.status, text }
-}
-
-async function waitFor(predicate, { timeoutMs, everyMs = 4000, what }) {
-  const t0 = Date.now()
-  while (Date.now() - t0 < timeoutMs) {
-    const v = await predicate()
-    if (v) return v
-    await sleep(everyMs)
-  }
-  throw new Error(`等待超时: ${what}(${timeoutMs}ms)`)
-}
-
-/* ════════ Stage 0 · 隔离平台 + 专用模拟器 ════════ */
-section('Stage 0 · 隔离平台(3005) + 专用模拟器(4011)')
-
-async function platformUp(base) {
-  try {
-    const r = await fetch(`${base}/api/users/setup-status`, { signal: AbortSignal.timeout(5000) })
-    const j = await r.json().catch(() => null)
-    return r.status === 200 && j?.data != null
-  }
-  catch { return false }
-}
-
-if (!(await platformUp(AW_BASE))) {
-  const child = spawn(process.execPath, [join(REPO, 'bin', 'aw.mjs'), 'start', '--port', String(AW_PORT)], {
-    cwd: REPO,
-    env: { ...process.env, AW_AML_DIR: AML_ROOT, PORT: String(AW_PORT) },
-    detached: true,
-    stdio: 'ignore',
-  })
-  child.unref()
-  log(`平台冷启: port=${AW_PORT} pid=${child.pid}(repo 配置根)`)
-  log(`  AW_AML_DIR=${AML_ROOT}(复用真实训练 venv;模型工件落 repo ./aml/models)`)
-}
-await waitFor(() => platformUp(AW_BASE), { timeoutMs: 180_000, what: '平台就绪' })
-log(`✔ 平台在线 ${AW_BASE}`)
-const sim = await ensureSimulator({ log: m => log(`[sim] ${m}`) })
-log(`✔ 模拟器在线: ${JSON.stringify(sim).slice(0, 120)}`)
-await applyPreset('cast-film-physics')
-const simDevices = await simNodes()
-log(`✔ cast-film 预设应用,设备 ${simDevices.length} 台: ${simDevices.map(d => `${d.id}(${d.protocol})`).join(', ')}`)
-
-/* ════════ Stage 1 · admin 注册 ════════ */
-section('Stage 1 · 用户注册 + 提权(repo 库既有用户体系,新注册=user;本链路需 admin 建 DCW/产线)')
-let token = String(process.env.AW_TOKEN ?? '').trim()
-if (token) {
-  log('✔ 复用调用方 admin token（隔离实例已初始化，不再注册新账号）')
-}
-else {
-  const reg = await api('POST', '/api/workshop/users/register', { name: `aml-live-${RUN_ID}` })
-  token = reg.json?.data?.token
-  if (!token) throw new Error(`注册失败: ${JSON.stringify(reg.json).slice(0, 200)}`)
-  log('✔ token 就绪')
-  // 提升 role → admin（repo 库首次注册为 user 时的一次性测试授权）
-  {
-    const { DatabaseSync } = await import('node:sqlite')
-    const usersDb = new DatabaseSync(join(REPO, '.AgentWorkShop', 'data', 'users.sqlite'))
-    const r = usersDb.prepare(`UPDATE users SET role = 'admin' WHERE email LIKE 'aml-live-${RUN_ID}%' AND role = 'user'`).run()
-    usersDb.close()
-    log(`✔ 提权 admin: ${Number(r.changes)} 行`)
-  }
-}
-
-/* ════════ Stage 2 · 产线 provision + 开跑（或复用已完成采样的线/数据集）════════ */
-section('Stage 2 · 产线 provisioning(6 DCW 执行器 + 5 DAQ 传感器,驱动配置取自模拟器 /export)')
-const sfx = RUN_ID.slice(5, 16).replace(/-/g, '')
-const REUSE_LINE = String(process.env.AML_REUSE_LINE ?? '').trim()
-let twin
-let batch1
-let dataset
-let spReadback
-if (REUSE_LINE) {
-  twin = {
-    lineId: REUSE_LINE,
-    productId: String(process.env.AML_REUSE_PRODUCT ?? ''),
-    dcw: JSON.parse(String(process.env.AML_REUSE_DCW ?? '{}')),
-    daq: JSON.parse(String(process.env.AML_REUSE_DAQ ?? '{}')),
-  }
-  batch1 = { recipeId: String(process.env.AML_REUSE_RECIPE ?? '') }
-  dataset = {
-    id: String(process.env.AML_REUSE_DATASET ?? ''),
-    rowCount: Number(process.env.AML_REUSE_ROWS ?? 0),
-    runIds: [1, 2, 3],
-    sha256: String(process.env.AML_REUSE_SHA ?? ''),
-  }
-  spReadback = JSON.parse(String(process.env.AML_REUSE_SP ?? '{}'))
-  log(`✔ 复用采样结果 line=${twin.lineId} recipe=${batch1.recipeId} dataset=${dataset.id} (${dataset.rowCount} 窗)`)
-}
-else {
-  twin = await provisionTwinLine({ call: async (m, p, b) => { const r = await api(m, p, b, token); return { status: r.status, ...(r.json ?? {}) } } }, { simDevices, sfx })
-  for (const e of twin.ev) log('  ' + e)
-  if (!twin.ok) throw new Error(`provision 失败: ${(twin.errors ?? []).join('; ')} | ev=${(twin.ev ?? []).join(' ; ')}`)
-  log(`✔ lineId=${twin.lineId} productId=${twin.productId}`)
-  log(`  DCW: ${JSON.stringify(twin.dcw)}`)
-  log(`  DAQ: ${JSON.stringify(twin.daq)}`)
-
-  batch1 = await startTwinBatch({ call: async (m, p, b) => { const r = await api(m, p, b, token); return { status: r.status, ...(r.json ?? {}) } } }, { lineId: twin.lineId, productId: twin.productId, dcw: twin.dcw, sfx })
-  if (!batch1.ok) throw new Error(`开跑失败: ${batch1.errors?.join('; ')}`)
-  log(`✔ 批次开跑 recipeId=${batch1.recipeId}`)
-
-  // ── SP 设定点回读 DAQ 节点:AML 数据集的 control 输入取自数采时序库,
-  //    执行器设定点必须有 DAQ 回读节点(同一信号,只读)才能进入训练。──
-  const { simExport } = await import('../bench/lib/sim.mjs')
-  spReadback = {}
-  const spSignalIds = Object.keys(twin.dcw) // zone1-sp / screw-sp / linespeed-sp / diegap-sp …
-  for (const dev of simDevices) {
-    const exp = await simExport(dev.id)
-    const items = exp?.items ?? []
-    for (const spId of spSignalIds) {
-      if (spReadback[spId]) continue
-      const sig = (dev.signals ?? []).find(s => s.id === spId)
-      if (!sig) continue
-      // export 的 item.signal = 信号【名】(如「加热区1SP」),按名对齐
-      const item = items.find(i => i.signal === sig.name)
-      if (!item?.driverConfig) continue
-      // 写配置 → 读配置转换:mqtt 写用 jsonKey(发布字段),DAQ 读订阅同主题用 jsonPath 取值
-      let cfg = { ...item.driverConfig }
-      if (cfg.jsonKey !== undefined && cfg.jsonPath === undefined) {
-        cfg.jsonPath = cfg.jsonKey
-        delete cfg.jsonKey
-      }
-      const r = await api('POST', '/api/workshop/daq', {
-        name: `${sig.name} 回读 ${sfx}`, templateRef: 'daq-temp-tc', driver: dev.protocol, driverConfig: cfg,
-        unit: sig.unit, min: sig.min, max: sig.max, lineId: twin.lineId, intervalMs: 1000, publishIntervalMs: 0,
-        semantics: `执行器设定点回读(AML control 输入) ${item.signal}`,
-      }, token)
-      if (r.json?.data?.node?.id) spReadback[spId] = r.json.data.node.id
-    }
-  }
-  log(`✔ SP 回读 DAQ 节点 ${Object.keys(spReadback).length}/6: ${JSON.stringify(spReadback)}`)
-
-  /* ════════ Stage 3 · 3 批次采样(每批 500s;批内两段安全阶跃激励) ════════ */
-  section('Stage 3 · 真实采样:3 批次 × 500s(批内阶跃激励:每个切分都覆盖 3 个设定点水平)')
-  // 实验设计:设定点只在批次间变化会导致 train/test 分布割裂(外推爆炸,G1/G3 必挂)。
-  // 批内阶跃:每个 run 依次经历 base → +Δ1 → +Δ2 三个水平,train/val/test 同分布;
-  // 各 run 的 Δ 略作变化增加丰富度,全部在配方窗口与安全上限内。
-  const BASE_SP = { 'linespeed-sp': 95, 'diegap-sp': 1.0 }
-  const EXCITE_RUNS = [
-    [{ ls: 2.0, dg: 0.03 }, { ls: -1.5, dg: 0.06 }],
-    [{ ls: -1.0, dg: 0.05 }, { ls: 2.5, dg: 0.02 }],
-    [{ ls: 1.5, dg: 0.04 }, { ls: -2.0, dg: 0.05 }],
-  ]
-  for (let run = 1; run <= 3; run++) {
-    if (run > 1) {
-    // stop 是异步收窗：立即 start 会撞「已在运行」。必须轮询到 active=false 再开。
-      for (let attempt = 0; attempt < 12; attempt++) {
-        await api('POST', `/api/workshop/dcw/lines/${twin.lineId}/stop`, {}, token)
-        await sleep(1000)
-        const snap = await api('GET', '/api/workshop/dcw', undefined, token)
-        const state = (snap.json?.data?.lineStates ?? []).find(x => x.lineId === twin.lineId)
-        if (!state?.active) break
-        if (attempt === 11) throw new Error(`批次 ${run} 前停线超时: ${JSON.stringify(state)}`)
-      }
-      log('  上一批次停止(active=false)')
-      // 复用同一配方 stop/start → 新 runId(数据集按 recipeId 取批次,须同配方)
-      let started = null
-      for (let attempt = 0; attempt < 8; attempt++) {
-        started = await api('POST', `/api/workshop/dcw/lines/${twin.lineId}/start`, { recipeId: batch1.recipeId }, token)
-        if (started.status === 200) break
-        if (!/已在运行/.test(String(started.json?.message ?? ''))) break
-        await sleep(1000)
-      }
-      if (started?.status !== 200) throw new Error(`批次 ${run} 开跑失败: ${JSON.stringify(started?.json).slice(0, 200)}`)
-      log(`  批次 ${run} 开跑(同配方 ${batch1.recipeId})`)
-    }
-    await sleep(20_000) // base 水平稳定（基准压缩：3×240s，仍≥720 行 1s 样本）
-    const [e1, e2] = EXCITE_RUNS[run - 1]
-    await api('POST', `/api/workshop/dcw/${twin.dcw['linespeed-sp']}/write`, { value: BASE_SP['linespeed-sp'] + e1.ls }, token)
-    await api('POST', `/api/workshop/dcw/${twin.dcw['diegap-sp']}/write`, { value: BASE_SP['diegap-sp'] + e1.dg }, token)
-    log(`  批次${run} 阶跃1: linespeed=${BASE_SP['linespeed-sp'] + e1.ls} diegap=${(BASE_SP['diegap-sp'] + e1.dg).toFixed(2)}`)
-    await sleep(80_000)
-    await api('POST', `/api/workshop/dcw/${twin.dcw['linespeed-sp']}/write`, { value: BASE_SP['linespeed-sp'] + e2.ls }, token)
-    await api('POST', `/api/workshop/dcw/${twin.dcw['diegap-sp']}/write`, { value: BASE_SP['diegap-sp'] + e2.dg }, token)
-    log(`  批次${run} 阶跃2: linespeed=${BASE_SP['linespeed-sp'] + e2.ls} diegap=${(BASE_SP['diegap-sp'] + e2.dg).toFixed(2)}`)
-    await sleep(140_000) // 第三水平保持至 240s
-    const pv = twin.daq['film-thickness']
-    const r = await api('GET', `/api/workshop/daq/${pv}/samples?bucketMs=1000&limit=3`, undefined, token)
-    const latest = r.json?.data?.points?.[0]?.avg ?? '?'
-    log(`  批次${run} 完成(500s) · 膜厚当前≈${typeof latest === 'number' ? latest.toFixed(2) : latest} μm`)
-  }
-  await api('POST', `/api/workshop/dcw/lines/${twin.lineId}/stop`, {}, token)
-  log('✔ 采样完成(3 批次 × 330s),产线停止')
-
-  /* ════════ Stage 4 · 构建数据集(输入=SP 回读 DAQ,目标=膜厚) ════════ */
-  section('Stage 4 · AML 数据集构建(输入=6 个 SP 回读 DAQ;输出目标=膜厚 goal;特征=熔温/熔压)')
-  if (Object.keys(spReadback).length < 6) throw new Error(`SP 回读节点不足: ${JSON.stringify(spReadback)}`)
-  // 采样预检:逐节点确认批次期间有样本(无样本的回读剔除并记录,防 0 窗数据集)
-  const sampledControls = []
-  for (const [sig, nodeId] of Object.entries(spReadback)) {
-    const r = await api('GET', `/api/workshop/daq/${nodeId}/samples?bucketMs=60000&limit=1`, undefined, token)
-    const has = (r.json?.data?.points ?? []).length > 0
-    log(`  回读采样检查 ${sig}(${nodeId}): ${has ? '✓ 有样本' : '✘ 无样本,剔除'}`)
-    if (has) sampledControls.push({ nodeId, role: 'control' })
-  }
-  if (sampledControls.length < 4) throw new Error(`可用 control 回读不足: ${sampledControls.length}`)
-  const dsBody = {
-    lineId: twin.lineId, productId: twin.productId, recipeId: batch1.recipeId,
-    nodes: [
-      ...sampledControls,
-      { nodeId: twin.daq['film-thickness'], role: 'target' },
-      { nodeId: twin.daq['melt-temp'], role: 'feature' },
-      { nodeId: twin.daq['melt-pressure'], role: 'feature' },
-    ],
-    beatMs: 1000, window: { historySteps: 8, horizonSteps: 4 },
-    split: { valRatio: 0.34, testRatio: 0.33, seed: 42 },
-    purpose: 'mpc_surrogate',
-    note: `live 闭环优化数据集 ${RUN_ID}`,
-  }
-  const ds = await api('POST', '/api/workshop/aml/datasets', dsBody, token)
-  dataset = ds.json?.data?.dataset
-  if (!dataset) throw new Error(`数据集构建失败: ${JSON.stringify(ds.json).slice(0, 300)}`)
-  log(`✔ 数据集 ${dataset.id}: ${dataset.rowCount} 窗 / ${dataset.runIds.length} 批 / sha=${String(dataset.sha256).slice(0, 12)}`)
-}
-
-/* ════════ Stage 5 · hybrid_twin Channel 实例化 + worker 绑定 ════════ */
-section('Stage 5 · Hybrid Twin Channel(模板实例化 + worker 节点绑定)')
-const templates = await api('GET', '/api/workshop/channel-templates', undefined, token)
-const tpl = templates.json?.data?.find(x => x.id === 'chtpl-hybrid-twin-mpc-default')
-if (!tpl) throw new Error('hybrid 模板缺失')
-const SCENE_ID = `castfilm-hold-opt-${sfx}`
-const inst = await api('POST', `/api/workshop/channel-templates/${tpl.id}/instantiate`, {
-  name: `AML Live OptLoop ${RUN_ID}`,
-  toolProfile: 'hybrid_twin',
-  scene: { sceneId: SCENE_ID, sceneVersion: '1.0.0', lineId: twin.lineId, productId: twin.productId, recipeId: batch1.recipeId },
-  objective: { objectiveId: 'thickness-50um', targets: { film_thickness: 50 } },
-  controlPolicy: 'recommendation_only',
-}, token)
-const channelId = inst.json?.data?.channelId
-const worker = inst.json?.data?.agents?.find(x => x.role === 'worker')
-if (!channelId || !worker) throw new Error(`实例化失败: ${JSON.stringify(inst.json).slice(0, 300)}`)
-log(`✔ channelId=${channelId} worker=${worker.id}`)
-// 权限模型 v2：dcw 节点直绑已废弃；hybrid 场景的控制面 = 绑定 recipe（参数节点展开为 DCW）。
-const recipeBind = await api('POST', '/api/workshop/agent-tools/bindings', { agentId: worker.id, nodeId: batch1.recipeId, kind: 'recipe', mode: 'auto' }, token)
-if (recipeBind.status !== 200) throw new Error(`配方绑定失败: ${JSON.stringify(recipeBind.json).slice(0, 200)}`)
-for (const nodeId of Object.values(twin.daq)) {
-  await api('POST', '/api/workshop/agent-tools/bindings', { agentId: worker.id, nodeId, kind: 'daq', mode: 'auto' }, token)
-}
-log(`✔ worker 绑定 recipe(${Object.keys(twin.dcw).length} 参数) + ${Object.keys(twin.daq).length} DAQ`)
-
-/* ════════ Stage 6 · 场景发现→编译→冻结;骨架 PhysicsSpec ════════ */
-section('Stage 6 · 场景发现→编译→冻结;骨架 PhysicsSpec(twin_physics_spec_draft)')
-// 膜厚节点语义补『厚度』关键词(inferNodeRole 目标词表;信号名『膜厚』缺『度』不命中)→
-// 不补语义膜厚会回退成 state,契约里没有 target,骨架物理模型就没有优化输出。
-await api('PATCH', `/api/workshop/daq/${twin.daq['film-thickness']}`, { semantics: '流延膜厚度测量(优化目标 goal,单位 μm;厚度质量输出)' }, token)
-const disco = await invoke('twin_scene_discover', {}, worker.id, token)
-log(disco.text.split('\n').slice(0, 14).map(x => '  ' + x).join('\n'))
-log('  膜厚目标识别: ' + (disco.text.includes('"target": 1') || disco.text.includes('"target":1') ? '✔ target=1' : '✘ 仍是 0(检查语义)'))
-const compiled = await invoke('twin_scene_compile', {
-  scene_id: SCENE_ID, scene_version: '1.0.0', line_id: twin.lineId, product_id: twin.productId, recipe_id: batch1.recipeId,
-  prompt: '流延膜厚度闭环优化:目标膜厚 50μm。控制=6 个执行器设定点,观测=膜厚(目标)/熔温/熔压。约束:熔温 195~225℃,熔压 ≤22MPa。膜厚为目标输出。',
-}, worker.id, token)
-log(compiled.text.split('\n').slice(0, 12).map(x => '  ' + x).join('\n'))
-const frozen = await invoke('twin_scene_freeze', {
-  scene_id: SCENE_ID, scene_version: '1.0.0', confirmation: 'USER_CONFIRMED_SCENE_CONTRACT', approved_by: 'live-admin',
-}, worker.id, token)
-log(frozen.text.split('\n').slice(0, 8).map(x => '  ' + x).join('\n'))
-
-// 冻结契约全文 = 孪生工件目录 scenes/<id>-<ver>-frozen/scenes.json(平台落盘,本地可读;
-// 结构 { scene, contractHash, status, approvedBy, approvedAt },契约在 .scene 下)
-const frozenArtifactPath = join(AML_ROOT, 'twins', 'scenes', `${SCENE_ID}-1.0.0-frozen`, 'scenes.json')
-if (!existsSync(frozenArtifactPath)) throw new Error(`冻结场景工件缺失: ${frozenArtifactPath}`)
-const frozenArtifact = JSON.parse(readFileSync(frozenArtifactPath, 'utf8'))
-const frozenScene = frozenArtifact.scene ?? frozenArtifact
-const sceneControls = frozenScene.controls ?? []
-log(`✔ 冻结契约: controls=${sceneControls.length} targets=${(frozenScene.observations ?? []).filter(v => v.role === 'target').length} 约束=${(frozenScene.constraints ?? []).length}`)
-
-const draft = await invoke('twin_physics_spec_draft', { dt_sec: 1 }, worker.id, token)
-log(draft.text.split('\n').slice(0, 6).map(x => '  ' + x).join('\n'))
-// 工具输出为摘要;spec 全文在落盘工件(aml/twins/physics-spec-draft/<id>/physics-spec-draft.json)
-const draftArtifactPath = (draft.text.match(/artifact:\s*([^\n]+)/) || [])[1]?.trim()
-if (!draftArtifactPath || !existsSync(draftArtifactPath)) throw new Error(`骨架 spec 工件缺失: ${draftArtifactPath}`)
-const physicsSpec = JSON.parse(readFileSync(draftArtifactPath, 'utf8'))
-log(`✔ 骨架 PhysicsSpec: model=${physicsSpec.modelId} 状态方程 ${physicsSpec.states.length} 观测 ${physicsSpec.observations.length} 可校准参数 ${physicsSpec.parameters.filter(p => p.min != null && p.max != null).length}/${physicsSpec.parameters.length}`)
-writeFileSync(join(OUT_DIR, 'physics-spec.json'), JSON.stringify(physicsSpec, null, 2))
-const validated = await invoke('twin_physics_spec_validate', { physics_spec: physicsSpec }, worker.id, token)
-log(validated.text.split('\n').slice(0, 5).map(x => '  ' + x).join('\n'))
-
-// ── 训练前对齐(关键):场景控制绑定的是 DCW 执行器,而数据集 control 列是 SP 回读
-//    DAQ 节点 → spec 控制变量必须改绑回读节点;且 spec 只能引用数据集包含的节点
-//    (场景可比数据集宽,宽出的变量取不到值 → 物理 NaN fail-closed)。──
-const datasetManifest = JSON.parse(readFileSync(join(AML_ROOT, 'datasets', dataset.id, 'manifest.json'), 'utf8'))
-const datasetNodes = new Set(datasetManifest.allNodes)
-const dwToDn = {}
-for (const [sig, dn] of Object.entries(spReadback)) dwToDn[twin.dcw[sig]] = dn
-let remapped = 0
-for (const v of physicsSpec.variables) {
-  if (v.nodeId && !datasetNodes.has(v.nodeId) && dwToDn[v.nodeId]) { v.nodeId = dwToDn[v.nodeId]; remapped++ }
-}
-const keptVars = physicsSpec.variables.filter(v => !v.nodeId || datasetNodes.has(v.nodeId))
-const keptIds = new Set(keptVars.map(v => v.id))
-const dropped = physicsSpec.variables.filter(v => !keptIds.has(v.id)).map(v => `${v.id}(${v.nodeId})`)
-physicsSpec.variables = keptVars
-physicsSpec.states = physicsSpec.states.filter((e) => { const base = e.lhs.endsWith('_next') ? e.lhs.slice(0, -5) : e.lhs; return keptIds.has(base) })
-physicsSpec.observations = physicsSpec.observations.filter(e => keptIds.has(e.lhs))
-physicsSpec.guards = (physicsSpec.guards ?? []).filter(e => keptIds.has(e.lhs))
-physicsSpec.constraints = physicsSpec.constraints.filter(c => keptIds.has(c.id))
-writeFileSync(join(OUT_DIR, 'physics-spec-remapped.json'), JSON.stringify(physicsSpec, null, 2))
-log(`✔ 训练对齐: 控制改绑 ${remapped};数据集裁剪剔除 ${dropped.length} 变量(${dropped.join(', ') || '无'});保留 ${keptVars.length}`)
-const revalidated = await invoke('twin_physics_spec_validate', { physics_spec: physicsSpec }, worker.id, token)
-log('  再校验: ' + revalidated.text.split('\n').filter(x => x.includes('"valid"')).join(' ').slice(0, 60))
-
-/* ════════ Stage 7 · hybrid 训练(平台参考训练器,无 code) ════════ */
-section('Stage 7 · 提交 hybrid_residual 训练(无 code → 平台参考训练器;物理校准+3 成员残差集成+UQ)')
-const jobRes = await api('POST', '/api/workshop/aml/jobs', {
-  datasetId: dataset.id,
-  purpose: 'mpc_surrogate',
-  changeNote: 'live 闭环优化:骨架物理+数据残差(全部 DCW → 膜厚 goal)',
-  params: { epochs: 220, hidden: 96, lr: 0.002, residual_scale: 1.2, ensemble: 3 },
-  seed: 42,
-  jobKind: 'hybrid_residual',
-  sceneId: SCENE_ID,
-  sceneVersion: '1.0.0',
-  objectiveId: 'thickness-50um',
-  physicsSpec,
-  providerId: physicsSpec.modelId,
-  providerVersion: '1.0.0',
-  providerHash: `sha256:${JSON.stringify(physicsSpec).length}`,
-  modelName: `流延膜厚闭环模型 ${sfx}`,
-  modelDescription: `输入=全部 6 个 DCW 执行器;输出=膜厚 goal(50μm);场景 ${SCENE_ID}@1.0.0;产线 ${twin.lineId}/配方 ${batch1.recipeId}`,
-}, token)
-const job = jobRes.json?.data?.job
-if (!job) throw new Error(`作业提交失败: ${JSON.stringify(jobRes.json).slice(0, 300)}`)
-log(`✔ 作业 ${job.id} 入队(jobKind=${job.budget?.jobKind}, scene=${job.budget?.sceneId})`)
-log(`  模型标识 name: ${job.budget?.modelName}`)
-log(`  描述: ${job.budget?.modelDescription}`)
-
-const jobLogsPath = join(OUT_DIR, 'training-run.log')
-let jobDone = null
-for (let i = 0; i < 150; i++) {
-  await sleep(10_000)
-  const st = await api('GET', `/api/workshop/aml/jobs/${job.id}`, undefined, token)
-  const j = st.json?.data?.job ?? st.json?.data
-  if (!j) continue
-  log(`  [${i * 10}s] ${j.status} ${j.stage} ${j.progress}%`)
-  const logFile = join(AML_ROOT, 'jobs', job.id, 'run.log')
-  if (existsSync(logFile)) {
-    const lines = readFileSync(logFile, 'utf8').split('\n').filter(Boolean)
-    writeFileSync(jobLogsPath, lines.join('\n'))
-    for (const t of lines.filter(l => l.includes('##AML')).slice(-3)) log(`    ${t.slice(0, 170)}`)
-  }
-  if (['done', 'failed', 'cancelled', 'timeout'].includes(j.status)) { jobDone = j; break }
-}
-if (!jobDone || jobDone.status !== 'done') throw new Error(`训练未成功: ${JSON.stringify(jobDone).slice(0, 300)}`)
-const gatesPassed = (() => {
-  try { return JSON.parse(jobDone.gatesJson || '{}').passed }
-  catch { return false }
-})()
-log(`✔ 训练完成 · 平台门禁 G1-G5: ${gatesPassed ? '全部通过' : '未过'}`)
-log(`  训练日志全文 → ${jobLogsPath}`)
-
-/* ════════ Stage 8 · 模型保存校验 + 晋升 shadow ════════ */
-section('Stage 8 · 模型保存校验(aml/models/<id> 工件 + label 元数据)+ 晋升 shadow')
-const modelsRes = await api('GET', '/api/workshop/aml/models', undefined, token)
-const model = (modelsRes.json?.data?.models ?? []).find(m => m.datasetId === dataset.id)
-if (!model) throw new Error('模型未登记')
-log(`✔ 模型 ${model.id} [${model.stage}]`)
-log(`  label: ${model.label}`)
-log(`  description: ${model.description}`)
-log(`  line=${model.lineId} recipe=${model.recipeId} purpose=${model.purpose} objective=${model.objectiveId}`)
-const modelDir = model.path
-if (existsSync(modelDir)) {
-  log(`  工件目录 ${modelDir}:`)
-  for (const f of readdirSync(modelDir)) log(`    - ${f}(${statSync(join(modelDir, f)).size}B)`)
-}
-else log(`  ⚠ 模型目录(服务器进程视角): ${modelDir}`)
-const metrics = model.metrics ?? {}
-log(`  G1 单步 NRMSE=${metrics.oneStepTest?.nrmse} · G2 滚动=${metrics.rolloutTest?.nrmse}`)
-log(`  hybrid: ${JSON.stringify(metrics.hybrid ?? {}).slice(0, 240)}`)
-log(`  uncertainty: ${JSON.stringify(metrics.uncertainty ?? {}).slice(0, 240)}`)
-log(`  physics: ${JSON.stringify(metrics.physics ?? {}).slice(0, 240)}`)
-const promo = await api('POST', `/api/workshop/aml/models/${model.id}/promote`, { toStage: 'shadow' }, token)
-log(`✔ 晋升 shadow: HTTP ${promo.status}`)
-
-/* ════════ Stage 9 · 孪生闭环:快照 → 10 试验 → 门禁 → MPC 推荐 ════════ */
-section('Stage 9 · 孪生闭环(auto_daq 快照 → 10 候选试验 → Twin Gate → MPC 推荐,训练模型驱动)')
-await api('POST', `/api/workshop/dcw/lines/${twin.lineId}/start`, { recipeId: batch1.recipeId }, token)
-log('  产线重启(真实工况下采快照)')
-await sleep(25_000)
-const snap = await invoke('twin_snapshot_create', { auto_daq: true, channel_id: channelId, phase: 'holding' }, worker.id, token)
-log(snap.text.split('\n').slice(0, 8).map(x => '  ' + x).join('\n'))
-const snapshotId = (snap.text.match(/snapshot_id:\s*([^\n]+)/) || [])[1]?.trim()
-if (!snapshotId) throw new Error('快照创建失败: ' + snap.text.slice(0, 300))
-
-// 候选轨迹:线速度小幅变化(量程 20~120,maxStep 由场景 writePolicy 派生)
-const speedCtl = sceneControls.find(c => (c.nodeId ?? '').includes('linespeed')) ?? sceneControls[0]
-const speedStep = speedCtl?.maxStep ?? 2
-for (let k = 0; k < 10; k++) {
-  const delta = (k % 2 === 0 ? 1 : -0.5) * speedStep * (0.3 + 0.07 * k)
-  const t = await invoke('twin_trial_run', {
-    snapshot_id: snapshotId,
-    baseline_controls: { [speedCtl.id]: 95 },
-    candidate_controls: [
-      { [speedCtl.id]: 95 + delta }, { [speedCtl.id]: 95 + delta },
-      { [speedCtl.id]: 95 + delta }, { [speedCtl.id]: 95 + delta },
-    ],
-  }, worker.id, token)
-  log(`  trial#${k + 1} Δlinespeed=${delta.toFixed(2)} → ${t.text.split('\n').filter(x => x.includes('constraints_passed') || x.includes('improvement')).join(' | ').slice(0, 160)}`)
-}
-const gate = await invoke('twin_gate_evaluate', { model_id: model.id }, worker.id, token)
-log('  Twin Gate(12 判据,回写 twinEligibility):')
-log(gate.text.split('\n').slice(2, 26).map(x => '    ' + x).join('\n'))
-const mpc = await invoke('mpc_optimize', {
-  snapshot_id: snapshotId,
-  model_id: model.id,
-  baseline_controls: { [speedCtl.id]: 95 },
-  horizon_steps: 4,
-}, worker.id, token)
-writeFileSync(join(OUT_DIR, 'mpc-recommendation.txt'), mpc.text)
-const mode = (mpc.text.match(/"mode":\s*"([^"]+)"/) || [])[1]
-const bestCandidate = (mpc.text.match(/"bestCandidate":\s*(\{[^}]*\})/) || [])[1]
-const bestCost = (mpc.text.match(/"bestCost":\s*([\d.eE+-]+)/) || [])[1]
-log(`✔ MPC 完成: mode=${mode} bestCandidate=${bestCandidate} bestCost=${bestCost}`)
-log(`  完整推荐 → ${join(OUT_DIR, 'mpc-recommendation.txt')}`)
-
-/* ════════ Stage 9b · 闭环执行:推荐设定点 → admin 写入(HITL 授权模拟)→ 复测 ════════ */
-section('Stage 9b · 闭环执行:把 MPC 推荐的设定点写入产线(admin 执行),等待响应后复测')
-let best
-try {
-  best = JSON.parse(bestCandidate ?? 'null')
-}
-catch {
-  best = null
-}
-if (best && Object.keys(best).length > 0) {
-  // spec 控制 id → 场景控制 nodeId(DCW 节点)
-  const idToNode = Object.fromEntries(sceneControls.map(c => [c.id, c.nodeId]))
-  const writes = []
-  for (const [specId, value] of Object.entries(best)) {
-    const nodeId = idToNode[specId]
-    if (!nodeId || !Number.isFinite(Number(value))) continue
-    const w = await api('POST', `/api/workshop/dcw/${nodeId}/write`, { value }, token)
-    writes.push({ specId, nodeId, value, status: w.status })
-    log(`  写入 ${nodeId} ← ${Number(value).toFixed(2)} → HTTP ${w.status}`)
-  }
-  log(`  等待 90s 工艺响应(settling)…`)
-  await sleep(90_000)
-  const pv = twin.daq['film-thickness']
-  const r = await api('GET', `/api/workshop/daq/${pv}/samples?bucketMs=1000&limit=60`, undefined, token)
-  const pts = (r.json?.data?.points ?? []).map(p => Number(p.avg ?? p.value)).filter(Number.isFinite).slice(0, 30)
-  const mean = pts.length ? pts.reduce((a, b) => a + b, 0) / pts.length : Number.NaN
-  log(`✔ 执行后膜厚均值(最近 30s): ${mean.toFixed(2)} μm(目标 50μm;写入前 ≈54 μm)`)
-  writeFileSync(join(OUT_DIR, 'closed-loop-write.json'), JSON.stringify({ writes, thicknessAfter: mean }, null, 2))
-}
-else {
-  log('  ⚠ MPC 未给出可执行的 bestCandidate(如 safe_small_step 阶段)——按平台治理,此阶段只推荐不写。')
-}
-
-/* ════════ Stage 10 · 透明度文档 ════════ */
-section('Stage 10 · 生成透明度文档')
-const doc = [
-  `# AML 混合孪生闭环优化 · 实机透明度记录`,
-  ``,
-  `- 运行标识: ${RUN_ID}`,
-  `- 平台: ${AW_BASE}(隔离实例,本日构建) · 模拟器: cast-film 预设(bench 影子实例 4011)`,
-  `- 产线: ${twin.lineId} · 产品 ${twin.productId} · 配方 ${batch1.recipeId}`,
-  `- 场景: ${SCENE_ID}@1.0.0(已冻结,用户确认令牌) · Hybrid Twin Channel \`${channelId}\``,
-  `- 模型: \`${model.id}\`(label: ${model.label})`,
-  ``,
-  `## 1. 模型身份与谱系(输入=全部 DCW;输出=膜厚 goal)`,
-  ``,
-  '```json',
-  JSON.stringify({
-    id: model.id, label: model.label, description: model.description,
-    lineId: model.lineId, recipeId: model.recipeId, purpose: model.purpose, objectiveId: model.objectiveId,
-    dataset: dataset.id, windows: dataset.rowCount, runs: dataset.runIds.length,
-    inputs: Object.keys(twin.dcw), output: 'film-thickness(目标 50μm)',
-  }, null, 2),
-  '```',
-  ``,
-  `## 2. 训练过程(平台参考训练器:物理参数校准 → 3 成员残差集成 → conformal UQ)`,
-  ``,
-  `- 作业: \`${job.id}\` · 平台门禁 G1-G5: ${gatesPassed ? '全部通过' : '未过'}`,
-  `- 训练日志(平台 ##AML 协议行 + 训练输出全文)→ training-run.log;阶段日志 → live-run.log`,
-  `- 平台权威指标:`,
-  ``,
-  '```json',
-  JSON.stringify({
-    oneStepTest: metrics.oneStepTest, rolloutTest: metrics.rolloutTest,
-    hybrid: metrics.hybrid, uncertainty: metrics.uncertainty, physics: metrics.physics,
-  }, null, 2).slice(0, 3600),
-  '```',
-  ``,
-  `## 3. 物理模型(骨架 + 训练校准后 θ 见模型工件 physics_parameters.json)`,
-  ``,
-  '```json',
-  JSON.stringify(physicsSpec, null, 2).slice(0, 3000),
-  '```',
-  ``,
-  `## 4. Twin Gate 判定(12 判据,回写模型 twinEligibility)`,
-  ``,
-  '```text',
-  gate.text.slice(0, 2400),
-  '```',
-  ``,
-  `## 5. MPC 推荐(recommendation-only;model-backed rollout)`,
-  ``,
-  '```text',
-  mpc.text.slice(0, 2400),
-  '```',
-  ``,
-  `## 6. 模型工件清单(aml/models/${model.id}/)`,
-  ``,
-  existsSync(modelDir) ? readdirSync(modelDir).map(f => `- ${f}(${statSync(join(modelDir, f)).size}B)`).join('\n') : `(服务器进程视角路径: ${modelDir})`,
-  ``,
-  `## 7. 全程事件日志`,
-  ``,
-  '```text',
-  L.join('\n').slice(-80_000),
-  '```',
-].join('\n')
-writeFileSync(DOC_PATH, doc)
-log(`✔ 透明度文档: ${DOC_PATH}`)
-log(`✔ 运行目录: ${OUT_DIR}`)
-log('ALL STAGES DONE')
+/**
+ * LIVE AML 闭环优化全链路验收(bench 方法论;真实平台 + 真实 PLC 模拟器 + 真实训练)。
+ *
+ * 链路:
+ *   隔离平台(3005,含本日构建) + 专用模拟器(4011 影子实例,cast-film 预设)
+ *   → admin 注册 → provisionTwinLine(6 DCW 执行器 + 5 DAQ 传感器,真实协议链路)
+ *   → 3 批次 × 330s 真实采样(DAQ 1s 节拍;val 批 ≥300 窗满足 Twin Gate G4)
+ *   → REST 构建数据集(输入=全部 6 个 DCW;输出目标=膜厚 goal;特征=熔温/熔压)
+ *   → hybrid_twin Channel(模板实例化 + worker 节点绑定)
+ *   → invoke 桥(POST /agent-tools/invoke): twin_scene_discover → twin_scene_compile
+ *     → twin_scene_freeze(用户确认令牌) → twin_physics_spec_draft(骨架物理模型)
+ *     → twin_physics_spec_validate
+ *   → REST 提交 hybrid_residual 训练(**无 code → 平台参考训练器**:物理参数校准 →
+ *     3 成员残差集成 → conformal 覆盖率),轮询状态/进度/日志尾(透明)
+ *   → 模型注册表校验(label/元数据/aml/models/<id> 工件) → REST 晋升 shadow
+ *   → invoke: twin_snapshot_create(auto_daq) → twin_trial_run×10 → twin_gate_evaluate
+ *     → mpc_optimize(model_id=训练模型驱动 rollout) → 读推荐
+ *   → 透明度文档 bench/results/<rid>/OPTIMIZATION-LOG.md
+ *
+ * 运行: SIM_BASE=http://127.0.0.1:4011 node scripts/live-aml-optloop.mjs(需先 aw build)
+ */
+import { mkdirSync, writeFileSync, appendFileSync, existsSync, readdirSync, statSync, readFileSync } from 'node:fs'
+import { spawn } from 'node:child_process'
+import { join, resolve, dirname } from 'node:path'
+import { fileURLToPath } from 'node:url'
+import { ensureSimulator, applyPreset, simNodes } from '../bench/lib/sim.mjs'
+import { provisionTwinLine, startTwinBatch } from '../bench/lib/closedloop.mjs'
+
+const REPO = resolve(dirname(fileURLToPath(import.meta.url)), '..')
+const AW_BASE = process.env.AW_BASE ?? 'http://127.0.0.1:3005'
+const AW_PORT = Number(new URL(AW_BASE).port)
+const RUN_ID = new Date().toISOString().replace(/[:.]/g, '-').slice(0, 19)
+const OUT_DIR = join(REPO, 'bench', 'results', `${RUN_ID}-aml-live-optloop`)
+mkdirSync(OUT_DIR, { recursive: true })
+const LOG_PATH = join(OUT_DIR, 'live-run.log')
+const DOC_PATH = join(OUT_DIR, 'OPTIMIZATION-LOG.md')
+const AML_ROOT = process.env.AW_AML_DIR_LIVE ?? join(REPO, 'aml')
+
+process.env.NO_PROXY = '127.0.0.1,localhost'
+process.env.no_proxy = '127.0.0.1,localhost'
+
+const L = []
+function log(line) {
+  const text = typeof line === 'string' ? line : JSON.stringify(line)
+  const stamped = `[${new Date().toISOString()}] ${text}`
+  console.log(stamped)
+  L.push(text)
+  appendFileSync(LOG_PATH, stamped + '\n')
+}
+function section(title) {
+  log('')
+  log('════════════════════════════════════════════════════════')
+  log(`  ${title}`)
+  log('════════════════════════════════════════════════════════')
+}
+const sleep = ms => new Promise(r => setTimeout(r, ms))
+
+async function api(method, path, body, token) {
+  const r = await fetch(AW_BASE + path, {
+    method,
+    headers: { 'content-type': 'application/json', ...(token ? { authorization: `Bearer ${token}` } : {}) },
+    body: body === undefined ? undefined : JSON.stringify(body),
+    signal: AbortSignal.timeout(120_000),
+  })
+  const json = await r.json().catch(() => null)
+  return { status: r.status, json }
+}
+
+async function invoke(tool, args, workerId, token) {
+  const r = await api('POST', '/api/workshop/agent-tools/invoke', { agentId: workerId, tool, args }, token)
+  const text = r.json?.data?.result?.text ?? r.json?.result?.text ?? JSON.stringify(r.json ?? {}).slice(0, 600)
+  log(`  [invoke] ${tool} → ${r.status}`)
+  return { status: r.status, text }
+}
+
+async function waitFor(predicate, { timeoutMs, everyMs = 4000, what }) {
+  const t0 = Date.now()
+  while (Date.now() - t0 < timeoutMs) {
+    const v = await predicate()
+    if (v) return v
+    await sleep(everyMs)
+  }
+  throw new Error(`等待超时: ${what}(${timeoutMs}ms)`)
+}
+
+/* ════════ Stage 0 · 隔离平台 + 专用模拟器 ════════ */
+section('Stage 0 · 隔离平台(3005) + 专用模拟器(4011)')
+
+async function platformUp(base) {
+  try {
+    const r = await fetch(`${base}/api/users/setup-status`, { signal: AbortSignal.timeout(5000) })
+    const j = await r.json().catch(() => null)
+    return r.status === 200 && j?.data != null
+  }
+  catch { return false }
+}
+
+if (!(await platformUp(AW_BASE))) {
+  const child = spawn(process.execPath, [join(REPO, 'bin', 'aw.mjs'), 'start', '--port', String(AW_PORT)], {
+    cwd: REPO,
+    env: { ...process.env, AW_AML_DIR: AML_ROOT, PORT: String(AW_PORT) },
+    detached: true,
+    stdio: 'ignore',
+  })
+  child.unref()
+  log(`平台冷启: port=${AW_PORT} pid=${child.pid}(repo 配置根)`)
+  log(`  AW_AML_DIR=${AML_ROOT}(复用真实训练 venv;模型工件落 repo ./aml/models)`)
+}
+await waitFor(() => platformUp(AW_BASE), { timeoutMs: 180_000, what: '平台就绪' })
+log(`✔ 平台在线 ${AW_BASE}`)
+const sim = await ensureSimulator({ log: m => log(`[sim] ${m}`) })
+log(`✔ 模拟器在线: ${JSON.stringify(sim).slice(0, 120)}`)
+await applyPreset('cast-film-physics')
+const simDevices = await simNodes()
+log(`✔ cast-film 预设应用,设备 ${simDevices.length} 台: ${simDevices.map(d => `${d.id}(${d.protocol})`).join(', ')}`)
+
+/* ════════ Stage 1 · admin 注册 ════════ */
+section('Stage 1 · 用户注册 + 提权(repo 库既有用户体系,新注册=user;本链路需 admin 建 DCW/产线)')
+let token = String(process.env.AW_TOKEN ?? '').trim()
+if (token) {
+  log('✔ 复用调用方 admin token（隔离实例已初始化，不再注册新账号）')
+}
+else {
+  const reg = await api('POST', '/api/workshop/users/register', { name: `aml-live-${RUN_ID}` })
+  token = reg.json?.data?.token
+  if (!token) throw new Error(`注册失败: ${JSON.stringify(reg.json).slice(0, 200)}`)
+  log('✔ token 就绪')
+  // 提升 role → admin（repo 库首次注册为 user 时的一次性测试授权）
+  {
+    const { DatabaseSync } = await import('node:sqlite')
+    const usersDb = new DatabaseSync(join(REPO, '.AgentWorkShop', 'data', 'users.sqlite'))
+    const r = usersDb.prepare(`UPDATE users SET role = 'admin' WHERE email LIKE 'aml-live-${RUN_ID}%' AND role = 'user'`).run()
+    usersDb.close()
+    log(`✔ 提权 admin: ${Number(r.changes)} 行`)
+  }
+}
+
+/* ════════ Stage 2 · 产线 provision + 开跑（或复用已完成采样的线/数据集）════════ */
+section('Stage 2 · 产线 provisioning(6 DCW 执行器 + 5 DAQ 传感器,驱动配置取自模拟器 /export)')
+const sfx = RUN_ID.slice(5, 16).replace(/-/g, '')
+const REUSE_LINE = String(process.env.AML_REUSE_LINE ?? '').trim()
+let twin
+let batch1
+let dataset
+let spReadback
+if (REUSE_LINE) {
+  twin = {
+    lineId: REUSE_LINE,
+    productId: String(process.env.AML_REUSE_PRODUCT ?? ''),
+    dcw: JSON.parse(String(process.env.AML_REUSE_DCW ?? '{}')),
+    daq: JSON.parse(String(process.env.AML_REUSE_DAQ ?? '{}')),
+  }
+  batch1 = { recipeId: String(process.env.AML_REUSE_RECIPE ?? '') }
+  dataset = {
+    id: String(process.env.AML_REUSE_DATASET ?? ''),
+    rowCount: Number(process.env.AML_REUSE_ROWS ?? 0),
+    runIds: [1, 2, 3],
+    sha256: String(process.env.AML_REUSE_SHA ?? ''),
+  }
+  spReadback = JSON.parse(String(process.env.AML_REUSE_SP ?? '{}'))
+  log(`✔ 复用采样结果 line=${twin.lineId} recipe=${batch1.recipeId} dataset=${dataset.id} (${dataset.rowCount} 窗)`)
+}
+else {
+  twin = await provisionTwinLine({ call: async (m, p, b) => { const r = await api(m, p, b, token); return { status: r.status, ...(r.json ?? {}) } } }, { simDevices, sfx })
+  for (const e of twin.ev) log('  ' + e)
+  if (!twin.ok) throw new Error(`provision 失败: ${(twin.errors ?? []).join('; ')} | ev=${(twin.ev ?? []).join(' ; ')}`)
+  log(`✔ lineId=${twin.lineId} productId=${twin.productId}`)
+  log(`  DCW: ${JSON.stringify(twin.dcw)}`)
+  log(`  DAQ: ${JSON.stringify(twin.daq)}`)
+
+  batch1 = await startTwinBatch({ call: async (m, p, b) => { const r = await api(m, p, b, token); return { status: r.status, ...(r.json ?? {}) } } }, { lineId: twin.lineId, productId: twin.productId, dcw: twin.dcw, sfx })
+  if (!batch1.ok) throw new Error(`开跑失败: ${batch1.errors?.join('; ')}`)
+  log(`✔ 批次开跑 recipeId=${batch1.recipeId}`)
+
+  // ── SP 设定点回读 DAQ 节点:AML 数据集的 control 输入取自数采时序库,
+  //    执行器设定点必须有 DAQ 回读节点(同一信号,只读)才能进入训练。──
+  const { simExport } = await import('../bench/lib/sim.mjs')
+  spReadback = {}
+  const spSignalIds = Object.keys(twin.dcw) // zone1-sp / screw-sp / linespeed-sp / diegap-sp …
+  for (const dev of simDevices) {
+    const exp = await simExport(dev.id)
+    const items = exp?.items ?? []
+    for (const spId of spSignalIds) {
+      if (spReadback[spId]) continue
+      const sig = (dev.signals ?? []).find(s => s.id === spId)
+      if (!sig) continue
+      // export 的 item.signal = 信号【名】(如「加热区1SP」),按名对齐
+      const item = items.find(i => i.signal === sig.name)
+      if (!item?.driverConfig) continue
+      // 写配置 → 读配置转换:mqtt 写用 jsonKey(发布字段),DAQ 读订阅同主题用 jsonPath 取值
+      let cfg = { ...item.driverConfig }
+      if (cfg.jsonKey !== undefined && cfg.jsonPath === undefined) {
+        cfg.jsonPath = cfg.jsonKey
+        delete cfg.jsonKey
+      }
+      const r = await api('POST', '/api/workshop/daq', {
+        name: `${sig.name} 回读 ${sfx}`, templateRef: 'daq-temp-tc', driver: dev.protocol, driverConfig: cfg,
+        unit: sig.unit, min: sig.min, max: sig.max, lineId: twin.lineId, intervalMs: 1000, publishIntervalMs: 0,
+        semantics: `执行器设定点回读(AML control 输入) ${item.signal}`,
+      }, token)
+      if (r.json?.data?.node?.id) spReadback[spId] = r.json.data.node.id
+    }
+  }
+  log(`✔ SP 回读 DAQ 节点 ${Object.keys(spReadback).length}/6: ${JSON.stringify(spReadback)}`)
+
+  /* ════════ Stage 3 · 3 批次采样(每批 500s;批内两段安全阶跃激励) ════════ */
+  section('Stage 3 · 真实采样:3 批次 × 500s(批内阶跃激励:每个切分都覆盖 3 个设定点水平)')
+  // 实验设计:设定点只在批次间变化会导致 train/test 分布割裂(外推爆炸,G1/G3 必挂)。
+  // 批内阶跃:每个 run 依次经历 base → +Δ1 → +Δ2 三个水平,train/val/test 同分布;
+  // 各 run 的 Δ 略作变化增加丰富度,全部在配方窗口与安全上限内。
+  const BASE_SP = { 'linespeed-sp': 95, 'diegap-sp': 1.0 }
+  const EXCITE_RUNS = [
+    [{ ls: 2.0, dg: 0.03 }, { ls: -1.5, dg: 0.06 }],
+    [{ ls: -1.0, dg: 0.05 }, { ls: 2.5, dg: 0.02 }],
+    [{ ls: 1.5, dg: 0.04 }, { ls: -2.0, dg: 0.05 }],
+  ]
+  for (let run = 1; run <= 3; run++) {
+    if (run > 1) {
+    // stop 是异步收窗：立即 start 会撞「已在运行」。必须轮询到 active=false 再开。
+      for (let attempt = 0; attempt < 12; attempt++) {
+        await api('POST', `/api/workshop/dcw/lines/${twin.lineId}/stop`, {}, token)
+        await sleep(1000)
+        const snap = await api('GET', '/api/workshop/dcw', undefined, token)
+        const state = (snap.json?.data?.lineStates ?? []).find(x => x.lineId === twin.lineId)
+        if (!state?.active) break
+        if (attempt === 11) throw new Error(`批次 ${run} 前停线超时: ${JSON.stringify(state)}`)
+      }
+      log('  上一批次停止(active=false)')
+      // 复用同一配方 stop/start → 新 runId(数据集按 recipeId 取批次,须同配方)
+      let started = null
+      for (let attempt = 0; attempt < 8; attempt++) {
+        started = await api('POST', `/api/workshop/dcw/lines/${twin.lineId}/start`, { recipeId: batch1.recipeId }, token)
+        if (started.status === 200) break
+        if (!/已在运行/.test(String(started.json?.message ?? ''))) break
+        await sleep(1000)
+      }
+      if (started?.status !== 200) throw new Error(`批次 ${run} 开跑失败: ${JSON.stringify(started?.json).slice(0, 200)}`)
+      log(`  批次 ${run} 开跑(同配方 ${batch1.recipeId})`)
+    }
+    await sleep(20_000) // base 水平稳定（基准压缩：3×240s，仍≥720 行 1s 样本）
+    const [e1, e2] = EXCITE_RUNS[run - 1]
+    await api('POST', `/api/workshop/dcw/${twin.dcw['linespeed-sp']}/write`, { value: BASE_SP['linespeed-sp'] + e1.ls }, token)
+    await api('POST', `/api/workshop/dcw/${twin.dcw['diegap-sp']}/write`, { value: BASE_SP['diegap-sp'] + e1.dg }, token)
+    log(`  批次${run} 阶跃1: linespeed=${BASE_SP['linespeed-sp'] + e1.ls} diegap=${(BASE_SP['diegap-sp'] + e1.dg).toFixed(2)}`)
+    await sleep(80_000)
+    await api('POST', `/api/workshop/dcw/${twin.dcw['linespeed-sp']}/write`, { value: BASE_SP['linespeed-sp'] + e2.ls }, token)
+    await api('POST', `/api/workshop/dcw/${twin.dcw['diegap-sp']}/write`, { value: BASE_SP['diegap-sp'] + e2.dg }, token)
+    log(`  批次${run} 阶跃2: linespeed=${BASE_SP['linespeed-sp'] + e2.ls} diegap=${(BASE_SP['diegap-sp'] + e2.dg).toFixed(2)}`)
+    await sleep(140_000) // 第三水平保持至 240s
+    const pv = twin.daq['film-thickness']
+    const r = await api('GET', `/api/workshop/daq/${pv}/samples?bucketMs=1000&limit=3`, undefined, token)
+    const latest = r.json?.data?.points?.[0]?.avg ?? '?'
+    log(`  批次${run} 完成(500s) · 膜厚当前≈${typeof latest === 'number' ? latest.toFixed(2) : latest} μm`)
+  }
+  await api('POST', `/api/workshop/dcw/lines/${twin.lineId}/stop`, {}, token)
+  log('✔ 采样完成(3 批次 × 330s),产线停止')
+
+  /* ════════ Stage 4 · 构建数据集(输入=SP 回读 DAQ,目标=膜厚) ════════ */
+  section('Stage 4 · AML 数据集构建(输入=6 个 SP 回读 DAQ;输出目标=膜厚 goal;特征=熔温/熔压)')
+  if (Object.keys(spReadback).length < 6) throw new Error(`SP 回读节点不足: ${JSON.stringify(spReadback)}`)
+  // 采样预检:逐节点确认批次期间有样本(无样本的回读剔除并记录,防 0 窗数据集)
+  const sampledControls = []
+  for (const [sig, nodeId] of Object.entries(spReadback)) {
+    const r = await api('GET', `/api/workshop/daq/${nodeId}/samples?bucketMs=60000&limit=1`, undefined, token)
+    const has = (r.json?.data?.points ?? []).length > 0
+    log(`  回读采样检查 ${sig}(${nodeId}): ${has ? '✓ 有样本' : '✘ 无样本,剔除'}`)
+    if (has) sampledControls.push({ nodeId, role: 'control' })
+  }
+  if (sampledControls.length < 4) throw new Error(`可用 control 回读不足: ${sampledControls.length}`)
+  const dsBody = {
+    lineId: twin.lineId, productId: twin.productId, recipeId: batch1.recipeId,
+    nodes: [
+      ...sampledControls,
+      { nodeId: twin.daq['film-thickness'], role: 'target' },
+      { nodeId: twin.daq['melt-temp'], role: 'feature' },
+      { nodeId: twin.daq['melt-pressure'], role: 'feature' },
+    ],
+    beatMs: 1000, window: { historySteps: 8, horizonSteps: 4 },
+    split: { valRatio: 0.34, testRatio: 0.33, seed: 42 },
+    purpose: 'mpc_surrogate',
+    note: `live 闭环优化数据集 ${RUN_ID}`,
+  }
+  const ds = await api('POST', '/api/workshop/aml/datasets', dsBody, token)
+  dataset = ds.json?.data?.dataset
+  if (!dataset) throw new Error(`数据集构建失败: ${JSON.stringify(ds.json).slice(0, 300)}`)
+  log(`✔ 数据集 ${dataset.id}: ${dataset.rowCount} 窗 / ${dataset.runIds.length} 批 / sha=${String(dataset.sha256).slice(0, 12)}`)
+}
+
+/* ════════ Stage 5 · hybrid_twin Channel 实例化 + worker 绑定 ════════ */
+section('Stage 5 · Hybrid Twin Channel(模板实例化 + worker 节点绑定)')
+const templates = await api('GET', '/api/workshop/channel-templates', undefined, token)
+const tpl = templates.json?.data?.find(x => x.id === 'chtpl-hybrid-twin-mpc-default')
+if (!tpl) throw new Error('hybrid 模板缺失')
+const SCENE_ID = `castfilm-hold-opt-${sfx}`
+const inst = await api('POST', `/api/workshop/channel-templates/${tpl.id}/instantiate`, {
+  name: `AML Live OptLoop ${RUN_ID}`,
+  toolProfile: 'hybrid_twin',
+  scene: { sceneId: SCENE_ID, sceneVersion: '1.0.0', lineId: twin.lineId, productId: twin.productId, recipeId: batch1.recipeId },
+  objective: { objectiveId: 'thickness-50um', targets: { film_thickness: 50 } },
+  controlPolicy: 'recommendation_only',
+}, token)
+const channelId = inst.json?.data?.channelId
+const worker = inst.json?.data?.agents?.find(x => x.role === 'worker')
+if (!channelId || !worker) throw new Error(`实例化失败: ${JSON.stringify(inst.json).slice(0, 300)}`)
+log(`✔ channelId=${channelId} worker=${worker.id}`)
+// 权限模型 v2：dcw 节点直绑已废弃；hybrid 场景的控制面 = 绑定 recipe（参数节点展开为 DCW）。
+const recipeBind = await api('POST', '/api/workshop/agent-tools/bindings', { agentId: worker.id, nodeId: batch1.recipeId, kind: 'recipe', mode: 'auto' }, token)
+if (recipeBind.status !== 200) throw new Error(`配方绑定失败: ${JSON.stringify(recipeBind.json).slice(0, 200)}`)
+for (const nodeId of Object.values(twin.daq)) {
+  await api('POST', '/api/workshop/agent-tools/bindings', { agentId: worker.id, nodeId, kind: 'daq', mode: 'auto' }, token)
+}
+log(`✔ worker 绑定 recipe(${Object.keys(twin.dcw).length} 参数) + ${Object.keys(twin.daq).length} DAQ`)
+
+/* ════════ Stage 6 · 场景发现→编译→冻结;骨架 PhysicsSpec ════════ */
+section('Stage 6 · 场景发现→编译→冻结;骨架 PhysicsSpec(twin_physics_spec_draft)')
+// 膜厚节点语义补『厚度』关键词(inferNodeRole 目标词表;信号名『膜厚』缺『度』不命中)→
+// 不补语义膜厚会回退成 state,契约里没有 target,骨架物理模型就没有优化输出。
+await api('PATCH', `/api/workshop/daq/${twin.daq['film-thickness']}`, { semantics: '流延膜厚度测量(优化目标 goal,单位 μm;厚度质量输出)' }, token)
+const disco = await invoke('twin_scene_discover', {}, worker.id, token)
+log(disco.text.split('\n').slice(0, 14).map(x => '  ' + x).join('\n'))
+log('  膜厚目标识别: ' + (disco.text.includes('"target": 1') || disco.text.includes('"target":1') ? '✔ target=1' : '✘ 仍是 0(检查语义)'))
+const compiled = await invoke('twin_scene_compile', {
+  scene_id: SCENE_ID, scene_version: '1.0.0', line_id: twin.lineId, product_id: twin.productId, recipe_id: batch1.recipeId,
+  prompt: '流延膜厚度闭环优化:目标膜厚 50μm。控制=6 个执行器设定点,观测=膜厚(目标)/熔温/熔压。约束:熔温 195~225℃,熔压 ≤22MPa。膜厚为目标输出。',
+}, worker.id, token)
+log(compiled.text.split('\n').slice(0, 12).map(x => '  ' + x).join('\n'))
+const frozen = await invoke('twin_scene_freeze', {
+  scene_id: SCENE_ID, scene_version: '1.0.0', confirmation: 'USER_CONFIRMED_SCENE_CONTRACT', approved_by: 'live-admin',
+}, worker.id, token)
+log(frozen.text.split('\n').slice(0, 8).map(x => '  ' + x).join('\n'))
+
+// 冻结契约全文 = 孪生工件目录 scenes/<id>-<ver>-frozen/scenes.json(平台落盘,本地可读;
+// 结构 { scene, contractHash, status, approvedBy, approvedAt },契约在 .scene 下)
+const frozenArtifactPath = join(AML_ROOT, 'twins', 'scenes', `${SCENE_ID}-1.0.0-frozen`, 'scenes.json')
+if (!existsSync(frozenArtifactPath)) throw new Error(`冻结场景工件缺失: ${frozenArtifactPath}`)
+const frozenArtifact = JSON.parse(readFileSync(frozenArtifactPath, 'utf8'))
+const frozenScene = frozenArtifact.scene ?? frozenArtifact
+const sceneControls = frozenScene.controls ?? []
+log(`✔ 冻结契约: controls=${sceneControls.length} targets=${(frozenScene.observations ?? []).filter(v => v.role === 'target').length} 约束=${(frozenScene.constraints ?? []).length}`)
+
+const draft = await invoke('twin_physics_spec_draft', { dt_sec: 1 }, worker.id, token)
+log(draft.text.split('\n').slice(0, 6).map(x => '  ' + x).join('\n'))
+// 工具输出为摘要;spec 全文在落盘工件(aml/twins/physics-spec-draft/<id>/physics-spec-draft.json)
+const draftArtifactPath = (draft.text.match(/artifact:\s*([^\n]+)/) || [])[1]?.trim()
+if (!draftArtifactPath || !existsSync(draftArtifactPath)) throw new Error(`骨架 spec 工件缺失: ${draftArtifactPath}`)
+const physicsSpec = JSON.parse(readFileSync(draftArtifactPath, 'utf8'))
+log(`✔ 骨架 PhysicsSpec: model=${physicsSpec.modelId} 状态方程 ${physicsSpec.states.length} 观测 ${physicsSpec.observations.length} 可校准参数 ${physicsSpec.parameters.filter(p => p.min != null && p.max != null).length}/${physicsSpec.parameters.length}`)
+writeFileSync(join(OUT_DIR, 'physics-spec.json'), JSON.stringify(physicsSpec, null, 2))
+const validated = await invoke('twin_physics_spec_validate', { physics_spec: physicsSpec }, worker.id, token)
+log(validated.text.split('\n').slice(0, 5).map(x => '  ' + x).join('\n'))
+
+// ── 训练前对齐(关键):场景控制绑定的是 DCW 执行器,而数据集 control 列是 SP 回读
+//    DAQ 节点 → spec 控制变量必须改绑回读节点;且 spec 只能引用数据集包含的节点
+//    (场景可比数据集宽,宽出的变量取不到值 → 物理 NaN fail-closed)。──
+const datasetManifest = JSON.parse(readFileSync(join(AML_ROOT, 'datasets', dataset.id, 'manifest.json'), 'utf8'))
+const datasetNodes = new Set(datasetManifest.allNodes)
+const dwToDn = {}
+for (const [sig, dn] of Object.entries(spReadback)) dwToDn[twin.dcw[sig]] = dn
+let remapped = 0
+for (const v of physicsSpec.variables) {
+  if (v.nodeId && !datasetNodes.has(v.nodeId) && dwToDn[v.nodeId]) { v.nodeId = dwToDn[v.nodeId]; remapped++ }
+}
+const keptVars = physicsSpec.variables.filter(v => !v.nodeId || datasetNodes.has(v.nodeId))
+const keptIds = new Set(keptVars.map(v => v.id))
+const dropped = physicsSpec.variables.filter(v => !keptIds.has(v.id)).map(v => `${v.id}(${v.nodeId})`)
+physicsSpec.variables = keptVars
+physicsSpec.states = physicsSpec.states.filter((e) => { const base = e.lhs.endsWith('_next') ? e.lhs.slice(0, -5) : e.lhs; return keptIds.has(base) })
+physicsSpec.observations = physicsSpec.observations.filter(e => keptIds.has(e.lhs))
+physicsSpec.guards = (physicsSpec.guards ?? []).filter(e => keptIds.has(e.lhs))
+physicsSpec.constraints = physicsSpec.constraints.filter(c => keptIds.has(c.id))
+writeFileSync(join(OUT_DIR, 'physics-spec-remapped.json'), JSON.stringify(physicsSpec, null, 2))
+log(`✔ 训练对齐: 控制改绑 ${remapped};数据集裁剪剔除 ${dropped.length} 变量(${dropped.join(', ') || '无'});保留 ${keptVars.length}`)
+const revalidated = await invoke('twin_physics_spec_validate', { physics_spec: physicsSpec }, worker.id, token)
+log('  再校验: ' + revalidated.text.split('\n').filter(x => x.includes('"valid"')).join(' ').slice(0, 60))
+
+/* ════════ Stage 7 · hybrid 训练(平台参考训练器,无 code) ════════ */
+section('Stage 7 · 提交 hybrid_residual 训练(无 code → 平台参考训练器;物理校准+3 成员残差集成+UQ)')
+const jobRes = await api('POST', '/api/workshop/aml/jobs', {
+  datasetId: dataset.id,
+  purpose: 'mpc_surrogate',
+  changeNote: 'live 闭环优化:骨架物理+数据残差(全部 DCW → 膜厚 goal)',
+  params: { epochs: 220, hidden: 96, lr: 0.002, residual_scale: 1.2, ensemble: 3 },
+  seed: 42,
+  jobKind: 'hybrid_residual',
+  sceneId: SCENE_ID,
+  sceneVersion: '1.0.0',
+  objectiveId: 'thickness-50um',
+  physicsSpec,
+  providerId: physicsSpec.modelId,
+  providerVersion: '1.0.0',
+  providerHash: `sha256:${JSON.stringify(physicsSpec).length}`,
+  modelName: `流延膜厚闭环模型 ${sfx}`,
+  modelDescription: `输入=全部 6 个 DCW 执行器;输出=膜厚 goal(50μm);场景 ${SCENE_ID}@1.0.0;产线 ${twin.lineId}/配方 ${batch1.recipeId}`,
+}, token)
+const job = jobRes.json?.data?.job
+if (!job) throw new Error(`作业提交失败: ${JSON.stringify(jobRes.json).slice(0, 300)}`)
+log(`✔ 作业 ${job.id} 入队(jobKind=${job.budget?.jobKind}, scene=${job.budget?.sceneId})`)
+log(`  模型标识 name: ${job.budget?.modelName}`)
+log(`  描述: ${job.budget?.modelDescription}`)
+
+const jobLogsPath = join(OUT_DIR, 'training-run.log')
+let jobDone = null
+for (let i = 0; i < 150; i++) {
+  await sleep(10_000)
+  const st = await api('GET', `/api/workshop/aml/jobs/${job.id}`, undefined, token)
+  const j = st.json?.data?.job ?? st.json?.data
+  if (!j) continue
+  log(`  [${i * 10}s] ${j.status} ${j.stage} ${j.progress}%`)
+  const logFile = join(AML_ROOT, 'jobs', job.id, 'run.log')
+  if (existsSync(logFile)) {
+    const lines = readFileSync(logFile, 'utf8').split('\n').filter(Boolean)
+    writeFileSync(jobLogsPath, lines.join('\n'))
+    for (const t of lines.filter(l => l.includes('##AML')).slice(-3)) log(`    ${t.slice(0, 170)}`)
+  }
+  if (['done', 'failed', 'cancelled', 'timeout'].includes(j.status)) { jobDone = j; break }
+}
+if (!jobDone || jobDone.status !== 'done') throw new Error(`训练未成功: ${JSON.stringify(jobDone).slice(0, 300)}`)
+const gatesPassed = (() => {
+  try { return JSON.parse(jobDone.gatesJson || '{}').passed }
+  catch { return false }
+})()
+log(`✔ 训练完成 · 平台门禁 G1-G5: ${gatesPassed ? '全部通过' : '未过'}`)
+log(`  训练日志全文 → ${jobLogsPath}`)
+
+/* ════════ Stage 8 · 模型保存校验 + 晋升 shadow ════════ */
+section('Stage 8 · 模型保存校验(aml/models/<id> 工件 + label 元数据)+ 晋升 shadow')
+const modelsRes = await api('GET', '/api/workshop/aml/models', undefined, token)
+const model = (modelsRes.json?.data?.models ?? []).find(m => m.datasetId === dataset.id)
+if (!model) throw new Error('模型未登记')
+log(`✔ 模型 ${model.id} [${model.stage}]`)
+log(`  label: ${model.label}`)
+log(`  description: ${model.description}`)
+log(`  line=${model.lineId} recipe=${model.recipeId} purpose=${model.purpose} objective=${model.objectiveId}`)
+const modelDir = model.path
+if (existsSync(modelDir)) {
+  log(`  工件目录 ${modelDir}:`)
+  for (const f of readdirSync(modelDir)) log(`    - ${f}(${statSync(join(modelDir, f)).size}B)`)
+}
+else log(`  ⚠ 模型目录(服务器进程视角): ${modelDir}`)
+const metrics = model.metrics ?? {}
+log(`  G1 单步 NRMSE=${metrics.oneStepTest?.nrmse} · G2 滚动=${metrics.rolloutTest?.nrmse}`)
+log(`  hybrid: ${JSON.stringify(metrics.hybrid ?? {}).slice(0, 240)}`)
+log(`  uncertainty: ${JSON.stringify(metrics.uncertainty ?? {}).slice(0, 240)}`)
+log(`  physics: ${JSON.stringify(metrics.physics ?? {}).slice(0, 240)}`)
+const promo = await api('POST', `/api/workshop/aml/models/${model.id}/promote`, { toStage: 'shadow' }, token)
+log(`✔ 晋升 shadow: HTTP ${promo.status}`)
+
+/* ════════ Stage 9 · 孪生闭环:快照 → 10 试验 → 门禁 → MPC 推荐 ════════ */
+section('Stage 9 · 孪生闭环(auto_daq 快照 → 10 候选试验 → Twin Gate → MPC 推荐,训练模型驱动)')
+await api('POST', `/api/workshop/dcw/lines/${twin.lineId}/start`, { recipeId: batch1.recipeId }, token)
+log('  产线重启(真实工况下采快照)')
+await sleep(25_000)
+const snap = await invoke('twin_snapshot_create', { auto_daq: true, channel_id: channelId, phase: 'holding' }, worker.id, token)
+log(snap.text.split('\n').slice(0, 8).map(x => '  ' + x).join('\n'))
+const snapshotId = (snap.text.match(/snapshot_id:\s*([^\n]+)/) || [])[1]?.trim()
+if (!snapshotId) throw new Error('快照创建失败: ' + snap.text.slice(0, 300))
+
+// 候选轨迹:线速度小幅变化(量程 20~120,maxStep 由场景 writePolicy 派生)
+const speedCtl = sceneControls.find(c => (c.nodeId ?? '').includes('linespeed')) ?? sceneControls[0]
+const speedStep = speedCtl?.maxStep ?? 2
+for (let k = 0; k < 10; k++) {
+  const delta = (k % 2 === 0 ? 1 : -0.5) * speedStep * (0.3 + 0.07 * k)
+  const t = await invoke('twin_trial_run', {
+    snapshot_id: snapshotId,
+    baseline_controls: { [speedCtl.id]: 95 },
+    candidate_controls: [
+      { [speedCtl.id]: 95 + delta }, { [speedCtl.id]: 95 + delta },
+      { [speedCtl.id]: 95 + delta }, { [speedCtl.id]: 95 + delta },
+    ],
+  }, worker.id, token)
+  log(`  trial#${k + 1} Δlinespeed=${delta.toFixed(2)} → ${t.text.split('\n').filter(x => x.includes('constraints_passed') || x.includes('improvement')).join(' | ').slice(0, 160)}`)
+}
+const gate = await invoke('twin_gate_evaluate', { model_id: model.id }, worker.id, token)
+log('  Twin Gate(12 判据,回写 twinEligibility):')
+log(gate.text.split('\n').slice(2, 26).map(x => '    ' + x).join('\n'))
+const mpc = await invoke('mpc_optimize', {
+  snapshot_id: snapshotId,
+  model_id: model.id,
+  baseline_controls: { [speedCtl.id]: 95 },
+  horizon_steps: 4,
+}, worker.id, token)
+writeFileSync(join(OUT_DIR, 'mpc-recommendation.txt'), mpc.text)
+const mode = (mpc.text.match(/"mode":\s*"([^"]+)"/) || [])[1]
+const bestCandidate = (mpc.text.match(/"bestCandidate":\s*(\{[^}]*\})/) || [])[1]
+const bestCost = (mpc.text.match(/"bestCost":\s*([\d.eE+-]+)/) || [])[1]
+log(`✔ MPC 完成: mode=${mode} bestCandidate=${bestCandidate} bestCost=${bestCost}`)
+log(`  完整推荐 → ${join(OUT_DIR, 'mpc-recommendation.txt')}`)
+
+/* ════════ Stage 9b · 闭环执行:推荐设定点 → admin 写入(HITL 授权模拟)→ 复测 ════════ */
+section('Stage 9b · 闭环执行:把 MPC 推荐的设定点写入产线(admin 执行),等待响应后复测')
+let best
+try {
+  best = JSON.parse(bestCandidate ?? 'null')
+}
+catch {
+  best = null
+}
+if (best && Object.keys(best).length > 0) {
+  // spec 控制 id → 场景控制 nodeId(DCW 节点)
+  const idToNode = Object.fromEntries(sceneControls.map(c => [c.id, c.nodeId]))
+  const writes = []
+  for (const [specId, value] of Object.entries(best)) {
+    const nodeId = idToNode[specId]
+    if (!nodeId || !Number.isFinite(Number(value))) continue
+    const w = await api('POST', `/api/workshop/dcw/${nodeId}/write`, { value }, token)
+    writes.push({ specId, nodeId, value, status: w.status })
+    log(`  写入 ${nodeId} ← ${Number(value).toFixed(2)} → HTTP ${w.status}`)
+  }
+  log(`  等待 90s 工艺响应(settling)…`)
+  await sleep(90_000)
+  const pv = twin.daq['film-thickness']
+  const r = await api('GET', `/api/workshop/daq/${pv}/samples?bucketMs=1000&limit=60`, undefined, token)
+  const pts = (r.json?.data?.points ?? []).map(p => Number(p.avg ?? p.value)).filter(Number.isFinite).slice(0, 30)
+  const mean = pts.length ? pts.reduce((a, b) => a + b, 0) / pts.length : Number.NaN
+  log(`✔ 执行后膜厚均值(最近 30s): ${mean.toFixed(2)} μm(目标 50μm;写入前 ≈54 μm)`)
+  writeFileSync(join(OUT_DIR, 'closed-loop-write.json'), JSON.stringify({ writes, thicknessAfter: mean }, null, 2))
+}
+else {
+  log('  ⚠ MPC 未给出可执行的 bestCandidate(如 safe_small_step 阶段)——按平台治理,此阶段只推荐不写。')
+}
+
+/* ════════ Stage 10 · 透明度文档 ════════ */
+section('Stage 10 · 生成透明度文档')
+const doc = [
+  `# AML 混合孪生闭环优化 · 实机透明度记录`,
+  ``,
+  `- 运行标识: ${RUN_ID}`,
+  `- 平台: ${AW_BASE}(隔离实例,本日构建) · 模拟器: cast-film 预设(bench 影子实例 4011)`,
+  `- 产线: ${twin.lineId} · 产品 ${twin.productId} · 配方 ${batch1.recipeId}`,
+  `- 场景: ${SCENE_ID}@1.0.0(已冻结,用户确认令牌) · Hybrid Twin Channel \`${channelId}\``,
+  `- 模型: \`${model.id}\`(label: ${model.label})`,
+  ``,
+  `## 1. 模型身份与谱系(输入=全部 DCW;输出=膜厚 goal)`,
+  ``,
+  '```json',
+  JSON.stringify({
+    id: model.id, label: model.label, description: model.description,
+    lineId: model.lineId, recipeId: model.recipeId, purpose: model.purpose, objectiveId: model.objectiveId,
+    dataset: dataset.id, windows: dataset.rowCount, runs: dataset.runIds.length,
+    inputs: Object.keys(twin.dcw), output: 'film-thickness(目标 50μm)',
+  }, null, 2),
+  '```',
+  ``,
+  `## 2. 训练过程(平台参考训练器:物理参数校准 → 3 成员残差集成 → conformal UQ)`,
+  ``,
+  `- 作业: \`${job.id}\` · 平台门禁 G1-G5: ${gatesPassed ? '全部通过' : '未过'}`,
+  `- 训练日志(平台 ##AML 协议行 + 训练输出全文)→ training-run.log;阶段日志 → live-run.log`,
+  `- 平台权威指标:`,
+  ``,
+  '```json',
+  JSON.stringify({
+    oneStepTest: metrics.oneStepTest, rolloutTest: metrics.rolloutTest,
+    hybrid: metrics.hybrid, uncertainty: metrics.uncertainty, physics: metrics.physics,
+  }, null, 2).slice(0, 3600),
+  '```',
+  ``,
+  `## 3. 物理模型(骨架 + 训练校准后 θ 见模型工件 physics_parameters.json)`,
+  ``,
+  '```json',
+  JSON.stringify(physicsSpec, null, 2).slice(0, 3000),
+  '```',
+  ``,
+  `## 4. Twin Gate 判定(12 判据,回写模型 twinEligibility)`,
+  ``,
+  '```text',
+  gate.text.slice(0, 2400),
+  '```',
+  ``,
+  `## 5. MPC 推荐(recommendation-only;model-backed rollout)`,
+  ``,
+  '```text',
+  mpc.text.slice(0, 2400),
+  '```',
+  ``,
+  `## 6. 模型工件清单(aml/models/${model.id}/)`,
+  ``,
+  existsSync(modelDir) ? readdirSync(modelDir).map(f => `- ${f}(${statSync(join(modelDir, f)).size}B)`).join('\n') : `(服务器进程视角路径: ${modelDir})`,
+  ``,
+  `## 7. 全程事件日志`,
+  ``,
+  '```text',
+  L.join('\n').slice(-80_000),
+  '```',
+].join('\n')
+writeFileSync(DOC_PATH, doc)
+log(`✔ 透明度文档: ${DOC_PATH}`)
+log(`✔ 运行目录: ${OUT_DIR}`)
+log('ALL STAGES DONE')