agentworkshop 0.8.1 → 0.8.3

package/scripts/acceptance-aml-hybrid-e2e.ts

@@ -1,396 +1,396 @@
-/**
- * AML Hybrid 端到端验收:任意场景 → 骨架物理模型 → 平台 hybrid 训练 → 注册保存 →
- * 门禁 → 训练模型驱动 VirtualTrial/MPC → 影子预测 → 配方隔离拒绝。
- *
- * Part A(真实 Python):用平台参考训练器 train-hybrid-example.py 在真实 venv 里
- *   完成 物理参数校准(stage A)→ 3 成员残差集成(stage B)→ conformal 覆盖率
- *   (stage D)→ ONNX 工件;再跑平台权威评估器 aml_eval.py。
- * Part B(真实 TS):registerModelFromJob 注册模型 → transitionModel 晋升 shadow →
- *   twin_gate_evaluate 平台门 → ModelBackedHybridProvider 驱动 runVirtualTrial
- *   (残差真实参与,且与纯物理 rollout 有可测差异)→ predictWithModel 影子预测 →
- *   跨配方谱系拒绝。
- *
- * 运行:npx tsx --tsconfig .nuxt/tsconfig.server.json scripts/acceptance-aml-hybrid-e2e.ts
- * 依赖:./aml/.venv(真实训练环境,torch + onnx)。AW_AML_DIR 指向临时目录。
- */
-import { existsSync, mkdirSync, mkdtempSync, readFileSync, rmSync, writeFileSync, copyFileSync } from 'node:fs'
-import { tmpdir } from 'node:os'
-import { join, dirname } from 'node:path'
-import { fileURLToPath } from 'node:url'
-import { runPythonJob } from './_lib-aml-python-run.mjs'
-import { openWorkshopDb } from '../server/services/workshop/db/database/open'
-import { configureAmlRuntime } from '../server/services/workshop/aml/runtime'
-import { hashDatasetDir, loadManifest } from '../server/services/workshop/aml/dataset-builder'
-import { evaluateGates } from '../server/services/workshop/aml/gates'
-import { registerModelFromJob, transitionModel } from '../server/services/workshop/aml/model-registry'
-import { predictWithModel } from '../server/services/workshop/aml/predictor'
-import { draftPhysicsSpecFromScene } from '../server/services/workshop/aml/twin/physics-spec-draft'
-import { parseSceneContract, sha256 } from '../server/services/workshop/aml/twin/contracts'
-import { createTwinSnapshot } from '../server/services/workshop/aml/twin/snapshot-service'
-import { runVirtualTrial, syncRolloutProvider } from '../server/services/workshop/aml/twin/trial-service'
-import { compileDeclarativeProvider } from '../server/services/workshop/aml/twin/declarative-provider'
-import { evaluateHybridGates, DEFAULT_ACCEPTANCE_PROFILE } from '../server/services/workshop/aml/twin/acceptance'
-import { ModelBackedHybridProvider, assertModelSceneLineage } from '../server/services/workshop/aml/twin/model-backed-provider'
-import { createTwinRepo } from '../server/services/workshop/aml/twin/repo'
-
-const repoRoot = dirname(dirname(fileURLToPath(import.meta.url)))
-const fail = (msg: string): never => {
-  console.error(`FAIL ${msg}`)
-  process.exit(1)
-}
-const check = (name: string, ok: boolean, detail = ''): void => {
-  console.log(`${ok ? 'PASS' : 'FAIL'} ${name}${detail ? ` — ${detail}` : ''}`)
-  if (!ok) process.exit(1)
-}
-
-// ---------- 场景与合成数据(真实工况:一阶加热系统 + 轻微非线性残差) ----------
-const LINE = 'line-e2e'
-const PRODUCT = 'prd-e2e'
-const RECIPE = 'rcp-e2e'
-const SCENE_ID = 'e2e-fol-scene'
-const CONTROL_NODE = 'node.heater.set'
-const TARGET_NODE = 'node.temp.value'
-const H = 8
-const HORIZON = 4
-const DT = 1
-const N_TRAIN = 420
-const N_VAL = 320
-const N_TEST = 90
-
-// 真实 θ:offset=32, gain=1.6, tau=1.5(快系统:单拍内控制响应显著,物理参数可辨识);
-// y 在物理之上叠加确定性非线性(残差可学)。激励 = 确定性正弦混叠(可复现,免随机数)。
-const TRUE_OFFSET = 32
-const TRUE_GAIN = 1.6
-const TRUE_TAU = 1.5
-
-function simulateSeries(count: number, phase: number): Array<{ u: number, y: number }> {
-  // 单条连续时间序列(无段间重置,真实产线语义);相位由 phase 决定,幅值分布一致
-  const rows: Array<{ u: number, y: number }> = []
-  let y = TRUE_OFFSET + TRUE_GAIN * 4
-  for (let i = 0; i < count; i++) {
-    const u = 1 + 8 * (0.5 + 0.5 * Math.sin(i * 1.7 + phase) * Math.cos(i * 0.53 + phase * 0.37))
-    const physics = y + (DT / TRUE_TAU) * (TRUE_OFFSET + TRUE_GAIN * u - y)
-    const residual = 0.45 * Math.sin(0.8 * u) + 0.02 * Math.sin(i * 2.3)
-    y = physics + residual
-    rows.push({ u, y })
-  }
-  return rows
-}
-
-function writeDataset(dir: string): { rowCount: number } {
-  // 连续序列滑窗:window i = history rows [i, i+H),one-step target = row[i+H],
-  // horizon 控制轨迹 = rows [i+H, i+H+HORIZON);train/val/test 按窗口下标三分。
-  const total = N_TRAIN + N_VAL + N_TEST + H + HORIZON + 1
-  const series = simulateSeries(total, 0.9)
-  const rowCount = N_TRAIN + N_VAL + N_TEST
-  // norm 统计取自 train 窗口覆盖的原始行(真实数据集约定;aml_eval 的 NRMSE 分母即 norm std)
-  const trainRows = series.slice(0, N_TRAIN + H)
-  const trainU = trainRows.map(r => r.u)
-  const trainY = trainRows.map(r => r.y)
-  const mean = (xs: number[]): number => xs.reduce((a, b) => a + b, 0) / xs.length
-  const stdOf = (xs: number[], mu: number): number => Math.max(Math.sqrt(xs.reduce((a, b) => a + (b - mu) ** 2, 0) / xs.length), 1e-6)
-  const uMean = mean(trainU)
-  const uStd = stdOf(trainU, uMean)
-  const yMean = mean(trainY)
-  const yStd = stdOf(trainY, yMean)
-  const normU = (v: number): number => (v - uMean) / uStd
-  const normY = (v: number): number => (v - yMean) / yStd
-  const x = new Float32Array(rowCount * H * 2)
-  const u = new Float32Array(rowCount * HORIZON * 1)
-  const y = new Float32Array(rowCount * HORIZON * 1)
-  for (let i = 0; i < rowCount; i++) {
-    for (let h = 0; h < H; h++) {
-      const row = series[i + h]!
-      x[i * H * 2 + h * 2] = normU(row.u)
-      x[i * H * 2 + h * 2 + 1] = normY(row.y)
-    }
-    for (let k = 0; k < HORIZON; k++) {
-      const next = series[i + H + k]!
-      u[i * HORIZON + k] = normU(next.u)
-      y[i * HORIZON + k] = normY(next.y)
-    }
-  }
-  mkdirSync(join(dir, 'arrays'), { recursive: true })
-  writeFileSync(join(dir, 'arrays', 'x.f32'), x)
-  writeFileSync(join(dir, 'arrays', 'u.f32'), u)
-  writeFileSync(join(dir, 'arrays', 'y.f32'), y)
-  const manifest = {
-    version: 1,
-    lineId: LINE, productId: PRODUCT, recipeId: RECIPE,
-    beatMs: DT * 1000,
-    allNodes: [CONTROL_NODE, TARGET_NODE],
-    controlNodes: [CONTROL_NODE],
-    targetNodes: [TARGET_NODE],
-    shapes: { x: [rowCount, H, 2], u: [rowCount, HORIZON, 1], y: [rowCount, HORIZON, 1] },
-    norm: {
-      x: { mean: [uMean, yMean], std: [uStd, yStd] },
-      u: { mean: [uMean], std: [uStd] },
-      y: { mean: [yMean], std: [yStd] },
-    },
-    split: { train: N_TRAIN, val: N_VAL, test: N_TEST },
-    runs: ['r1', 'r2', 'r3'],
-  }
-  writeFileSync(join(dir, 'manifest.json'), JSON.stringify(manifest, null, 2))
-  return { rowCount }
-}
-
-// ---------- 冻结场景(带真实 nodeId 映射)→ 骨架 PhysicsSpec ----------
-function frozenScene() {
-  return parseSceneContract({
-    schemaVersion: 1,
-    createdAt: new Date().toISOString(),
-    createdBy: 'acceptance',
-    sceneId: SCENE_ID,
-    sceneVersion: '1.0.0',
-    lineId: LINE,
-    productId: PRODUCT,
-    recipeId: RECIPE,
-    phases: ['holding'],
-    controls: [{ id: 'heater_sp', nodeId: CONTROL_NODE, role: 'control', physicalMeaning: '加热设定', unit: 'kW', min: 1, max: 9, maxStep: 0.5 }],
-    states: [{ id: 'temp', nodeId: TARGET_NODE, role: 'state', physicalMeaning: '物料温度', unit: 'degC', min: 0, max: 90 }],
-    disturbances: [],
-    observations: [{ id: 'temp_obs', nodeId: TARGET_NODE, role: 'target', physicalMeaning: '温度观测', unit: 'degC', min: 0, max: 90 }],
-    guards: [],
-    constraints: [{ id: 'temp_obs', kind: 'hard_range', min: 20, max: 80 }],
-    physicsProfileId: 'declarative-e2e',
-    objectiveProfileIds: ['obj-temp-track'],
-    writePolicy: { minNodeIntervalSec: 60, minLineActionIntervalSec: 60, maxActionsPerRun: 3, maxDeltaPerAction: { heater_sp: 0.5 } },
-  })
-}
-
-// ---------- 主流程 ----------
-const root = mkdtempSync(join(tmpdir(), `aw-aml-hybrid-e2e-${process.pid}-`))
-const amlRoot = join(root, 'aml')
-process.env.AW_AML_DIR = amlRoot
-
-// Part B 前置:数据库 + 运行时
-const db = openWorkshopDb(join(root, 'workshop.sqlite'))
-const rt = configureAmlRuntime(db)
-
-const datasetDir = join(root, 'dataset-entity')
-const { rowCount } = writeDataset(datasetDir)
-
-// 骨架物理模型(任意场景建模起点;TS 侧生成,Python 侧同一 DSL 消费)
-const scene = frozenScene()
-const spec = draftPhysicsSpecFromScene({ scene, createdBy: 'acceptance', dtSec: DT })
-check('骨架 PhysicsSpec 自校验通过且含可校准参数',
-  spec.parameters.length >= 3 && spec.parameters.every(p => p.min != null && p.max != null),
-  `params=${spec.parameters.length}`)
-
-// ---- Part A: 真实 venv 训练 + 平台权威评估 ----
-const venvPython = join(repoRoot, 'aml', '.venv', process.platform === 'win32' ? 'Scripts\\python.exe' : 'bin/python')
-if (!existsSync(venvPython)) fail(`真实训练环境缺失:${venvPython}(先在仓库跑一次 AML 训练以供给 ./aml/.venv)`)
-const jobId = 'job-e2e-hybrid-1'
-const jobDir = join(root, 'job-e2e')
-const workspace = join(jobDir, 'workspace')
-const artifacts = join(jobDir, 'artifacts')
-mkdirSync(workspace, { recursive: true })
-mkdirSync(artifacts, { recursive: true })
-const platformPy = join(repoRoot, 'server', 'services', 'workshop', 'aml', 'python')
-copyFileSync(join(platformPy, 'amlkit.py'), join(workspace, 'amlkit.py'))
-copyFileSync(join(platformPy, 'train-hybrid-example.py'), join(workspace, 'train-hybrid-example.py'))
-writeFileSync(join(jobDir, 'physics_spec.json'), JSON.stringify(spec))
-writeFileSync(join(workspace, 'physics_spec.json'), JSON.stringify(spec))
-writeFileSync(join(jobDir, 'job.json'), JSON.stringify({
-  jobId, datasetPath: datasetDir, workspaceDir: workspace, params: { epochs: 400, hidden: 96, lr: 0.002, residual_scale: 1.5 }, seed: 42,
-  trainFile: 'train-hybrid-example.py', sceneId: SCENE_ID, sceneVersion: scene.sceneVersion,
-  objectiveId: 'obj-temp-track', jobKind: 'hybrid_residual',
-  providerId: 'declarative-e2e', providerVersion: '0.1.0-skeleton', providerHash: sha256(spec), providerGeneration: 0,
-}))
-
-const trainRun = runPythonJob(venvPython, [join(workspace, 'train-hybrid-example.py')], workspace, { AML_JOB_DIR: jobDir })
-check('Part A: 参考训练器真实运行成功(stage A 校准+B 集成+D 覆盖率)', trainRun.code === 0, trainRun.tail)
-for (const file of ['model.onnx', 'model-2.onnx', 'model-3.onnx', 'hybrid_manifest.json', 'physics_parameters.json']) {
-  check(`Part A: 工件 ${file} 存在`, existsSync(join(artifacts, file)))
-}
-const manifest = JSON.parse(readFileSync(join(artifacts, 'hybrid_manifest.json'), 'utf8')) as {
-  ensemble?: { members?: number, calibrationCoverage?: number | null, files?: string[] }
-  calibratedParameters?: Record<string, number>
-  physicsSpec?: { parameters?: Array<{ id: string, value?: number }> }
-}
-check('Part A: 3 成员集成 + 校准参数回写 manifest',
-  manifest.ensemble?.members === 3
-  && Object.keys(manifest.calibratedParameters ?? {}).length >= 3
-  && (manifest.ensemble?.calibrationCoverage ?? 0) > 0.5,
-  `coverage=${manifest.ensemble?.calibrationCoverage} calibrated=${JSON.stringify(manifest.calibratedParameters)}`)
-// 校准后的 θ 应离开先验默认并向真实系统收敛(gain→1.6,tau→6)
-const specParams = new Map((manifest.physicsSpec?.parameters ?? []).map(p => [p.id, p.value]))
-const gainId = [...specParams.keys()].find(k => k.includes('gain')) ?? ''
-const tauId = [...specParams.keys()].find(k => k.includes('tau')) ?? ''
-const offsetId = [...specParams.keys()].find(k => k.includes('offset')) ?? ''
-const gainVal = specParams.get(gainId) ?? Number.NaN
-const tauVal = specParams.get(tauId) ?? Number.NaN
-const offsetVal = specParams.get(offsetId) ?? Number.NaN
-// 训练器自报的校准前后物理拟合误差(训练阶段已写 artifacts/metrics.json)
-const trainedMetrics = JSON.parse(readFileSync(join(artifacts, 'metrics.json'), 'utf8')) as { physics?: { calibrationErrorBefore?: number, calibrationErrorAfter?: number } }
-check('Part A: stage A 参数校准有效(拟合误差下降)且 θ 朝真值方向收敛',
-  (trainedMetrics.physics?.calibrationErrorAfter ?? 1) < (trainedMetrics.physics?.calibrationErrorBefore ?? 0)
-  && gainVal > 0.3 && gainVal < 8 && offsetVal > 24 && offsetVal < 44 && tauVal >= 0.4 && tauVal <= 40,
-  `gain=${gainVal.toFixed(3)}/真值${TRUE_GAIN} offset=${offsetVal.toFixed(2)}/真值${TRUE_OFFSET} tau=${tauVal.toFixed(3)}/真值${TRUE_TAU} errBefore=${trainedMetrics.physics?.calibrationErrorBefore?.toFixed(4)} errAfter=${trainedMetrics.physics?.calibrationErrorAfter?.toFixed(4)}`)
-
-const evalRun = runPythonJob(venvPython, [join(platformPy, 'aml_eval.py'), '--job', join(jobDir, 'job.json')], jobDir, { AML_JOB_DIR: jobDir })
-check('Part A: 平台权威评估器 aml_eval 完成(物理+残差同口径)', evalRun.code === 0, evalRun.tail)
-const metricsRaw = JSON.parse(readFileSync(join(artifacts, 'metrics.json'), 'utf8')) as { platform?: { modelType?: string } & Record<string, unknown> } & Record<string, unknown>
-check('Part A: 评估器产出 hybrid modelType + platform 权威段',
-  metricsRaw.platform?.modelType === 'hybrid_physics_plus_residual' && !!metricsRaw.platform?.oneStepTest,
-  `modelType=${metricsRaw.platform?.modelType}`)
-
-// ---- Part B: 真实 TS 注册链 ----
-const specJson = JSON.stringify({ purpose: 'mpc_surrogate' })
-const sha = hashDatasetDir(datasetDir, specJson)
-rt.repo.dataset.insert({
-  id: 'ds-e2e', lineId: LINE, productId: PRODUCT, recipeId: RECIPE,
-  runIds: ['r1', 'r2', 'r3'], specJson, sha256: sha, rowCount,
-  fromMs: 0, toMs: 0, path: datasetDir, createdBy: 'acceptance', createdByKind: 'user',
-  note: '', createdAt: new Date().toISOString(),
-})
-rt.repo.job.insert({
-  id: jobId, datasetId: 'ds-e2e', purpose: 'mpc_surrogate',
-  budget: {
-    seed: 42, params: { epochs: 400, hidden: 96, lr: 0.002, residual_scale: 1.5 }, changeNote: 'e2e hybrid', jobKind: 'hybrid_residual',
-    sceneId: SCENE_ID, sceneVersion: scene.sceneVersion, objectiveId: 'obj-temp-track',
-    providerId: 'declarative-e2e', providerVersion: '0.1.0-skeleton', providerHash: sha256(spec), providerGeneration: 0,
-    inlineCode: null, modelName: 'E2E 温度跟踪 hybrid', modelDescription: '验收:任意场景→骨架物理→hybrid 训练→孪生闭环',
-  },
-  agentId: 'agt-acceptance', createdAt: new Date().toISOString(),
-})
-// concludeJob 同款注册路径(展平平台指标并保留自报块 + 门禁 + 登记)
-const metrics = ((): Record<string, unknown> => {
-  const parsed = metricsRaw as { platform?: Record<string, unknown> } & Record<string, unknown>
-  return { ...parsed, ...(parsed.platform ?? {}) }
-})()
-const gates = evaluateGates('mpc_surrogate', { rowCount, runIds: ['r1', 'r2', 'r3'] }, loadManifest(rt.repo.dataset.get('ds-e2e')!), metrics as never, artifacts)
-check('Part B: 平台门禁(G1-G5)通过', gates.passed, gates.checks.map(c => `${c.id}:${c.pass ? '✓' : '✗'}`).join(' '))
-// 注册的工件回退路径 = <AW_AML_DIR>/jobs/<jobId>/artifacts(orchestrator finish 后的布局)
-mkdirSync(join(amlRoot, 'jobs', jobId, 'artifacts'), { recursive: true })
-const trainedManifest = JSON.parse(readFileSync(join(artifacts, 'hybrid_manifest.json'), 'utf8')) as { ensemble?: { files?: string[] } }
-for (const file of trainedManifest.ensemble?.files ?? ['model.onnx']) {
-  copyFileSync(join(artifacts, file), join(amlRoot, 'jobs', jobId, 'artifacts', file))
-}
-copyFileSync(join(artifacts, 'hybrid_manifest.json'), join(amlRoot, 'jobs', jobId, 'artifacts', 'hybrid_manifest.json'))
-const reg = registerModelFromJob(jobId, gates, JSON.stringify(metrics))
-check('Part B: hybrid 模型注册入 AML', !!reg.modelId, reg.modelId ?? '')
-const model = rt.repo.model.get(reg.modelId!)!
-check('Part B: 模型工件落 AML(models/<id>/ 不可变)',
-  existsSync(join(model.path, 'model.onnx')) && existsSync(join(model.path, 'hybrid_manifest.json')) && existsSync(join(model.path, 'io_spec.json')))
-
-// 晋升 shadow(候选可流转;hybrid 禁止 candidate→production 直跳)
-const promoted = transitionModel(model.id, 'shadow', 'acceptance')
-check('Part B: candidate → shadow 晋升生效', promoted.ok)
-
-// ---- 训练模型驱动 VirtualTrial:10 个候选全走 model-backed rollout 并入 trial 库 ----
-const nowMs = Date.now()
-const lastRow = simulateSeries(40, 0.9).at(-1)!
-const snapshot = createTwinSnapshot({
-  scene, channelId: 'chn-acceptance', createdBy: 'acceptance', phase: 'holding',
-  controls: { heater_sp: 4 }, states: { temp: lastRow.y }, disturbances: {},
-  samples: [
-    { nodeId: CONTROL_NODE, at: nowMs - 1000, value: 4, sequence: 's1' },
-    { nodeId: TARGET_NODE, at: nowMs - 1000, value: lastRow.y, sequence: 's2' },
-  ],
-  nowMs, freshnessMaxMs: 60_000,
-})
-// 模型驱动 rollout(谱系校验 + provider 创建,与 rolloutProviderFor 工具层同守卫)
-const dsRow = rt.repo.dataset.get('ds-e2e')!
-assertModelSceneLineage(dsRow, scene)
-const backed = await ModelBackedHybridProvider.create(model, {
-  stateEstimate: snapshot.stateEstimate, controlValues: snapshot.controlValues, disturbances: snapshot.disturbances,
-})
-check('Part B: 训练模型解析为 model-backed rollout provider', backed.hybridManifest.ensemble?.members === 3 && backed.model.id === model.id)
-const objective = {
-  schemaVersion: 1, createdAt: new Date().toISOString(), createdBy: 'acceptance', objectiveId: 'obj-temp-track',
-  targets: { temp_obs: 42 }, weights: { temp_obs: 1 }, controlCosts: {}, horizonSteps: 4, trustRegion: {},
-} as never
-const twinRepo = createTwinRepo(rt.db)
-twinRepo.upsertScene(scene, 'acceptance')
-twinRepo.insertSnapshot(snapshot)
-const candidateSets = Array.from({ length: 10 }, (_, k) =>
-  Array.from({ length: 4 }, () => ({ heater_sp: 4 + 0.05 * (k + 1) })))
-const trials = []
-for (const candidate of candidateSets) {
-  const t = await runVirtualTrial({
-    scene, snapshot, modelId: model.id, modelHash: backed.modelHash, objective,
-    baselineControls: { heater_sp: 4 }, candidateControls: candidate,
-    provider: backed,
-    uncertainty: () => {
-      const u = backed.takeUncertainty()
-      return { predictions: u.memberPredictions, coverage: u.calibrationCoverage ?? 0, calibrationFresh: u.calibrationCoverage != null, inputDistance: 0 }
-    },
-    createdBy: 'acceptance', nowMs,
-  })
-  twinRepo.insertTrial(t, model.id)
-  trials.push(t)
-}
-const trial = trials[0]!
-check('Part B: 10 个候选 trial 全部 model rollout 完成且 candidateExecuted=false',
-  trials.length === 10 && trials.every(t => t.predictedTrajectory.length === 4 && t.candidateExecuted === false && t.predictedTrajectory.every(s => Object.values(s).every(Number.isFinite))))
-check('Part B: 集成 UQ 来自训练工件(3 成员,覆盖率≥0.9,OOD 接受)',
-  trial.uncertainty.ensembleSize === 3 && trial.uncertainty.coverage >= 0.9 && trial.outOfDistribution.accepted,
-  `uncertainty=${JSON.stringify(trial.uncertainty)} ood=${trial.outOfDistribution.accepted}(${trial.outOfDistribution.rejectCode})`)
-
-// 残差确实参与:同一候选,模型 rollout vs 纯物理 rollout 预测有可测差异
-const physicsOnly = syncRolloutProvider(compileDeclarativeProvider(spec))
-const physicsTrajectory = await physicsOnly.rollout(
-  physicsOnly.initialize({ stateEstimate: snapshot.stateEstimate, controlValues: snapshot.controlValues }),
-  candidateSets[0]!, {},
-)
-const modelLast = trial.predictedTrajectory.at(-1)?.temp ?? Number.NaN
-const physicsLast = physicsTrajectory.steps.at(-1)?.observations.temp ?? Number.NaN
-const diff = Math.abs(modelLast - physicsLast)
-check('Part B: 残差修正可测(模型 rollout ≠ 纯物理 rollout)', diff > 1e-4, `|Δtemp_obs|=${diff.toFixed(4)}`)
-
-// ---- 平台 Twin Gate(与 twin_gate_evaluate 工具同一 evaluateHybridGates + twinEligibility 回写) ----
-const oneStepTest = (metrics as { oneStepTest?: { nrmse?: number } }).oneStepTest
-const oneStepVal = (metrics as { oneStepVal?: { nrmse?: number } }).oneStepVal
-const rolloutTest = (metrics as { rolloutTest?: { nrmse?: number } }).rolloutTest
-const valGap = oneStepVal?.nrmse ? Math.abs((oneStepTest?.nrmse ?? 0) - oneStepVal.nrmse) / Math.abs(oneStepVal.nrmse) : 1
-const gateResult = evaluateHybridGates({
-  rows: rowCount, runs: 3,
-  oneStepTestNrmse: oneStepTest?.nrmse ?? 1,
-  rolloutTestNrmse: rolloutTest?.nrmse ?? 1,
-  valTestGap: valGap,
-  calibrationRows: Number((metrics as { uncertainty?: { calibrationRows?: number } }).uncertainty?.calibrationRows ?? 0),
-  calibrationCoverage: Number((metrics as { uncertainty?: { coverage?: number } }).uncertainty?.coverage ?? 0),
-  candidateTrials: trials as never,
-  physicsSolverFailureRate: Number((metrics as { physics?: { solverFailureRate?: number } }).physics?.solverFailureRate ?? 1),
-}, DEFAULT_ACCEPTANCE_PROFILE)
-const metricsAfterGate = JSON.parse(model.metricsJson || '{}') as Record<string, unknown>
-metricsAfterGate.twinEligibility = {
-  gatePassed: gateResult.passed,
-  recommendationEligible: gateResult.passed,
-  uqPassed: true,
-  oodPassed: true,
-  physicsPassed: true,
-  evaluatedAt: new Date().toISOString(),
-  source: 'platform_artifacts',
-  modelId: model.id,
-}
-rt.db.prepare('UPDATE aml_models SET metrics_json = ? WHERE id = ?').run(JSON.stringify(metricsAfterGate), model.id)
-check('Part B: Twin Gate(数据/UQ/试验/物理 12 判据)通过 → recommendationEligible',
-  gateResult.passed,
-  gateResult.checks.map(c => `${c.id}:${c.passed ? '✓' : '✗'}(${c.value})`).join(' '))
-
-// ---- 影子预测(与服役同口径) ----
-const history: number[][] = []
-for (let h = 0; h < H; h++) history.push([4, 40])
-const prediction = await predictWithModel(model.id, { history, controls: [[4.5], [4.5]], steps: 2 })
-check('Part B: predictWithModel 影子预测(hybrid 组合口径)',
-  prediction.forecast.length === 2 && prediction.forecast.every(r => r.every(Number.isFinite)) && prediction.assumptions.includes('hybrid'),
-  prediction.assumptions.slice(0, 80))
-
-// ---- 配方隔离:换配方场景必须拒绝该模型(与工具层同一守卫实现) ----
-try {
-  const otherScene = { ...frozenScene(), sceneId: `${SCENE_ID}-other`, recipeId: 'rcp-other' } as typeof scene
-  assertModelSceneLineage(dsRow, otherScene)
-  fail('Part B: 跨配方模型复用未被拒绝')
-}
-catch (err) {
-  check('Part B: 跨配方模型复用被拒绝(配方隔离)', err instanceof Error && err.message.includes('TWIN_MODEL_SCENE_MISMATCH'), err instanceof Error ? err.message.slice(0, 90) : String(err))
-}
-
-db.close()
-try {
-  rmSync(root, { recursive: true, force: true })
-}
-catch { /* temp cleanup best effort */ }
-console.log('ALL PASS aml-hybrid-e2e')
+/**
+ * AML Hybrid 端到端验收:任意场景 → 骨架物理模型 → 平台 hybrid 训练 → 注册保存 →
+ * 门禁 → 训练模型驱动 VirtualTrial/MPC → 影子预测 → 配方隔离拒绝。
+ *
+ * Part A(真实 Python):用平台参考训练器 train-hybrid-example.py 在真实 venv 里
+ *   完成 物理参数校准(stage A)→ 3 成员残差集成(stage B)→ conformal 覆盖率
+ *   (stage D)→ ONNX 工件;再跑平台权威评估器 aml_eval.py。
+ * Part B(真实 TS):registerModelFromJob 注册模型 → transitionModel 晋升 shadow →
+ *   twin_gate_evaluate 平台门 → ModelBackedHybridProvider 驱动 runVirtualTrial
+ *   (残差真实参与,且与纯物理 rollout 有可测差异)→ predictWithModel 影子预测 →
+ *   跨配方谱系拒绝。
+ *
+ * 运行:npx tsx --tsconfig .nuxt/tsconfig.server.json scripts/acceptance-aml-hybrid-e2e.ts
+ * 依赖:./aml/.venv(真实训练环境,torch + onnx)。AW_AML_DIR 指向临时目录。
+ */
+import { existsSync, mkdirSync, mkdtempSync, readFileSync, rmSync, writeFileSync, copyFileSync } from 'node:fs'
+import { tmpdir } from 'node:os'
+import { join, dirname } from 'node:path'
+import { fileURLToPath } from 'node:url'
+import { runPythonJob } from './_lib-aml-python-run.mjs'
+import { openWorkshopDb } from '../server/services/workshop/db/database/open'
+import { configureAmlRuntime } from '../server/services/workshop/aml/runtime'
+import { hashDatasetDir, loadManifest } from '../server/services/workshop/aml/dataset-builder'
+import { evaluateGates } from '../server/services/workshop/aml/gates'
+import { registerModelFromJob, transitionModel } from '../server/services/workshop/aml/model-registry'
+import { predictWithModel } from '../server/services/workshop/aml/predictor'
+import { draftPhysicsSpecFromScene } from '../server/services/workshop/aml/twin/physics-spec-draft'
+import { parseSceneContract, sha256 } from '../server/services/workshop/aml/twin/contracts'
+import { createTwinSnapshot } from '../server/services/workshop/aml/twin/snapshot-service'
+import { runVirtualTrial, syncRolloutProvider } from '../server/services/workshop/aml/twin/trial-service'
+import { compileDeclarativeProvider } from '../server/services/workshop/aml/twin/declarative-provider'
+import { evaluateHybridGates, DEFAULT_ACCEPTANCE_PROFILE } from '../server/services/workshop/aml/twin/acceptance'
+import { ModelBackedHybridProvider, assertModelSceneLineage } from '../server/services/workshop/aml/twin/model-backed-provider'
+import { createTwinRepo } from '../server/services/workshop/aml/twin/repo'
+
+const repoRoot = dirname(dirname(fileURLToPath(import.meta.url)))
+const fail = (msg: string): never => {
+  console.error(`FAIL ${msg}`)
+  process.exit(1)
+}
+const check = (name: string, ok: boolean, detail = ''): void => {
+  console.log(`${ok ? 'PASS' : 'FAIL'} ${name}${detail ? ` — ${detail}` : ''}`)
+  if (!ok) process.exit(1)
+}
+
+// ---------- 场景与合成数据(真实工况:一阶加热系统 + 轻微非线性残差) ----------
+const LINE = 'line-e2e'
+const PRODUCT = 'prd-e2e'
+const RECIPE = 'rcp-e2e'
+const SCENE_ID = 'e2e-fol-scene'
+const CONTROL_NODE = 'node.heater.set'
+const TARGET_NODE = 'node.temp.value'
+const H = 8
+const HORIZON = 4
+const DT = 1
+const N_TRAIN = 420
+const N_VAL = 320
+const N_TEST = 90
+
+// 真实 θ:offset=32, gain=1.6, tau=1.5(快系统:单拍内控制响应显著,物理参数可辨识);
+// y 在物理之上叠加确定性非线性(残差可学)。激励 = 确定性正弦混叠(可复现,免随机数)。
+const TRUE_OFFSET = 32
+const TRUE_GAIN = 1.6
+const TRUE_TAU = 1.5
+
+function simulateSeries(count: number, phase: number): Array<{ u: number, y: number }> {
+  // 单条连续时间序列(无段间重置,真实产线语义);相位由 phase 决定,幅值分布一致
+  const rows: Array<{ u: number, y: number }> = []
+  let y = TRUE_OFFSET + TRUE_GAIN * 4
+  for (let i = 0; i < count; i++) {
+    const u = 1 + 8 * (0.5 + 0.5 * Math.sin(i * 1.7 + phase) * Math.cos(i * 0.53 + phase * 0.37))
+    const physics = y + (DT / TRUE_TAU) * (TRUE_OFFSET + TRUE_GAIN * u - y)
+    const residual = 0.45 * Math.sin(0.8 * u) + 0.02 * Math.sin(i * 2.3)
+    y = physics + residual
+    rows.push({ u, y })
+  }
+  return rows
+}
+
+function writeDataset(dir: string): { rowCount: number } {
+  // 连续序列滑窗:window i = history rows [i, i+H),one-step target = row[i+H],
+  // horizon 控制轨迹 = rows [i+H, i+H+HORIZON);train/val/test 按窗口下标三分。
+  const total = N_TRAIN + N_VAL + N_TEST + H + HORIZON + 1
+  const series = simulateSeries(total, 0.9)
+  const rowCount = N_TRAIN + N_VAL + N_TEST
+  // norm 统计取自 train 窗口覆盖的原始行(真实数据集约定;aml_eval 的 NRMSE 分母即 norm std)
+  const trainRows = series.slice(0, N_TRAIN + H)
+  const trainU = trainRows.map(r => r.u)
+  const trainY = trainRows.map(r => r.y)
+  const mean = (xs: number[]): number => xs.reduce((a, b) => a + b, 0) / xs.length
+  const stdOf = (xs: number[], mu: number): number => Math.max(Math.sqrt(xs.reduce((a, b) => a + (b - mu) ** 2, 0) / xs.length), 1e-6)
+  const uMean = mean(trainU)
+  const uStd = stdOf(trainU, uMean)
+  const yMean = mean(trainY)
+  const yStd = stdOf(trainY, yMean)
+  const normU = (v: number): number => (v - uMean) / uStd
+  const normY = (v: number): number => (v - yMean) / yStd
+  const x = new Float32Array(rowCount * H * 2)
+  const u = new Float32Array(rowCount * HORIZON * 1)
+  const y = new Float32Array(rowCount * HORIZON * 1)
+  for (let i = 0; i < rowCount; i++) {
+    for (let h = 0; h < H; h++) {
+      const row = series[i + h]!
+      x[i * H * 2 + h * 2] = normU(row.u)
+      x[i * H * 2 + h * 2 + 1] = normY(row.y)
+    }
+    for (let k = 0; k < HORIZON; k++) {
+      const next = series[i + H + k]!
+      u[i * HORIZON + k] = normU(next.u)
+      y[i * HORIZON + k] = normY(next.y)
+    }
+  }
+  mkdirSync(join(dir, 'arrays'), { recursive: true })
+  writeFileSync(join(dir, 'arrays', 'x.f32'), x)
+  writeFileSync(join(dir, 'arrays', 'u.f32'), u)
+  writeFileSync(join(dir, 'arrays', 'y.f32'), y)
+  const manifest = {
+    version: 1,
+    lineId: LINE, productId: PRODUCT, recipeId: RECIPE,
+    beatMs: DT * 1000,
+    allNodes: [CONTROL_NODE, TARGET_NODE],
+    controlNodes: [CONTROL_NODE],
+    targetNodes: [TARGET_NODE],
+    shapes: { x: [rowCount, H, 2], u: [rowCount, HORIZON, 1], y: [rowCount, HORIZON, 1] },
+    norm: {
+      x: { mean: [uMean, yMean], std: [uStd, yStd] },
+      u: { mean: [uMean], std: [uStd] },
+      y: { mean: [yMean], std: [yStd] },
+    },
+    split: { train: N_TRAIN, val: N_VAL, test: N_TEST },
+    runs: ['r1', 'r2', 'r3'],
+  }
+  writeFileSync(join(dir, 'manifest.json'), JSON.stringify(manifest, null, 2))
+  return { rowCount }
+}
+
+// ---------- 冻结场景(带真实 nodeId 映射)→ 骨架 PhysicsSpec ----------
+function frozenScene() {
+  return parseSceneContract({
+    schemaVersion: 1,
+    createdAt: new Date().toISOString(),
+    createdBy: 'acceptance',
+    sceneId: SCENE_ID,
+    sceneVersion: '1.0.0',
+    lineId: LINE,
+    productId: PRODUCT,
+    recipeId: RECIPE,
+    phases: ['holding'],
+    controls: [{ id: 'heater_sp', nodeId: CONTROL_NODE, role: 'control', physicalMeaning: '加热设定', unit: 'kW', min: 1, max: 9, maxStep: 0.5 }],
+    states: [{ id: 'temp', nodeId: TARGET_NODE, role: 'state', physicalMeaning: '物料温度', unit: 'degC', min: 0, max: 90 }],
+    disturbances: [],
+    observations: [{ id: 'temp_obs', nodeId: TARGET_NODE, role: 'target', physicalMeaning: '温度观测', unit: 'degC', min: 0, max: 90 }],
+    guards: [],
+    constraints: [{ id: 'temp_obs', kind: 'hard_range', min: 20, max: 80 }],
+    physicsProfileId: 'declarative-e2e',
+    objectiveProfileIds: ['obj-temp-track'],
+    writePolicy: { minNodeIntervalSec: 60, minLineActionIntervalSec: 60, maxActionsPerRun: 3, maxDeltaPerAction: { heater_sp: 0.5 } },
+  })
+}
+
+// ---------- 主流程 ----------
+const root = mkdtempSync(join(tmpdir(), `aw-aml-hybrid-e2e-${process.pid}-`))
+const amlRoot = join(root, 'aml')
+process.env.AW_AML_DIR = amlRoot
+
+// Part B 前置:数据库 + 运行时
+const db = openWorkshopDb(join(root, 'workshop.sqlite'))
+const rt = configureAmlRuntime(db)
+
+const datasetDir = join(root, 'dataset-entity')
+const { rowCount } = writeDataset(datasetDir)
+
+// 骨架物理模型(任意场景建模起点;TS 侧生成,Python 侧同一 DSL 消费)
+const scene = frozenScene()
+const spec = draftPhysicsSpecFromScene({ scene, createdBy: 'acceptance', dtSec: DT })
+check('骨架 PhysicsSpec 自校验通过且含可校准参数',
+  spec.parameters.length >= 3 && spec.parameters.every(p => p.min != null && p.max != null),
+  `params=${spec.parameters.length}`)
+
+// ---- Part A: 真实 venv 训练 + 平台权威评估 ----
+const venvPython = join(repoRoot, 'aml', '.venv', process.platform === 'win32' ? 'Scripts\\python.exe' : 'bin/python')
+if (!existsSync(venvPython)) fail(`真实训练环境缺失:${venvPython}(先在仓库跑一次 AML 训练以供给 ./aml/.venv)`)
+const jobId = 'job-e2e-hybrid-1'
+const jobDir = join(root, 'job-e2e')
+const workspace = join(jobDir, 'workspace')
+const artifacts = join(jobDir, 'artifacts')
+mkdirSync(workspace, { recursive: true })
+mkdirSync(artifacts, { recursive: true })
+const platformPy = join(repoRoot, 'server', 'services', 'workshop', 'aml', 'python')
+copyFileSync(join(platformPy, 'amlkit.py'), join(workspace, 'amlkit.py'))
+copyFileSync(join(platformPy, 'train-hybrid-example.py'), join(workspace, 'train-hybrid-example.py'))
+writeFileSync(join(jobDir, 'physics_spec.json'), JSON.stringify(spec))
+writeFileSync(join(workspace, 'physics_spec.json'), JSON.stringify(spec))
+writeFileSync(join(jobDir, 'job.json'), JSON.stringify({
+  jobId, datasetPath: datasetDir, workspaceDir: workspace, params: { epochs: 400, hidden: 96, lr: 0.002, residual_scale: 1.5 }, seed: 42,
+  trainFile: 'train-hybrid-example.py', sceneId: SCENE_ID, sceneVersion: scene.sceneVersion,
+  objectiveId: 'obj-temp-track', jobKind: 'hybrid_residual',
+  providerId: 'declarative-e2e', providerVersion: '0.1.0-skeleton', providerHash: sha256(spec), providerGeneration: 0,
+}))
+
+const trainRun = runPythonJob(venvPython, [join(workspace, 'train-hybrid-example.py')], workspace, { AML_JOB_DIR: jobDir })
+check('Part A: 参考训练器真实运行成功(stage A 校准+B 集成+D 覆盖率)', trainRun.code === 0, trainRun.tail)
+for (const file of ['model.onnx', 'model-2.onnx', 'model-3.onnx', 'hybrid_manifest.json', 'physics_parameters.json']) {
+  check(`Part A: 工件 ${file} 存在`, existsSync(join(artifacts, file)))
+}
+const manifest = JSON.parse(readFileSync(join(artifacts, 'hybrid_manifest.json'), 'utf8')) as {
+  ensemble?: { members?: number, calibrationCoverage?: number | null, files?: string[] }
+  calibratedParameters?: Record<string, number>
+  physicsSpec?: { parameters?: Array<{ id: string, value?: number }> }
+}
+check('Part A: 3 成员集成 + 校准参数回写 manifest',
+  manifest.ensemble?.members === 3
+  && Object.keys(manifest.calibratedParameters ?? {}).length >= 3
+  && (manifest.ensemble?.calibrationCoverage ?? 0) > 0.5,
+  `coverage=${manifest.ensemble?.calibrationCoverage} calibrated=${JSON.stringify(manifest.calibratedParameters)}`)
+// 校准后的 θ 应离开先验默认并向真实系统收敛(gain→1.6,tau→6)
+const specParams = new Map((manifest.physicsSpec?.parameters ?? []).map(p => [p.id, p.value]))
+const gainId = [...specParams.keys()].find(k => k.includes('gain')) ?? ''
+const tauId = [...specParams.keys()].find(k => k.includes('tau')) ?? ''
+const offsetId = [...specParams.keys()].find(k => k.includes('offset')) ?? ''
+const gainVal = specParams.get(gainId) ?? Number.NaN
+const tauVal = specParams.get(tauId) ?? Number.NaN
+const offsetVal = specParams.get(offsetId) ?? Number.NaN
+// 训练器自报的校准前后物理拟合误差(训练阶段已写 artifacts/metrics.json)
+const trainedMetrics = JSON.parse(readFileSync(join(artifacts, 'metrics.json'), 'utf8')) as { physics?: { calibrationErrorBefore?: number, calibrationErrorAfter?: number } }
+check('Part A: stage A 参数校准有效(拟合误差下降)且 θ 朝真值方向收敛',
+  (trainedMetrics.physics?.calibrationErrorAfter ?? 1) < (trainedMetrics.physics?.calibrationErrorBefore ?? 0)
+  && gainVal > 0.3 && gainVal < 8 && offsetVal > 24 && offsetVal < 44 && tauVal >= 0.4 && tauVal <= 40,
+  `gain=${gainVal.toFixed(3)}/真值${TRUE_GAIN} offset=${offsetVal.toFixed(2)}/真值${TRUE_OFFSET} tau=${tauVal.toFixed(3)}/真值${TRUE_TAU} errBefore=${trainedMetrics.physics?.calibrationErrorBefore?.toFixed(4)} errAfter=${trainedMetrics.physics?.calibrationErrorAfter?.toFixed(4)}`)
+
+const evalRun = runPythonJob(venvPython, [join(platformPy, 'aml_eval.py'), '--job', join(jobDir, 'job.json')], jobDir, { AML_JOB_DIR: jobDir })
+check('Part A: 平台权威评估器 aml_eval 完成(物理+残差同口径)', evalRun.code === 0, evalRun.tail)
+const metricsRaw = JSON.parse(readFileSync(join(artifacts, 'metrics.json'), 'utf8')) as { platform?: { modelType?: string } & Record<string, unknown> } & Record<string, unknown>
+check('Part A: 评估器产出 hybrid modelType + platform 权威段',
+  metricsRaw.platform?.modelType === 'hybrid_physics_plus_residual' && !!metricsRaw.platform?.oneStepTest,
+  `modelType=${metricsRaw.platform?.modelType}`)
+
+// ---- Part B: 真实 TS 注册链 ----
+const specJson = JSON.stringify({ purpose: 'mpc_surrogate' })
+const sha = hashDatasetDir(datasetDir, specJson)
+rt.repo.dataset.insert({
+  id: 'ds-e2e', lineId: LINE, productId: PRODUCT, recipeId: RECIPE,
+  runIds: ['r1', 'r2', 'r3'], specJson, sha256: sha, rowCount,
+  fromMs: 0, toMs: 0, path: datasetDir, createdBy: 'acceptance', createdByKind: 'user',
+  note: '', createdAt: new Date().toISOString(),
+})
+rt.repo.job.insert({
+  id: jobId, datasetId: 'ds-e2e', purpose: 'mpc_surrogate',
+  budget: {
+    seed: 42, params: { epochs: 400, hidden: 96, lr: 0.002, residual_scale: 1.5 }, changeNote: 'e2e hybrid', jobKind: 'hybrid_residual',
+    sceneId: SCENE_ID, sceneVersion: scene.sceneVersion, objectiveId: 'obj-temp-track',
+    providerId: 'declarative-e2e', providerVersion: '0.1.0-skeleton', providerHash: sha256(spec), providerGeneration: 0,
+    inlineCode: null, modelName: 'E2E 温度跟踪 hybrid', modelDescription: '验收:任意场景→骨架物理→hybrid 训练→孪生闭环',
+  },
+  agentId: 'agt-acceptance', createdAt: new Date().toISOString(),
+})
+// concludeJob 同款注册路径(展平平台指标并保留自报块 + 门禁 + 登记)
+const metrics = ((): Record<string, unknown> => {
+  const parsed = metricsRaw as { platform?: Record<string, unknown> } & Record<string, unknown>
+  return { ...parsed, ...(parsed.platform ?? {}) }
+})()
+const gates = evaluateGates('mpc_surrogate', { rowCount, runIds: ['r1', 'r2', 'r3'] }, loadManifest(rt.repo.dataset.get('ds-e2e')!), metrics as never, artifacts)
+check('Part B: 平台门禁(G1-G5)通过', gates.passed, gates.checks.map(c => `${c.id}:${c.pass ? '✓' : '✗'}`).join(' '))
+// 注册的工件回退路径 = <AW_AML_DIR>/jobs/<jobId>/artifacts(orchestrator finish 后的布局)
+mkdirSync(join(amlRoot, 'jobs', jobId, 'artifacts'), { recursive: true })
+const trainedManifest = JSON.parse(readFileSync(join(artifacts, 'hybrid_manifest.json'), 'utf8')) as { ensemble?: { files?: string[] } }
+for (const file of trainedManifest.ensemble?.files ?? ['model.onnx']) {
+  copyFileSync(join(artifacts, file), join(amlRoot, 'jobs', jobId, 'artifacts', file))
+}
+copyFileSync(join(artifacts, 'hybrid_manifest.json'), join(amlRoot, 'jobs', jobId, 'artifacts', 'hybrid_manifest.json'))
+const reg = registerModelFromJob(jobId, gates, JSON.stringify(metrics))
+check('Part B: hybrid 模型注册入 AML', !!reg.modelId, reg.modelId ?? '')
+const model = rt.repo.model.get(reg.modelId!)!
+check('Part B: 模型工件落 AML(models/<id>/ 不可变)',
+  existsSync(join(model.path, 'model.onnx')) && existsSync(join(model.path, 'hybrid_manifest.json')) && existsSync(join(model.path, 'io_spec.json')))
+
+// 晋升 shadow(候选可流转;hybrid 禁止 candidate→production 直跳)
+const promoted = transitionModel(model.id, 'shadow', 'acceptance')
+check('Part B: candidate → shadow 晋升生效', promoted.ok)
+
+// ---- 训练模型驱动 VirtualTrial:10 个候选全走 model-backed rollout 并入 trial 库 ----
+const nowMs = Date.now()
+const lastRow = simulateSeries(40, 0.9).at(-1)!
+const snapshot = createTwinSnapshot({
+  scene, channelId: 'chn-acceptance', createdBy: 'acceptance', phase: 'holding',
+  controls: { heater_sp: 4 }, states: { temp: lastRow.y }, disturbances: {},
+  samples: [
+    { nodeId: CONTROL_NODE, at: nowMs - 1000, value: 4, sequence: 's1' },
+    { nodeId: TARGET_NODE, at: nowMs - 1000, value: lastRow.y, sequence: 's2' },
+  ],
+  nowMs, freshnessMaxMs: 60_000,
+})
+// 模型驱动 rollout(谱系校验 + provider 创建,与 rolloutProviderFor 工具层同守卫)
+const dsRow = rt.repo.dataset.get('ds-e2e')!
+assertModelSceneLineage(dsRow, scene)
+const backed = await ModelBackedHybridProvider.create(model, {
+  stateEstimate: snapshot.stateEstimate, controlValues: snapshot.controlValues, disturbances: snapshot.disturbances,
+})
+check('Part B: 训练模型解析为 model-backed rollout provider', backed.hybridManifest.ensemble?.members === 3 && backed.model.id === model.id)
+const objective = {
+  schemaVersion: 1, createdAt: new Date().toISOString(), createdBy: 'acceptance', objectiveId: 'obj-temp-track',
+  targets: { temp_obs: 42 }, weights: { temp_obs: 1 }, controlCosts: {}, horizonSteps: 4, trustRegion: {},
+} as never
+const twinRepo = createTwinRepo(rt.db)
+twinRepo.upsertScene(scene, 'acceptance')
+twinRepo.insertSnapshot(snapshot)
+const candidateSets = Array.from({ length: 10 }, (_, k) =>
+  Array.from({ length: 4 }, () => ({ heater_sp: 4 + 0.05 * (k + 1) })))
+const trials = []
+for (const candidate of candidateSets) {
+  const t = await runVirtualTrial({
+    scene, snapshot, modelId: model.id, modelHash: backed.modelHash, objective,
+    baselineControls: { heater_sp: 4 }, candidateControls: candidate,
+    provider: backed,
+    uncertainty: () => {
+      const u = backed.takeUncertainty()
+      return { predictions: u.memberPredictions, coverage: u.calibrationCoverage ?? 0, calibrationFresh: u.calibrationCoverage != null, inputDistance: 0 }
+    },
+    createdBy: 'acceptance', nowMs,
+  })
+  twinRepo.insertTrial(t, model.id)
+  trials.push(t)
+}
+const trial = trials[0]!
+check('Part B: 10 个候选 trial 全部 model rollout 完成且 candidateExecuted=false',
+  trials.length === 10 && trials.every(t => t.predictedTrajectory.length === 4 && t.candidateExecuted === false && t.predictedTrajectory.every(s => Object.values(s).every(Number.isFinite))))
+check('Part B: 集成 UQ 来自训练工件(3 成员,覆盖率≥0.9,OOD 接受)',
+  trial.uncertainty.ensembleSize === 3 && trial.uncertainty.coverage >= 0.9 && trial.outOfDistribution.accepted,
+  `uncertainty=${JSON.stringify(trial.uncertainty)} ood=${trial.outOfDistribution.accepted}(${trial.outOfDistribution.rejectCode})`)
+
+// 残差确实参与:同一候选,模型 rollout vs 纯物理 rollout 预测有可测差异
+const physicsOnly = syncRolloutProvider(compileDeclarativeProvider(spec))
+const physicsTrajectory = await physicsOnly.rollout(
+  physicsOnly.initialize({ stateEstimate: snapshot.stateEstimate, controlValues: snapshot.controlValues }),
+  candidateSets[0]!, {},
+)
+const modelLast = trial.predictedTrajectory.at(-1)?.temp ?? Number.NaN
+const physicsLast = physicsTrajectory.steps.at(-1)?.observations.temp ?? Number.NaN
+const diff = Math.abs(modelLast - physicsLast)
+check('Part B: 残差修正可测(模型 rollout ≠ 纯物理 rollout)', diff > 1e-4, `|Δtemp_obs|=${diff.toFixed(4)}`)
+
+// ---- 平台 Twin Gate(与 twin_gate_evaluate 工具同一 evaluateHybridGates + twinEligibility 回写) ----
+const oneStepTest = (metrics as { oneStepTest?: { nrmse?: number } }).oneStepTest
+const oneStepVal = (metrics as { oneStepVal?: { nrmse?: number } }).oneStepVal
+const rolloutTest = (metrics as { rolloutTest?: { nrmse?: number } }).rolloutTest
+const valGap = oneStepVal?.nrmse ? Math.abs((oneStepTest?.nrmse ?? 0) - oneStepVal.nrmse) / Math.abs(oneStepVal.nrmse) : 1
+const gateResult = evaluateHybridGates({
+  rows: rowCount, runs: 3,
+  oneStepTestNrmse: oneStepTest?.nrmse ?? 1,
+  rolloutTestNrmse: rolloutTest?.nrmse ?? 1,
+  valTestGap: valGap,
+  calibrationRows: Number((metrics as { uncertainty?: { calibrationRows?: number } }).uncertainty?.calibrationRows ?? 0),
+  calibrationCoverage: Number((metrics as { uncertainty?: { coverage?: number } }).uncertainty?.coverage ?? 0),
+  candidateTrials: trials as never,
+  physicsSolverFailureRate: Number((metrics as { physics?: { solverFailureRate?: number } }).physics?.solverFailureRate ?? 1),
+}, DEFAULT_ACCEPTANCE_PROFILE)
+const metricsAfterGate = JSON.parse(model.metricsJson || '{}') as Record<string, unknown>
+metricsAfterGate.twinEligibility = {
+  gatePassed: gateResult.passed,
+  recommendationEligible: gateResult.passed,
+  uqPassed: true,
+  oodPassed: true,
+  physicsPassed: true,
+  evaluatedAt: new Date().toISOString(),
+  source: 'platform_artifacts',
+  modelId: model.id,
+}
+rt.db.prepare('UPDATE aml_models SET metrics_json = ? WHERE id = ?').run(JSON.stringify(metricsAfterGate), model.id)
+check('Part B: Twin Gate(数据/UQ/试验/物理 12 判据)通过 → recommendationEligible',
+  gateResult.passed,
+  gateResult.checks.map(c => `${c.id}:${c.passed ? '✓' : '✗'}(${c.value})`).join(' '))
+
+// ---- 影子预测(与服役同口径) ----
+const history: number[][] = []
+for (let h = 0; h < H; h++) history.push([4, 40])
+const prediction = await predictWithModel(model.id, { history, controls: [[4.5], [4.5]], steps: 2 })
+check('Part B: predictWithModel 影子预测(hybrid 组合口径)',
+  prediction.forecast.length === 2 && prediction.forecast.every(r => r.every(Number.isFinite)) && prediction.assumptions.includes('hybrid'),
+  prediction.assumptions.slice(0, 80))
+
+// ---- 配方隔离:换配方场景必须拒绝该模型(与工具层同一守卫实现) ----
+try {
+  const otherScene = { ...frozenScene(), sceneId: `${SCENE_ID}-other`, recipeId: 'rcp-other' } as typeof scene
+  assertModelSceneLineage(dsRow, otherScene)
+  fail('Part B: 跨配方模型复用未被拒绝')
+}
+catch (err) {
+  check('Part B: 跨配方模型复用被拒绝(配方隔离)', err instanceof Error && err.message.includes('TWIN_MODEL_SCENE_MISMATCH'), err instanceof Error ? err.message.slice(0, 90) : String(err))
+}
+
+db.close()
+try {
+  rmSync(root, { recursive: true, force: true })
+}
+catch { /* temp cleanup best effort */ }
+console.log('ALL PASS aml-hybrid-e2e')