agentworkshop 0.8.6 → 0.8.7

package/server/services/workshop/aml/twin/model-backed-provider.ts

@@ -1,332 +1,332 @@
-/**
- * Model-backed rollout provider:让**训练注册后的 hybrid 模型**直接驱动 VirtualTrial/MPC。
- *
- * 组成 = hybrid_manifest.json 里冻结(含校准后 θ)的 PhysicsSpec 声明式物理主干
- * + 残差集成 ONNX(默认 3 成员,平台 UQ 依据)。物理先走、残差只在 residualScale
- * 边界内修正目标观测(归一化域,与 predictor 服役路径同一数学);历史窗口按
- * persistence 语义维护(与 io_spec.assumptions 一致)。
- *
- * 这是"训练出的模型进入孪生闭环"的落实点:此前 trial/MPC 只拿 physics-only
- * provider,model_id 只是标签;现在 model_id 背后的真实工件参与每一次 rollout。
- */
-import { existsSync, readFileSync } from 'node:fs'
-import { join } from 'node:path'
-import { AppError } from '../../../../utils/errors'
-import { parseIoSpec } from '../model-registry'
-import { onnxSessionFor, ortRuntime, type IoSpec, type OrtSession } from '../predictor'
-import { compileDeclarativeProvider } from './declarative-provider'
-import { parsePhysicsSpec, type PhysicsSpec } from './physics-spec'
-import type { AmlModelRow } from '../aml.repo'
-import type { PhysicsState, PhysicsStepResult, PhysicsTrajectory } from './physics-runtime'
-import type { RolloutProvider } from './trial-service'
-
-interface HybridManifest {
-  physicsSpec?: unknown
-  residualScale?: number
-  composition?: string
-  physicsSpecHash?: string
-  ensemble?: {
-    members?: number
-    files?: string[]
-    calibrationCoverage?: number | null
-    coverageTarget?: number
-    calibrationRows?: number | null
-  }
-}
-
-export interface ModelBackedUncertainty {
-  /** 每步各目标归一化域的集成间最大标准差(与 residualScale 同量纲) */
-  maxStdByStep: number[]
-  /** 各成员最后一拍的目标预测(原始物理量纲;相对分歧对量纲不敏感) */
-  memberPredictions: Array<Record<string, number>>
-  calibrationCoverage: number | null
-  coverageTarget: number
-  members: number
-}
-
-export interface ModelSnapshotValues {
-  stateEstimate: Record<string, number>
-  controlValues: Record<string, number>
-  disturbances: Record<string, number>
-}
-
-/**
- * 配方隔离守卫:模型的训练数据集 line/product/recipe 必须与场景谱系一致,
- * 不同配方的模型禁止驱动彼此的孪生 rollout(单一实现,工具层与验收共用)。
- */
-export function assertModelSceneLineage(
-  dataset: { lineId: string, productId: string, recipeId: string },
-  scene: { lineId?: string, productId?: string, recipeId?: string },
-): void {
-  const mismatch: string[] = []
-  if (scene.lineId && dataset.lineId !== scene.lineId) mismatch.push(`line ${dataset.lineId}≠${scene.lineId}`)
-  if (scene.productId && dataset.productId !== scene.productId) mismatch.push(`product ${dataset.productId}≠${scene.productId}`)
-  if (scene.recipeId && dataset.recipeId !== scene.recipeId) mismatch.push(`recipe ${dataset.recipeId}≠${scene.recipeId}`)
-  if (mismatch.length > 0) {
-    throw new AppError(409, 'TWIN_MODEL_SCENE_MISMATCH', `TWIN_MODEL_SCENE_MISMATCH:${mismatch.join(';')}(模型与场景/配方必须同谱系,禁止跨配方复用)`)
-  }
-}
-
-type BaseProvider = ReturnType<typeof compileDeclarativeProvider>
-
-export class ModelBackedHybridProvider implements RolloutProvider {
-  readonly base: BaseProvider
-  readonly spec: PhysicsSpec
-  readonly io: IoSpec
-  readonly hybridManifest: HybridManifest
-  readonly model: AmlModelRow
-  private readonly sessions: OrtSession[]
-  private readonly aliases: Map<string, string>
-  /** 训练期控制均值(按 varId);baseline 未提供全部控制输入时的回退先验,防 spec 求值 NaN */
-  private readonly defaultControls: Record<string, number>
-  private readonly residualScale: number
-  private rawHistory: number[][] = []
-  private readonly uncertainty: ModelBackedUncertainty
-
-  private constructor(
-    model: AmlModelRow,
-    manifest: HybridManifest,
-    sessions: OrtSession[],
-    snapshot: ModelSnapshotValues,
-  ) {
-    this.model = model
-    this.hybridManifest = manifest
-    if (!manifest.physicsSpec) throw new AppError(409, 'AML_HYBRID_PHYSICS_MISSING', `模型 ${model.id} 缺少冻结 PhysicsSpec,拒绝把 residual 当完整预测`)
-    this.spec = parsePhysicsSpec(manifest.physicsSpec)
-    this.io = parseIoSpec(model) as unknown as IoSpec
-    this.base = compileDeclarativeProvider(this.spec)
-    this.residualScale = Math.abs(Number(manifest.residualScale ?? 0.25))
-    this.aliases = new Map(this.spec.variables.filter(v => v.nodeId).map(v => [v.nodeId as string, v.id]))
-    this.defaultControls = Object.fromEntries(this.io.controlNodes.map((node, j) => {
-      const varId = this.aliases.get(node) ?? node
-      return [varId, this.io.norm.u.mean[j] ?? 0]
-    }))
-    this.sessions = sessions
-    this.uncertainty = {
-      maxStdByStep: [],
-      memberPredictions: [],
-      calibrationCoverage: manifest.ensemble?.calibrationCoverage ?? null,
-      coverageTarget: manifest.ensemble?.coverageTarget ?? 0.9,
-      members: manifest.ensemble?.members ?? sessions.length,
-    }
-    this.primeHistory(snapshot)
-  }
-
-  /** 工厂:加载 manifest + 残差集成会话(会话经 predictor LRU 复用) */
-  static async create(model: AmlModelRow, snapshot: ModelSnapshotValues): Promise<ModelBackedHybridProvider> {
-    const manifestPath = join(model.path, 'hybrid_manifest.json')
-    if (!existsSync(manifestPath)) {
-      throw new AppError(409, 'AML_MODEL_NOT_HYBRID', `模型 ${model.id} 不是 hybrid 工件(缺 hybrid_manifest.json),不能作为孪生 rollout 模型`)
-    }
-    let manifest: HybridManifest
-    try {
-      manifest = JSON.parse(readFileSync(manifestPath, 'utf8')) as HybridManifest
-    }
-    catch {
-      throw new AppError(409, 'AML_HYBRID_MANIFEST_INVALID', `模型 ${model.id} 的 hybrid_manifest.json 无法解析`)
-    }
-    if (!ortRuntime()) throw new AppError(503, 'AML_PREDICTOR_UNAVAILABLE', '预测服务依赖 onnxruntime-node 未安装:执行 pnpm install(可选依赖)后重启')
-    const files = manifest.ensemble?.files?.length ? manifest.ensemble.files : ['model.onnx']
-    const sessions: OrtSession[] = []
-    for (const [index, file] of files.entries()) {
-      const path = join(model.path, file)
-      if (!existsSync(path)) {
-        if (index === 0) throw new AppError(410, 'AML_ARTIFACT_MISSING', `模型 ${model.id} 残差工件缺失(${file})`)
-        continue
-      }
-      sessions.push(await onnxSessionFor(path, `${model.id}::${file}`))
-    }
-    return new ModelBackedHybridProvider(model, manifest, sessions, snapshot)
-  }
-
-  private valueForNode(node: string, snapshot: ModelSnapshotValues): number {
-    const varId = this.aliases.get(node) ?? node
-    const variable = this.spec.variables.find(v => v.id === varId)
-    return snapshot.stateEstimate[varId]
-      ?? snapshot.stateEstimate[node]
-      ?? snapshot.controlValues[varId]
-      ?? snapshot.controlValues[node]
-      ?? snapshot.disturbances[varId]
-      ?? snapshot.disturbances[node]
-      ?? variable?.defaultValue
-      ?? variable?.min
-      ?? 0
-  }
-
-  private primeHistory(snapshot: ModelSnapshotValues): void {
-    const row = this.io.allNodes.map(node => this.valueForNode(node, snapshot))
-    this.rawHistory = Array.from({ length: this.io.historySteps }, () => [...row])
-  }
-
-  private normalizeWindow(): number[][] {
-    const mx = this.io.norm.x.mean
-    const sx = this.io.norm.x.std
-    return this.rawHistory.map(row => row.map((v, j) => (v - (mx[j] ?? 0)) / (sx[j] ?? 1)))
-  }
-
-  private normalizeControls(controls: Record<string, number>, lastRow: number[]): number[] {
-    return this.io.controlNodes.map((node, j) => {
-      const varId = this.aliases.get(node) ?? node
-      const raw = controls[varId] ?? controls[node] ?? lastRow[this.io.allNodes.indexOf(node)] ?? 0
-      return (raw - (this.io.norm.u.mean[j] ?? 0)) / (this.io.norm.u.std[j] ?? 1)
-    })
-  }
-
-  /** 残差集成前向:归一化窗口(末拍控制列替换为本拍控制,与 predictor 同口径)→ 各成员 residual */
-  private async residualForward(historyNorm: number[][], controlsNorm: number[]): Promise<{ mean: number[], std: number[], members: number[][] }> {
-    const H = this.io.historySteps
-    const nAll = this.io.allNodes.length
-    // 末拍控制列替换为本拍控制(predictor 服役路径同语义;否则残差看到的是上一拍控制)
-    const window = historyNorm.map(row => [...row])
-    const last = window[H - 1]
-    if (last) {
-      this.io.controlNodes.forEach((node, c) => {
-        const j = this.io.allNodes.indexOf(node)
-        if (j >= 0) last[j] = controlsNorm[c] ?? last[j]!
-      })
-    }
-    const flat = new Float32Array(H * nAll)
-    for (let i = 0; i < H; i++) {
-      window[i]?.forEach((v, j) => {
-        flat[i * nAll + j] = v
-      })
-    }
-    const ort = ortRuntime()
-    if (!ort) throw new AppError(503, 'AML_PREDICTOR_UNAVAILABLE', 'onnxruntime-node 不可用')
-    const nTgt = this.io.targetNodes.length
-    const perMember: number[][] = []
-    for (const session of this.sessions) {
-      const inpName = session.inputNames[0]
-      if (!inpName) throw new AppError(503, 'AML_PREDICTOR_UNAVAILABLE', `残差 session 无输入名(模型 ${this.model.id})`)
-      const outName = session.outputNames[0] ?? inpName
-      const out = await session.run({ [inpName]: new ort.Tensor('float32', flat, [1, H, nAll]) })
-      const data = Array.from((out[outName] as { data: ArrayLike<number> }).data) as number[]
-      perMember.push(data.slice(0, nTgt))
-    }
-    const mean: number[] = []
-    const std: number[] = []
-    for (let t = 0; t < nTgt; t++) {
-      const series = perMember.map(row => row[t] ?? 0)
-      const mu = series.reduce((a, b) => a + b, 0) / Math.max(1, series.length)
-      const variance = series.length > 1 ? series.reduce((a, b) => a + (b - mu) ** 2, 0) / series.length : 0
-      mean.push(mu)
-      std.push(Math.sqrt(variance))
-    }
-    return { mean, std, members: perMember }
-  }
-
-  initialize(snapshot: { stateEstimate: Record<string, number>, controlValues: Record<string, number> }): PhysicsState {
-    return this.base.initialize(snapshot)
-  }
-
-  /** 模型驱动 rollout:物理主干 → 有界残差修正目标 → 历史窗口推进 */
-  async rollout(initial: PhysicsState, controls: Array<Record<string, number>>, disturbances: Record<string, number> = {}): Promise<PhysicsTrajectory> {
-    void disturbances
-    let state = { ...initial }
-    const steps: PhysicsStepResult[] = []
-    const failures: string[] = []
-    for (const control of controls) {
-      // 控制输入并集:baseline/candidate 常只含被优化变量,物理方程引用其余控制时回退训练均值(否则 ref miss → NaN)
-      const physics = this.base.step({ state, controls: { ...this.defaultControls, ...control }, disturbances, dtSec: this.spec.solver.dtSec })
-      const lastRow = this.rawHistory[this.rawHistory.length - 1] ?? []
-      // 目标物理值 → 归一化 → 加有界残差(与 predictor 服役路径同一数学)
-      const my = this.io.norm.y.mean
-      const sy = this.io.norm.y.std
-      const yPhysNorm = this.io.targetNodes.map((node, t) => {
-        const varId = this.aliases.get(node) ?? node
-        const raw: number | undefined = physics.observations[varId] ?? physics.state[varId]
-        if (raw === undefined || !Number.isFinite(raw)) {
-          throw new AppError(409, 'AML_HYBRID_PHYSICS_OUTPUT_MISSING', `PhysicsSpec 未生成目标 ${node}(模型 ${this.model.id};常见原因:spec 方程求值得到非有限值 —— 检查 spec 引用的控制/状态变量是否都有输入)`)
-        }
-        return (raw - (my[t] ?? 0)) / (sy[t] ?? 1)
-      })
-      const controlsNorm = this.normalizeControls(control, lastRow)
-      const { mean, std, members } = await this.residualForward(this.normalizeWindow(), controlsNorm)
-      const boundedOf = (value: number | undefined): number => Math.max(-this.residualScale, Math.min(this.residualScale, value ?? 0))
-      const correctedRaw = this.io.targetNodes.map((node, t) => (yPhysNorm[t]! + boundedOf(mean[t])) * (sy[t] ?? 1) + (my[t] ?? 0))
-      const observations = { ...physics.observations }
-      this.io.targetNodes.forEach((node, t) => {
-        const varId = this.aliases.get(node) ?? node
-        observations[varId] = correctedRaw[t]!
-      })
-      // 历史窗口推进(persistence:feature 持最后观测;control 取本拍;target 回填修正值;state 回填演化)
-      const nextRow = [...lastRow]
-      this.io.controlNodes.forEach((node) => {
-        const pos = this.io.allNodes.indexOf(node)
-        const varId = this.aliases.get(node) ?? node
-        if (pos >= 0) nextRow[pos] = control[varId] ?? control[node] ?? nextRow[pos] ?? 0
-      })
-      this.io.targetNodes.forEach((node, t) => {
-        const pos = this.io.allNodes.indexOf(node)
-        if (pos >= 0) nextRow[pos] = correctedRaw[t]!
-      })
-      for (const variable of this.spec.variables) {
-        if (variable.role !== 'state') continue
-        const pos = this.io.allNodes.indexOf(variable.nodeId ?? variable.id)
-        if (pos >= 0) nextRow[pos] = physics.state[variable.id] ?? nextRow[pos] ?? 0
-      }
-      this.rawHistory.push(nextRow)
-      if (this.rawHistory.length > this.io.historySteps) this.rawHistory.shift()
-      this.uncertainty.maxStdByStep.push(Math.max(...std, 0))
-      // 每成员"本拍目标会是什么"(原始量纲,各自残差而非均值)——相对分歧评估的输入
-      this.uncertainty.memberPredictions = members.map((row) => {
-        const entry: Record<string, number> = {}
-        this.io.targetNodes.forEach((node, t) => {
-          entry[node] = (yPhysNorm[t]! + boundedOf(row[t])) * (sy[t] ?? 1) + (my[t] ?? 0)
-        })
-        return entry
-      })
-      const result: PhysicsStepResult = { state: physics.state, observations, guards: physics.guards }
-      if ([...Object.values(result.state), ...Object.values(result.observations), ...Object.values(result.guards)].some(value => !Number.isFinite(value))) failures.push('HYBRID_NON_FINITE')
-      for (const variable of this.spec.variables) {
-        const value = result.state[variable.id] ?? result.observations[variable.id] ?? result.guards[variable.id]
-        if (value !== undefined && ((variable.min !== undefined && value < variable.min) || (variable.max !== undefined && value > variable.max))) failures.push(`PHYSICS_BOUND:${variable.id}`)
-      }
-      state = physics.state
-      steps.push(result)
-    }
-    return { steps, failures: [...new Set(failures)] }
-  }
-
-  evaluateConstraints(scene: Parameters<BaseProvider['evaluateConstraints']>[0], trajectory: Parameters<BaseProvider['evaluateConstraints']>[1]) {
-    // 场景约束 id ↔ spec 变量 id 翻译:骨架去重后 spec id 可能与场景变量 id 不同
-    // (如场景 state 与 target 指向同一节点),约束按 nodeId 映射回 spec 观测键;
-    // 结果 id 翻译回场景 id,保证 trial.constraintResults 与 SceneContract 对齐。
-    const sceneVars = [...scene.controls, ...scene.states, ...scene.disturbances, ...scene.observations, ...scene.guards]
-    const specIdOf = new Map(sceneVars.map((v) => {
-      const nodeId = v.nodeId ?? v.id
-      return [v.id, this.aliases.get(nodeId) ?? nodeId]
-    }))
-    const originalIdOf = new Map(scene.constraints.map((c) => {
-      const specId = specIdOf.get(c.id) ?? c.id
-      return [specId, c.id]
-    }))
-    const mappedScene = {
-      ...scene,
-      constraints: scene.constraints.map(c => ({ ...c, id: specIdOf.get(c.id) ?? c.id })),
-    }
-    // 可观测性豁免:训练数据裁剪可能把约束引用的量测节点剔出模型 IO 面(如 feature 角色的
-    // defect guard),模型轨迹里永远找不到该键 → 基类 fail-closed。这不是「违规」而是
-    // 「无证据」:按可观测性豁免并显式标注,治理上由 recommendationOnly(不直接写 DCW)兜底。
-    const observable = new Set<string>()
-    for (const step of trajectory.steps) {
-      for (const key of [...Object.keys(step.observations ?? {}), ...Object.keys(step.guards ?? {}), ...Object.keys(step.state ?? {})]) observable.add(key)
-    }
-    return this.base.evaluateConstraints(mappedScene as typeof scene, trajectory).map((r, i) => {
-      const specId = mappedScene.constraints[i]?.id ?? r.id
-      if (!r.passed && !observable.has(specId)) {
-        return {
-          ...r,
-          id: originalIdOf.get(r.id) ?? r.id,
-          passed: true,
-          detail: `UNOBSERVABLE:约束 ${r.id} 的量测节点不在模型 IO 面(训练数据裁剪),模型无法评估;按可观测性豁免(recommendationOnly 兜底;建议场景版本升级时同步裁剪该约束)`,
-        }
-      }
-      return { ...r, id: originalIdOf.get(r.id) ?? r.id }
-    })
-  }
-
-  takeUncertainty(): ModelBackedUncertainty {
-    return this.uncertainty
-  }
-}
+/**
+ * Model-backed rollout provider:让**训练注册后的 hybrid 模型**直接驱动 VirtualTrial/MPC。
+ *
+ * 组成 = hybrid_manifest.json 里冻结(含校准后 θ)的 PhysicsSpec 声明式物理主干
+ * + 残差集成 ONNX(默认 3 成员,平台 UQ 依据)。物理先走、残差只在 residualScale
+ * 边界内修正目标观测(归一化域,与 predictor 服役路径同一数学);历史窗口按
+ * persistence 语义维护(与 io_spec.assumptions 一致)。
+ *
+ * 这是"训练出的模型进入孪生闭环"的落实点:此前 trial/MPC 只拿 physics-only
+ * provider,model_id 只是标签;现在 model_id 背后的真实工件参与每一次 rollout。
+ */
+import { existsSync, readFileSync } from 'node:fs'
+import { join } from 'node:path'
+import { AppError } from '../../../../utils/errors'
+import { parseIoSpec } from '../model-registry'
+import { onnxSessionFor, ortRuntime, type IoSpec, type OrtSession } from '../predictor'
+import { compileDeclarativeProvider } from './declarative-provider'
+import { parsePhysicsSpec, type PhysicsSpec } from './physics-spec'
+import type { AmlModelRow } from '../aml.repo'
+import type { PhysicsState, PhysicsStepResult, PhysicsTrajectory } from './physics-runtime'
+import type { RolloutProvider } from './trial-service'
+
+interface HybridManifest {
+  physicsSpec?: unknown
+  residualScale?: number
+  composition?: string
+  physicsSpecHash?: string
+  ensemble?: {
+    members?: number
+    files?: string[]
+    calibrationCoverage?: number | null
+    coverageTarget?: number
+    calibrationRows?: number | null
+  }
+}
+
+export interface ModelBackedUncertainty {
+  /** 每步各目标归一化域的集成间最大标准差(与 residualScale 同量纲) */
+  maxStdByStep: number[]
+  /** 各成员最后一拍的目标预测(原始物理量纲;相对分歧对量纲不敏感) */
+  memberPredictions: Array<Record<string, number>>
+  calibrationCoverage: number | null
+  coverageTarget: number
+  members: number
+}
+
+export interface ModelSnapshotValues {
+  stateEstimate: Record<string, number>
+  controlValues: Record<string, number>
+  disturbances: Record<string, number>
+}
+
+/**
+ * 配方隔离守卫:模型的训练数据集 line/product/recipe 必须与场景谱系一致,
+ * 不同配方的模型禁止驱动彼此的孪生 rollout(单一实现,工具层与验收共用)。
+ */
+export function assertModelSceneLineage(
+  dataset: { lineId: string, productId: string, recipeId: string },
+  scene: { lineId?: string, productId?: string, recipeId?: string },
+): void {
+  const mismatch: string[] = []
+  if (scene.lineId && dataset.lineId !== scene.lineId) mismatch.push(`line ${dataset.lineId}≠${scene.lineId}`)
+  if (scene.productId && dataset.productId !== scene.productId) mismatch.push(`product ${dataset.productId}≠${scene.productId}`)
+  if (scene.recipeId && dataset.recipeId !== scene.recipeId) mismatch.push(`recipe ${dataset.recipeId}≠${scene.recipeId}`)
+  if (mismatch.length > 0) {
+    throw new AppError(409, 'TWIN_MODEL_SCENE_MISMATCH', `TWIN_MODEL_SCENE_MISMATCH:${mismatch.join(';')}(模型与场景/配方必须同谱系,禁止跨配方复用)`)
+  }
+}
+
+type BaseProvider = ReturnType<typeof compileDeclarativeProvider>
+
+export class ModelBackedHybridProvider implements RolloutProvider {
+  readonly base: BaseProvider
+  readonly spec: PhysicsSpec
+  readonly io: IoSpec
+  readonly hybridManifest: HybridManifest
+  readonly model: AmlModelRow
+  private readonly sessions: OrtSession[]
+  private readonly aliases: Map<string, string>
+  /** 训练期控制均值(按 varId);baseline 未提供全部控制输入时的回退先验,防 spec 求值 NaN */
+  private readonly defaultControls: Record<string, number>
+  private readonly residualScale: number
+  private rawHistory: number[][] = []
+  private readonly uncertainty: ModelBackedUncertainty
+
+  private constructor(
+    model: AmlModelRow,
+    manifest: HybridManifest,
+    sessions: OrtSession[],
+    snapshot: ModelSnapshotValues,
+  ) {
+    this.model = model
+    this.hybridManifest = manifest
+    if (!manifest.physicsSpec) throw new AppError(409, 'AML_HYBRID_PHYSICS_MISSING', `模型 ${model.id} 缺少冻结 PhysicsSpec,拒绝把 residual 当完整预测`)
+    this.spec = parsePhysicsSpec(manifest.physicsSpec)
+    this.io = parseIoSpec(model) as unknown as IoSpec
+    this.base = compileDeclarativeProvider(this.spec)
+    this.residualScale = Math.abs(Number(manifest.residualScale ?? 0.25))
+    this.aliases = new Map(this.spec.variables.filter(v => v.nodeId).map(v => [v.nodeId as string, v.id]))
+    this.defaultControls = Object.fromEntries(this.io.controlNodes.map((node, j) => {
+      const varId = this.aliases.get(node) ?? node
+      return [varId, this.io.norm.u.mean[j] ?? 0]
+    }))
+    this.sessions = sessions
+    this.uncertainty = {
+      maxStdByStep: [],
+      memberPredictions: [],
+      calibrationCoverage: manifest.ensemble?.calibrationCoverage ?? null,
+      coverageTarget: manifest.ensemble?.coverageTarget ?? 0.9,
+      members: manifest.ensemble?.members ?? sessions.length,
+    }
+    this.primeHistory(snapshot)
+  }
+
+  /** 工厂:加载 manifest + 残差集成会话(会话经 predictor LRU 复用) */
+  static async create(model: AmlModelRow, snapshot: ModelSnapshotValues): Promise<ModelBackedHybridProvider> {
+    const manifestPath = join(model.path, 'hybrid_manifest.json')
+    if (!existsSync(manifestPath)) {
+      throw new AppError(409, 'AML_MODEL_NOT_HYBRID', `模型 ${model.id} 不是 hybrid 工件(缺 hybrid_manifest.json),不能作为孪生 rollout 模型`)
+    }
+    let manifest: HybridManifest
+    try {
+      manifest = JSON.parse(readFileSync(manifestPath, 'utf8')) as HybridManifest
+    }
+    catch {
+      throw new AppError(409, 'AML_HYBRID_MANIFEST_INVALID', `模型 ${model.id} 的 hybrid_manifest.json 无法解析`)
+    }
+    if (!ortRuntime()) throw new AppError(503, 'AML_PREDICTOR_UNAVAILABLE', '预测服务依赖 onnxruntime-node 未安装:执行 pnpm install(可选依赖)后重启')
+    const files = manifest.ensemble?.files?.length ? manifest.ensemble.files : ['model.onnx']
+    const sessions: OrtSession[] = []
+    for (const [index, file] of files.entries()) {
+      const path = join(model.path, file)
+      if (!existsSync(path)) {
+        if (index === 0) throw new AppError(410, 'AML_ARTIFACT_MISSING', `模型 ${model.id} 残差工件缺失(${file})`)
+        continue
+      }
+      sessions.push(await onnxSessionFor(path, `${model.id}::${file}`))
+    }
+    return new ModelBackedHybridProvider(model, manifest, sessions, snapshot)
+  }
+
+  private valueForNode(node: string, snapshot: ModelSnapshotValues): number {
+    const varId = this.aliases.get(node) ?? node
+    const variable = this.spec.variables.find(v => v.id === varId)
+    return snapshot.stateEstimate[varId]
+      ?? snapshot.stateEstimate[node]
+      ?? snapshot.controlValues[varId]
+      ?? snapshot.controlValues[node]
+      ?? snapshot.disturbances[varId]
+      ?? snapshot.disturbances[node]
+      ?? variable?.defaultValue
+      ?? variable?.min
+      ?? 0
+  }
+
+  private primeHistory(snapshot: ModelSnapshotValues): void {
+    const row = this.io.allNodes.map(node => this.valueForNode(node, snapshot))
+    this.rawHistory = Array.from({ length: this.io.historySteps }, () => [...row])
+  }
+
+  private normalizeWindow(): number[][] {
+    const mx = this.io.norm.x.mean
+    const sx = this.io.norm.x.std
+    return this.rawHistory.map(row => row.map((v, j) => (v - (mx[j] ?? 0)) / (sx[j] ?? 1)))
+  }
+
+  private normalizeControls(controls: Record<string, number>, lastRow: number[]): number[] {
+    return this.io.controlNodes.map((node, j) => {
+      const varId = this.aliases.get(node) ?? node
+      const raw = controls[varId] ?? controls[node] ?? lastRow[this.io.allNodes.indexOf(node)] ?? 0
+      return (raw - (this.io.norm.u.mean[j] ?? 0)) / (this.io.norm.u.std[j] ?? 1)
+    })
+  }
+
+  /** 残差集成前向:归一化窗口(末拍控制列替换为本拍控制,与 predictor 同口径)→ 各成员 residual */
+  private async residualForward(historyNorm: number[][], controlsNorm: number[]): Promise<{ mean: number[], std: number[], members: number[][] }> {
+    const H = this.io.historySteps
+    const nAll = this.io.allNodes.length
+    // 末拍控制列替换为本拍控制(predictor 服役路径同语义;否则残差看到的是上一拍控制)
+    const window = historyNorm.map(row => [...row])
+    const last = window[H - 1]
+    if (last) {
+      this.io.controlNodes.forEach((node, c) => {
+        const j = this.io.allNodes.indexOf(node)
+        if (j >= 0) last[j] = controlsNorm[c] ?? last[j]!
+      })
+    }
+    const flat = new Float32Array(H * nAll)
+    for (let i = 0; i < H; i++) {
+      window[i]?.forEach((v, j) => {
+        flat[i * nAll + j] = v
+      })
+    }
+    const ort = ortRuntime()
+    if (!ort) throw new AppError(503, 'AML_PREDICTOR_UNAVAILABLE', 'onnxruntime-node 不可用')
+    const nTgt = this.io.targetNodes.length
+    const perMember: number[][] = []
+    for (const session of this.sessions) {
+      const inpName = session.inputNames[0]
+      if (!inpName) throw new AppError(503, 'AML_PREDICTOR_UNAVAILABLE', `残差 session 无输入名(模型 ${this.model.id})`)
+      const outName = session.outputNames[0] ?? inpName
+      const out = await session.run({ [inpName]: new ort.Tensor('float32', flat, [1, H, nAll]) })
+      const data = Array.from((out[outName] as { data: ArrayLike<number> }).data) as number[]
+      perMember.push(data.slice(0, nTgt))
+    }
+    const mean: number[] = []
+    const std: number[] = []
+    for (let t = 0; t < nTgt; t++) {
+      const series = perMember.map(row => row[t] ?? 0)
+      const mu = series.reduce((a, b) => a + b, 0) / Math.max(1, series.length)
+      const variance = series.length > 1 ? series.reduce((a, b) => a + (b - mu) ** 2, 0) / series.length : 0
+      mean.push(mu)
+      std.push(Math.sqrt(variance))
+    }
+    return { mean, std, members: perMember }
+  }
+
+  initialize(snapshot: { stateEstimate: Record<string, number>, controlValues: Record<string, number> }): PhysicsState {
+    return this.base.initialize(snapshot)
+  }
+
+  /** 模型驱动 rollout:物理主干 → 有界残差修正目标 → 历史窗口推进 */
+  async rollout(initial: PhysicsState, controls: Array<Record<string, number>>, disturbances: Record<string, number> = {}): Promise<PhysicsTrajectory> {
+    void disturbances
+    let state = { ...initial }
+    const steps: PhysicsStepResult[] = []
+    const failures: string[] = []
+    for (const control of controls) {
+      // 控制输入并集:baseline/candidate 常只含被优化变量,物理方程引用其余控制时回退训练均值(否则 ref miss → NaN)
+      const physics = this.base.step({ state, controls: { ...this.defaultControls, ...control }, disturbances, dtSec: this.spec.solver.dtSec })
+      const lastRow = this.rawHistory[this.rawHistory.length - 1] ?? []
+      // 目标物理值 → 归一化 → 加有界残差(与 predictor 服役路径同一数学)
+      const my = this.io.norm.y.mean
+      const sy = this.io.norm.y.std
+      const yPhysNorm = this.io.targetNodes.map((node, t) => {
+        const varId = this.aliases.get(node) ?? node
+        const raw: number | undefined = physics.observations[varId] ?? physics.state[varId]
+        if (raw === undefined || !Number.isFinite(raw)) {
+          throw new AppError(409, 'AML_HYBRID_PHYSICS_OUTPUT_MISSING', `PhysicsSpec 未生成目标 ${node}(模型 ${this.model.id};常见原因:spec 方程求值得到非有限值 —— 检查 spec 引用的控制/状态变量是否都有输入)`)
+        }
+        return (raw - (my[t] ?? 0)) / (sy[t] ?? 1)
+      })
+      const controlsNorm = this.normalizeControls(control, lastRow)
+      const { mean, std, members } = await this.residualForward(this.normalizeWindow(), controlsNorm)
+      const boundedOf = (value: number | undefined): number => Math.max(-this.residualScale, Math.min(this.residualScale, value ?? 0))
+      const correctedRaw = this.io.targetNodes.map((node, t) => (yPhysNorm[t]! + boundedOf(mean[t])) * (sy[t] ?? 1) + (my[t] ?? 0))
+      const observations = { ...physics.observations }
+      this.io.targetNodes.forEach((node, t) => {
+        const varId = this.aliases.get(node) ?? node
+        observations[varId] = correctedRaw[t]!
+      })
+      // 历史窗口推进(persistence:feature 持最后观测;control 取本拍;target 回填修正值;state 回填演化)
+      const nextRow = [...lastRow]
+      this.io.controlNodes.forEach((node) => {
+        const pos = this.io.allNodes.indexOf(node)
+        const varId = this.aliases.get(node) ?? node
+        if (pos >= 0) nextRow[pos] = control[varId] ?? control[node] ?? nextRow[pos] ?? 0
+      })
+      this.io.targetNodes.forEach((node, t) => {
+        const pos = this.io.allNodes.indexOf(node)
+        if (pos >= 0) nextRow[pos] = correctedRaw[t]!
+      })
+      for (const variable of this.spec.variables) {
+        if (variable.role !== 'state') continue
+        const pos = this.io.allNodes.indexOf(variable.nodeId ?? variable.id)
+        if (pos >= 0) nextRow[pos] = physics.state[variable.id] ?? nextRow[pos] ?? 0
+      }
+      this.rawHistory.push(nextRow)
+      if (this.rawHistory.length > this.io.historySteps) this.rawHistory.shift()
+      this.uncertainty.maxStdByStep.push(Math.max(...std, 0))
+      // 每成员"本拍目标会是什么"(原始量纲,各自残差而非均值)——相对分歧评估的输入
+      this.uncertainty.memberPredictions = members.map((row) => {
+        const entry: Record<string, number> = {}
+        this.io.targetNodes.forEach((node, t) => {
+          entry[node] = (yPhysNorm[t]! + boundedOf(row[t])) * (sy[t] ?? 1) + (my[t] ?? 0)
+        })
+        return entry
+      })
+      const result: PhysicsStepResult = { state: physics.state, observations, guards: physics.guards }
+      if ([...Object.values(result.state), ...Object.values(result.observations), ...Object.values(result.guards)].some(value => !Number.isFinite(value))) failures.push('HYBRID_NON_FINITE')
+      for (const variable of this.spec.variables) {
+        const value = result.state[variable.id] ?? result.observations[variable.id] ?? result.guards[variable.id]
+        if (value !== undefined && ((variable.min !== undefined && value < variable.min) || (variable.max !== undefined && value > variable.max))) failures.push(`PHYSICS_BOUND:${variable.id}`)
+      }
+      state = physics.state
+      steps.push(result)
+    }
+    return { steps, failures: [...new Set(failures)] }
+  }
+
+  evaluateConstraints(scene: Parameters<BaseProvider['evaluateConstraints']>[0], trajectory: Parameters<BaseProvider['evaluateConstraints']>[1]) {
+    // 场景约束 id ↔ spec 变量 id 翻译:骨架去重后 spec id 可能与场景变量 id 不同
+    // (如场景 state 与 target 指向同一节点),约束按 nodeId 映射回 spec 观测键;
+    // 结果 id 翻译回场景 id,保证 trial.constraintResults 与 SceneContract 对齐。
+    const sceneVars = [...scene.controls, ...scene.states, ...scene.disturbances, ...scene.observations, ...scene.guards]
+    const specIdOf = new Map(sceneVars.map((v) => {
+      const nodeId = v.nodeId ?? v.id
+      return [v.id, this.aliases.get(nodeId) ?? nodeId]
+    }))
+    const originalIdOf = new Map(scene.constraints.map((c) => {
+      const specId = specIdOf.get(c.id) ?? c.id
+      return [specId, c.id]
+    }))
+    const mappedScene = {
+      ...scene,
+      constraints: scene.constraints.map(c => ({ ...c, id: specIdOf.get(c.id) ?? c.id })),
+    }
+    // 可观测性豁免:训练数据裁剪可能把约束引用的量测节点剔出模型 IO 面(如 feature 角色的
+    // defect guard),模型轨迹里永远找不到该键 → 基类 fail-closed。这不是「违规」而是
+    // 「无证据」:按可观测性豁免并显式标注,治理上由 recommendationOnly(不直接写 DCW)兜底。
+    const observable = new Set<string>()
+    for (const step of trajectory.steps) {
+      for (const key of [...Object.keys(step.observations ?? {}), ...Object.keys(step.guards ?? {}), ...Object.keys(step.state ?? {})]) observable.add(key)
+    }
+    return this.base.evaluateConstraints(mappedScene as typeof scene, trajectory).map((r, i) => {
+      const specId = mappedScene.constraints[i]?.id ?? r.id
+      if (!r.passed && !observable.has(specId)) {
+        return {
+          ...r,
+          id: originalIdOf.get(r.id) ?? r.id,
+          passed: true,
+          detail: `UNOBSERVABLE:约束 ${r.id} 的量测节点不在模型 IO 面(训练数据裁剪),模型无法评估;按可观测性豁免(recommendationOnly 兜底;建议场景版本升级时同步裁剪该约束)`,
+        }
+      }
+      return { ...r, id: originalIdOf.get(r.id) ?? r.id }
+    })
+  }
+
+  takeUncertainty(): ModelBackedUncertainty {
+    return this.uncertainty
+  }
+}