agentworkshop 0.7.46 → 0.7.48

package/server/services/workshop/agents/industrial/twin-tools.ts

@@ -10,7 +10,9 @@ import { defaultInjectionScene, InjectionGreyboxProvider, smallStepCandidates }
 import { createTwinSnapshot } from '../../aml/twin/snapshot-service'
 import { evaluateHybridGates, DEFAULT_ACCEPTANCE_PROFILE } from '../../aml/twin/acceptance'
 import { issueRecommendationCertificate, runVirtualTrial } from '../../aml/twin/trial-service'
-import { sha256 } from '../../aml/twin/contracts'
+import { parseSceneContract, sha256, type SceneContract } from '../../aml/twin/contracts'
+import { getDaqController } from '../../daq/daq-controller'
+import { getAgentNodeBindingRepo } from '../node-bindings.repo'
 import type { HostToolResult } from '../host-tool-bridge/types'
 import { recordOps } from '../../ops/ops'
 import { createTwinRepo } from '../../aml/twin/repo'
@@ -46,17 +48,46 @@ async function persistTwinArtifact(kind: string, id: string, payload: unknown):
 }
 
 export async function toolTwinSceneRead(_agentId: string, args: Record<string, unknown>): Promise<HostToolResult> {
-  const scene = defaultInjectionScene('twin-scene-tool')
+  const defaultScene = defaultInjectionScene('twin-scene-tool')
+  const supplied = jsonArg<Record<string, unknown>>(args, 'scene_json')
+  const scene = supplied ? parseSceneContract(supplied) : defaultScene
   const sceneId = String(args.scene_id ?? scene.sceneId)
-  if (sceneId !== scene.sceneId) return fail(`当前 MVP 仅内置可执行场景 ${scene.sceneId}；未知 scene_id=${sceneId} 请先注册 PhysicsModelProvider。`)
-  return ok(`场景契约 ${scene.sceneId}@${scene.sceneVersion}\n${JSON.stringify(scene, null, 2)}\n\n控制策略：recommendation-only；模型门禁未通过时只能 safe_small_step，禁止直接 DCW。`)
+  if (sceneId !== scene.sceneId) return fail(`场景契约未提供: scene_id=${sceneId}；请通过 scene_json 注册该场景，或先注册 PhysicsModelProvider。`)
+  return ok(`场景契约 ${scene.sceneId}@${scene.sceneVersion}\n${JSON.stringify(scene, null, 2)}\n\n控制策略：recommendation-only；模型门禁未通过时只能 safe_small_step，禁止直接 DCW。\n快照前置：scene 的 observations/states 必须带 nodeId，且与 Agent 已绑定的真实 DAQ 节点一致；内置默认场景不带 nodeId，直接调用只会在 TwinSnapshot 里得到 fresh=false 并在 VirtualTrial 抛 SNAPSHOT_STALE —— 请用 scene_json 注入带 nodeId 的场景契约。`)
+}
+
+async function autoDaqSamples(agentId: string, scene: SceneContract, nowMs: number, freshnessMaxMs: number): Promise<Array<{ nodeId: string, at: number, value: number, sequence?: string }>> {
+  const required = [...scene.observations, ...scene.states]
+    .map(item => item.nodeId)
+    .filter((nodeId): nodeId is string => Boolean(nodeId))
+  const daqBindings = new Set(getAgentNodeBindingRepo().byAgent(agentId).filter(binding => binding.kind === 'daq').map(binding => binding.nodeId))
+  const missingBindings = required.filter(nodeId => !daqBindings.has(nodeId))
+  if (missingBindings.length > 0) throw new Error(`AUTO_DAQ_UNBOUND:${missingBindings.join(',')}`)
+  const samples: Array<{ nodeId: string, at: number, value: number, sequence?: string }> = []
+  for (const nodeId of required) {
+    const points = await getDaqController().samples(nodeId, { fromMs: nowMs - freshnessMaxMs, toMs: nowMs, bucketMs: 1000, limit: 64 }) as Array<{ at?: number, value?: number, avg?: number }>
+    const latest = points
+      .map(point => ({ at: Number(point.at), value: Number(point.value ?? point.avg) }))
+      .filter(point => Number.isFinite(point.at) && Number.isFinite(point.value))
+      .sort((a, b) => a.at - b.at)
+      .at(-1)
+    if (!latest) throw new Error(`AUTO_DAQ_SAMPLE_MISSING:${nodeId}`)
+    samples.push({ nodeId, at: latest.at, value: latest.value, sequence: `daq:${nodeId}:${latest.at}` })
+  }
+  return samples
 }
 
 export async function toolTwinSnapshotCreate(agentId: string, args: Record<string, unknown>): Promise<HostToolResult> {
   try {
-    const scene = (jsonArg(args, 'scene_json') ?? defaultInjectionScene('twin-snapshot-tool')) as ReturnType<typeof defaultInjectionScene>
+    const scene = parseSceneContract(jsonArg(args, 'scene_json') ?? defaultInjectionScene('twin-snapshot-tool'))
     const twinRepo = createTwinRepo(getAmlRuntime().db)
     twinRepo.upsertScene(scene, agentId)
+    const nowMs = Number(args.now_ms) || Date.now()
+    const freshnessMaxMs = Number(args.freshness_max_ms) || 60_000
+    const suppliedSamples = jsonArg<Array<{ nodeId: string, at: number, value: number, sequence?: string }>>(args, 'samples', []) ?? []
+    const samples = args.auto_daq === true || suppliedSamples.length === 0
+      ? await autoDaqSamples(agentId, scene, nowMs, freshnessMaxMs)
+      : suppliedSamples
     const snapshot = createTwinSnapshot({
       scene,
       channelId: String(args.channel_id ?? ''),
@@ -65,14 +96,24 @@ export async function toolTwinSnapshotCreate(agentId: string, args: Record<strin
       controls: jsonArg<Record<string, number>>(args, 'controls', {}) ?? {},
       states: jsonArg<Record<string, number>>(args, 'states', {}) ?? {},
       disturbances: jsonArg<Record<string, number>>(args, 'disturbances', {}) ?? {},
-      samples: jsonArg<Array<{ nodeId: string, at: number, value: number, sequence?: string }>>(args, 'samples', []) ?? [],
-      nowMs: Number(args.now_ms) || Date.now(),
-      freshnessMaxMs: Number(args.freshness_max_ms) || 60_000,
+      samples,
+      nowMs,
+      freshnessMaxMs,
     })
     twinRepo.insertSnapshot(snapshot)
     const path = await persistTwinArtifact('snapshots', snapshot.snapshotId, snapshot)
     recordOps({ actor: agentId, actorName: agentBadgeLabel(agentId), actorKind: 'agent', action: 'aml.twin.snapshot_create', kind: 'aml' as 'system', summary: `创建 TwinSnapshot ${snapshot.snapshotId}`, targetKind: 'aml_twin_snapshot', targetId: snapshot.snapshotId, lineId: scene.lineId, productId: scene.productId, recipeId: scene.recipeId })
-    return ok(`TwinSnapshot 已创建\n  snapshot_id: ${snapshot.snapshotId}\n  hash: ${snapshot.snapshotHash}\n  fresh: ${snapshot.dataQuality.fresh}\n  completeness: ${snapshot.dataQuality.completeness}\n  artifact: ${path}`)
+    const quality = snapshot.dataQuality
+    // 全部必需节点都没有样本(watermark=0)时不是"数据不新鲜",而是场景映射/绑定配置错了:
+    // 旧行为只回 fresh=false,调用方要到 VirtualTrial 才看到 SNAPSHOT_STALE(实测踩过)。
+    const requiredCount = [...scene.observations, ...scene.states].filter(v => v.nodeId).length
+    if (snapshot.daqWatermark <= 0) {
+      const missing = requiredCount === 0
+        ? '场景 observations/states 没有任何 nodeId(内置默认场景即如此)'
+        : `以下节点没有可用样本: ${quality.staleNodeIds.join(', ')}`
+      return fail(`TwinSnapshot 创建失败:无可用 DAQ 样本(watermark=0)。${missing}。请用 scene_json 把场景观测/状态映射到已绑定的真实 DAQ 节点,并确认产线在跑。\n  snapshot_id: ${snapshot.snapshotId}\n  artifact: ${path}`)
+    }
+    return ok(`TwinSnapshot 已创建\n  source: ${args.auto_daq === true || suppliedSamples.length === 0 ? 'bound DAQ' : 'caller samples (legacy/test)'}\n  snapshot_id: ${snapshot.snapshotId}\n  hash: ${snapshot.snapshotHash}\n  fresh: ${quality.fresh}\n  completeness: ${quality.completeness}\n  stale_nodes: ${quality.staleNodeIds.length ? quality.staleNodeIds.join(', ') : '(none)'}\n  artifact: ${path}`)
   }
   catch (err) {
     return fail(`TwinSnapshot 创建失败:${err instanceof Error ? err.message : String(err)}`)
@@ -84,6 +125,13 @@ export async function toolTwinTrialRun(agentId: string, args: Record<string, unk
     const scene = (jsonArg(args, 'scene_json') ?? defaultInjectionScene('twin-trial-tool')) as ReturnType<typeof defaultInjectionScene>
     const snapshot = jsonArg<Record<string, unknown>>(args, 'snapshot_json') as never
     if (!snapshot) return fail('snapshot_json 必填；必须使用当前 DAQ 生成的 TwinSnapshot，禁止 Agent 伪造执行状态。')
+    // 形状校验:传半截 JSON / 工具结果信封 / 纯字符串时旧行为是原生 TypeError
+    // (Cannot read properties of undefined (reading 'fresh')),调用方看不出该怎么修。
+    const snap = snapshot as { dataQuality?: { fresh?: boolean }, snapshotHash?: string, snapshotId?: string }
+    if (!snap.dataQuality || typeof snap.dataQuality.fresh !== 'boolean' || !snap.snapshotId) {
+      return fail('snapshot_json 不是 TwinSnapshot 工件:缺少 snapshotId/dataQuality。请先调用 twin_snapshot_create,并把其 artifact 文件的 JSON 原文整段回传(不要传工具结果文本或截断片段)。')
+    }
+    if (!snap.snapshotHash) return fail('snapshot_json 缺少 snapshotHash(快照被篡改或截断),拒绝执行 VirtualTrial。')
     const provider = new InjectionGreyboxProvider()
     const baseline = jsonArg<Record<string, number>>(args, 'baseline_controls', {}) ?? {}
     const candidate = jsonArg<Array<Record<string, number>>>(args, 'candidate_controls', []) ?? []
@@ -147,7 +195,34 @@ export async function toolTwinGateEvaluate(_agentId: string, args: Record<string
     const result = evaluateHybridGates({
       rows: Number(args.rows ?? 0), runs: Number(args.runs ?? 0), oneStepTestNrmse: Number(args.one_step_nrmse ?? 1), rolloutTestNrmse: Number(args.rollout_nrmse ?? 1), valTestGap: Number(args.val_test_gap ?? 1), calibrationRows: Number(args.calibration_rows ?? 0), calibrationCoverage: Number(args.coverage ?? 0), candidateTrials: jsonArg(args, 'candidate_trials', []) as never, physicsSolverFailureRate: Number(args.physics_failure_rate ?? 1),
     }, DEFAULT_ACCEPTANCE_PROFILE)
-    return ok(JSON.stringify(result, null, 2))
+    const modelId = String(args.model_id ?? '').trim()
+    let modelUpdate = ''
+    if (modelId) {
+      const rt = getAmlRuntime()
+      const model = rt.repo.model.get(modelId)
+      if (model) {
+        let metrics: Record<string, unknown> = {}
+        try {
+          metrics = JSON.parse(model.metricsJson || '{}') as Record<string, unknown>
+        }
+        catch {
+          metrics = {}
+        }
+        const twinEligibility = {
+          gatePassed: result.passed,
+          recommendationEligible: result.passed,
+          uqPassed: result.checks.filter(c => c.id === 'G7_UQ_OOD').every(c => c.passed),
+          oodPassed: result.checks.filter(c => c.id === 'G7_UQ_OOD').every(c => c.passed),
+          physicsPassed: result.checks.filter(c => c.id === 'G5_PHYSICS').every(c => c.passed),
+          evaluatedAt: new Date().toISOString(),
+        }
+        metrics.twinEligibility = twinEligibility
+        rt.db.prepare('UPDATE aml_models SET metrics_json = ? WHERE id = ?').run(JSON.stringify(metrics), modelId)
+        modelUpdate = `\nmodel_id: ${modelId}\nmodel_twin_eligibility: ${JSON.stringify(twinEligibility)}`
+      }
+      else modelUpdate = `\nmodel_id: ${modelId}\nmodel_twin_eligibility: MODEL_NOT_FOUND`
+    }
+    return ok(JSON.stringify({ ...result, modelUpdate }, null, 2))
   }
   catch (err) {
     return fail(`Twin Gate 评估失败:${err instanceof Error ? err.message : String(err)}`)

package/server/services/workshop/aml/job-orchestrator/start-job.ts

@@ -1,124 +1,140 @@
-/**
- * 启动单个作业(环境探测 / 参数组装 / 分支到 stub 或真实进程)
- * (由 server/services/workshop/aml/job-orchestrator.ts 按职责拆出;内容逐行原文搬运)
- */
-import type { AmlJobRow } from '../aml.repo'
-import type { RunningJob } from './shared'
-import { concludeJob, runStub } from './stages'
-import { ensureVenv, platformPythonDir } from '../python-runtime'
-import { existsSync, mkdirSync, writeFileSync } from 'node:fs'
-import { experimentFail, failPermanentOrRetry, finishRun, lastErrLine, safeParse } from './failure'
-import { getAmlRuntime } from '../runtime'
-import { hashDatasetDir } from '../dataset-builder'
-import { join } from 'node:path'
-import { runPython } from './process'
-import { state } from './shared'
-import { wsStage } from './broadcast'
-
-export async function startJob(row: AmlJobRow): Promise<void> {
-  const rt = getAmlRuntime()
-  const st = state()
-  const dataset = rt.repo.dataset.get(row.datasetId)
-  if (!dataset) {
-    rt.repo.job.fail(row.id, 'failed', '数据集已被删除(永久错误)', new Date().toISOString())
-    return
-  }
-  // 快照完整性校验(防篡改;sha 失配 = 永久错误不重试)
-  const actualSha = hashDatasetDir(dataset.path, dataset.specJson)
-  if (actualSha !== dataset.sha256) {
-    rt.repo.job.fail(row.id, 'failed', `数据集快照校验失败(sha256 不匹配,疑似被改动或损坏):${dataset.id}(永久错误)`, new Date().toISOString())
-    return
-  }
-  const jobDir = join(rt.jobsDir, row.id)
-  const workspace = join(jobDir, 'workspace')
-  const artifacts = join(jobDir, 'artifacts')
-  mkdirSync(workspace, { recursive: true })
-  mkdirSync(artifacts, { recursive: true })
-  let trainFile = existsSync(join(workspace, 'train.py')) ? 'train.py' : null
-  if (!trainFile) {
-    // REST 内联代码路径:预算袋携带 inlineCode 时落盘(≤500KB,提交入口已限)
-    const inline = safeParse(row.budget).inlineCode
-    if (typeof inline === 'string' && inline.length > 0 && inline.length <= 512_000) {
-      writeFileSync(join(workspace, 'train.py'), inline)
-      trainFile = 'train.py'
-    }
-    else {
-      rt.repo.job.fail(row.id, 'failed', 'workspace/train.py 不存在(Agent 用 aml_job_submit 携带 code 提交,或 REST 提交时带 code)', new Date().toISOString())
-      return
-    }
-  }
-
-  writeFileSync(join(jobDir, 'job.json'), JSON.stringify({
-    jobId: row.id,
-    datasetPath: dataset.path,
-    workspaceDir: workspace,
-    params: safeParse(row.budget).params ?? {},
-    seed: safeParse(row.budget).seed ?? 42,
-    trainFile,
-  }, null, 2))
-  const run: RunningJob = {
-    row,
-    jobDir,
-    child: null,
-    startedAt: Date.now(),
-    lastProtocolAt: Date.now(),
-    lastWsAt: 0,
-    logLines: [],
-    protocolBadLines: 0,
-    killed: false,
-  }
-  st.running.set(row.id, run)
-  rt.repo.job.markStarted(row.id, 'provisioning', 'provision', new Date().toISOString())
-  wsStage(row, 'provision')
-
-  try {
-    if (process.env.AML_STUB === '1') {
-      await runStub(run, artifacts)
-      return
-    }
-    const vpy = await ensureVenv(rt)
-    if (run.killed) return
-    rt.repo.job.setState(row.id, 'training', 'train', 5)
-    wsStage(row, 'train')
-    // 第一拍:Agent 训练代码(独立进程;崩溃不伤评估器)
-    const exit1 = await runPython(run, vpy, [join(workspace, trainFile)], workspace)
-    if (run.killed) return
-    if (exit1.code !== 0) {
-      failPermanentOrRetry(run, `训练进程退出(${exit1.code}):${lastErrLine(exit1.errTail)}`, exit1.code)
-      return
-    }
-    if (!existsSync(join(artifacts, 'model.onnx'))) {
-      rt.repo.job.fail(row.id, 'failed', '训练未产出 artifacts/model.onnx(契约缺失,永久错误不重试)', new Date().toISOString())
-      experimentFail(run, '契约缺失:model.onnx 未产出')
-      finishRun(run)
-      return
-    }
-    // 第二拍:平台权威评估
-    rt.repo.job.setState(row.id, 'evaluating', 'evaluate', 80)
-    wsStage(row, 'evaluate')
-    const evalScript = join(platformPythonDir(), 'aml_eval.py')
-    const exit2 = await runPython(run, vpy, [evalScript, '--job', join(jobDir, 'job.json')], jobDir)
-    if (run.killed) return
-    if (exit2.code !== 0) {
-      rt.repo.job.fail(row.id, 'failed', `评估失败:${lastErrLine(exit2.errTail)}(永久错误不重试)`, new Date().toISOString())
-      experimentFail(run, `评估失败:${lastErrLine(exit2.errTail)}`)
-      finishRun(run)
-      return
-    }
-    concludeJob(run, artifacts)
-  }
-  catch (err) {
-    if (!run.killed) {
-      const msg = err instanceof Error ? err.message : String(err)
-      const permanent = String((err as { code?: string })?.code ?? '').startsWith('AML_')
-      if (permanent) {
-        rt.repo.job.fail(row.id, 'failed', msg, new Date().toISOString())
-        experimentFail(run, msg)
-        finishRun(run)
-      }
-      else {
-        failPermanentOrRetry(run, msg, 1)
-      }
-    }
-  }
-}
+/**
+ * 启动单个作业(环境探测 / 参数组装 / 分支到 stub 或真实进程)
+ * (由 server/services/workshop/aml/job-orchestrator.ts 按职责拆出;内容逐行原文搬运)
+ */
+import type { AmlJobRow } from '../aml.repo'
+import type { RunningJob } from './shared'
+import { concludeJob, runStub } from './stages'
+import { ensureVenv, platformPythonDir } from '../python-runtime'
+import { existsSync, mkdirSync, writeFileSync } from 'node:fs'
+import { experimentFail, failPermanentOrRetry, finishRun, lastErrLine, safeParse } from './failure'
+import { getAmlRuntime } from '../runtime'
+import { hashDatasetDir } from '../dataset-builder'
+import { join } from 'node:path'
+import { runPython } from './process'
+import { state } from './shared'
+import { wsStage } from './broadcast'
+
+export async function startJob(row: AmlJobRow): Promise<void> {
+  const rt = getAmlRuntime()
+  const st = state()
+  const dataset = rt.repo.dataset.get(row.datasetId)
+  if (!dataset) {
+    rt.repo.job.fail(row.id, 'failed', '数据集已被删除(永久错误)', new Date().toISOString())
+    return
+  }
+  // 快照完整性校验(防篡改;sha 失配 = 永久错误不重试)
+  const actualSha = hashDatasetDir(dataset.path, dataset.specJson)
+  if (actualSha !== dataset.sha256) {
+    rt.repo.job.fail(row.id, 'failed', `数据集快照校验失败(sha256 不匹配,疑似被改动或损坏):${dataset.id}(永久错误)`, new Date().toISOString())
+    return
+  }
+  const jobDir = join(rt.jobsDir, row.id)
+  const workspace = join(jobDir, 'workspace')
+  const artifacts = join(jobDir, 'artifacts')
+  mkdirSync(workspace, { recursive: true })
+  mkdirSync(artifacts, { recursive: true })
+  let trainFile = existsSync(join(workspace, 'train.py')) ? 'train.py' : null
+  if (!trainFile) {
+    // REST 内联代码路径:预算袋携带 inlineCode 时落盘(≤500KB,提交入口已限)
+    const inline = safeParse(row.budget).inlineCode
+    if (typeof inline === 'string' && inline.length > 0 && inline.length <= 512_000) {
+      writeFileSync(join(workspace, 'train.py'), inline)
+      trainFile = 'train.py'
+    }
+    else {
+      rt.repo.job.fail(row.id, 'failed', 'workspace/train.py 不存在(Agent 用 aml_job_submit 携带 code 提交,或 REST 提交时带 code)', new Date().toISOString())
+      return
+    }
+  }
+
+  const budget = safeParse(row.budget) as Record<string, unknown>
+  writeFileSync(join(jobDir, 'job.json'), JSON.stringify({
+    jobId: row.id,
+    datasetPath: dataset.path,
+    workspaceDir: workspace,
+    params: budget.params ?? {},
+    seed: budget.seed ?? 42,
+    trainFile,
+    sceneId: budget.sceneId ?? null,
+    sceneVersion: budget.sceneVersion ?? null,
+    objectiveId: budget.objectiveId ?? null,
+    jobKind: budget.jobKind ?? 'supervised',
+  }, null, 2))
+  // Hybrid Twin lineage is immutable job input. Persist these sidecars in the
+  // isolated workspace so amlkit.load_* can consume them without trusting REST
+  // callers or reaching into the live database during training.
+  for (const [name, key] of [
+    ['physics_manifest.json', 'physicsManifest'],
+    ['twin_snapshot.json', 'twinSnapshot'],
+    ['objective_profile.json', 'objectiveProfile'],
+  ] as const) {
+    const value = budget[key]
+    if (value && typeof value === 'object') writeFileSync(join(workspace, name), JSON.stringify(value, null, 2))
+  }
+  const run: RunningJob = {
+    row,
+    jobDir,
+    child: null,
+    startedAt: Date.now(),
+    lastProtocolAt: Date.now(),
+    lastWsAt: 0,
+    logLines: [],
+    protocolBadLines: 0,
+    killed: false,
+  }
+  st.running.set(row.id, run)
+  rt.repo.job.markStarted(row.id, 'provisioning', 'provision', new Date().toISOString())
+  wsStage(row, 'provision')
+
+  try {
+    if (process.env.AML_STUB === '1') {
+      await runStub(run, artifacts)
+      return
+    }
+    const vpy = await ensureVenv(rt)
+    if (run.killed) return
+    rt.repo.job.setState(row.id, 'training', 'train', 5)
+    wsStage(row, 'train')
+    // 第一拍:Agent 训练代码(独立进程;崩溃不伤评估器)
+    const exit1 = await runPython(run, vpy, [join(workspace, trainFile)], workspace)
+    if (run.killed) return
+    if (exit1.code !== 0) {
+      failPermanentOrRetry(run, `训练进程退出(${exit1.code}):${lastErrLine(exit1.errTail)}`, exit1.code)
+      return
+    }
+    if (!existsSync(join(artifacts, 'model.onnx'))) {
+      rt.repo.job.fail(row.id, 'failed', '训练未产出 artifacts/model.onnx(契约缺失,永久错误不重试)', new Date().toISOString())
+      experimentFail(run, '契约缺失:model.onnx 未产出')
+      finishRun(run)
+      return
+    }
+    // 第二拍:平台权威评估
+    rt.repo.job.setState(row.id, 'evaluating', 'evaluate', 80)
+    wsStage(row, 'evaluate')
+    const evalScript = join(platformPythonDir(), 'aml_eval.py')
+    const exit2 = await runPython(run, vpy, [evalScript, '--job', join(jobDir, 'job.json')], jobDir)
+    if (run.killed) return
+    if (exit2.code !== 0) {
+      rt.repo.job.fail(row.id, 'failed', `评估失败:${lastErrLine(exit2.errTail)}(永久错误不重试)`, new Date().toISOString())
+      experimentFail(run, `评估失败:${lastErrLine(exit2.errTail)}`)
+      finishRun(run)
+      return
+    }
+    concludeJob(run, artifacts)
+  }
+  catch (err) {
+    if (!run.killed) {
+      const msg = err instanceof Error ? err.message : String(err)
+      const permanent = String((err as { code?: string })?.code ?? '').startsWith('AML_')
+      if (permanent) {
+        rt.repo.job.fail(row.id, 'failed', msg, new Date().toISOString())
+        experimentFail(run, msg)
+        finishRun(run)
+      }
+      else {
+        failPermanentOrRetry(run, msg, 1)
+      }
+    }
+  }
+}

package/server/services/workshop/aml/twin/physics-runtime.ts

@@ -94,14 +94,29 @@ export class InjectionGreyboxProvider implements PhysicsModelProvider {
   }
 
   evaluateConstraints(scene: SceneContract, trajectory: PhysicsTrajectory): Array<{ id: string, passed: boolean, detail: string, firstViolationStep?: number }> {
+    // 失败关闭(fail-closed):约束 id 必须能映射到轨迹观测量/守卫量。
+    // 旧行为对映射不到的 id 直接跳过并返回「全轨迹通过」(value == null 分支),
+    // 于是 scene_json 里写成 part_weight / 质量窗口 这类 id 时,硬约束门禁形同虚设
+    // —— 实测被真实 Agent 团队两次抓出:预测 weight 30.3 g 明显低于窗口下界仍报通过。
+    const known = [...new Set([
+      ...Object.keys(trajectory.steps.at(-1)?.observations ?? {}),
+      ...Object.keys(trajectory.steps.at(-1)?.guards ?? {}),
+    ])]
+    if (trajectory.steps.length === 0) {
+      return scene.constraints.map(constraint => ({ id: constraint.id, passed: false, detail: '空轨迹:未执行任何校验' }))
+    }
     return scene.constraints.map((constraint) => {
+      let unmapped = true
       for (let i = 0; i < trajectory.steps.length; i++) {
         const obs = trajectory.steps[i]?.observations ?? {}
         const value = obs[constraint.id] ?? trajectory.steps[i]?.guards[constraint.id]
-        if (constraint.kind === 'hard_range' && value != null && ((constraint.min != null && value < constraint.min) || (constraint.max != null && value > constraint.max))) {
+        if (value == null) continue
+        unmapped = false
+        if (constraint.kind === 'hard_range' && ((constraint.min != null && value < constraint.min) || (constraint.max != null && value > constraint.max))) {
           return { id: constraint.id, passed: false, detail: `${constraint.id}=${value} 越过 ${constraint.min ?? '-∞'}~${constraint.max ?? '∞'}`, firstViolationStep: i }
         }
       }
+      if (unmapped) return { id: constraint.id, passed: false, detail: `约束 ${constraint.id} 无法映射到观测量/守卫量(可用:${known.join(', ') || '无'}),按失败关闭处理` }
       return { id: constraint.id, passed: true, detail: '全轨迹通过' }
     })
   }

package/server/services/workshop/aml/twin/repo.ts

@@ -8,19 +8,24 @@ export function createTwinRepo(db: DatabaseSync) {
   const insertSnapshotStmt = db.prepare(`INSERT OR IGNORE INTO twin_snapshots(id,scene_id,scene_version,line_id,product_id,recipe_id,phase,snapshot_hash,watermark_ms,freshness_json,payload_json,created_by,created_at) VALUES(?,?,?,?,?,?,?,?,?,?,?,?,?)`)
   const insertTrialStmt = db.prepare(`INSERT OR IGNORE INTO twin_trials(id,snapshot_id,twin_model_id,objective_id,candidate_executed,status,result_json,result_hash,created_by,created_at) VALUES(?,?,?,?,0,?,?,?,?,?)`)
   const insertRecStmt = db.prepare(`INSERT OR IGNORE INTO twin_recommendations(id,trial_id,certificate_hash,status,certificate_json,expires_at,created_by,created_at) VALUES(?,?,?,?,?,?,?,?)`)
+  // 绑定值直接来自调用方(scene_json/snapshot_json 都是外部输入):对象/undefined 会以
+  // "Provided value cannot be bound to SQLite parameter N" 这种驱动级报错冒出来(实测踩过),
+  // 这里统一收敛成字符串/有限数字,让参数形状问题不再污染错误面。
+  const text = (v: unknown): string => (v == null ? '' : typeof v === 'string' ? v : String(v))
+  const num = (v: unknown): number => (Number.isFinite(Number(v)) ? Number(v) : 0)
   return {
     upsertScene(scene: SceneContract, createdBy: string): void {
       const now = new Date().toISOString()
-      upsertSceneStmt.run(scene.sceneId, scene.sceneVersion, scene.schemaVersion, JSON.stringify(scene), sha256(scene), scene.lineId, scene.productId ?? '', scene.recipeId ?? '', 'draft', createdBy, scene.createdAt || now, now)
+      upsertSceneStmt.run(text(scene.sceneId), text(scene.sceneVersion), num(scene.schemaVersion), JSON.stringify(scene), sha256(scene), text(scene.lineId), text(scene.productId), text(scene.recipeId), 'draft', text(createdBy), text(scene.createdAt) || now, now)
     },
     insertSnapshot(snapshot: TwinSnapshot): void {
-      insertSnapshotStmt.run(snapshot.snapshotId, snapshot.sceneId, snapshot.sceneVersion, snapshot.lineId, snapshot.productId ?? '', snapshot.recipeId ?? '', snapshot.phase, snapshot.snapshotHash, snapshot.daqWatermark, JSON.stringify(snapshot.dataQuality), JSON.stringify(snapshot), snapshot.createdBy, snapshot.createdAt)
+      insertSnapshotStmt.run(text(snapshot.snapshotId), text(snapshot.sceneId), text(snapshot.sceneVersion), text(snapshot.lineId), text(snapshot.productId), text(snapshot.recipeId), text(snapshot.phase), text(snapshot.snapshotHash), num(snapshot.daqWatermark), JSON.stringify(snapshot.dataQuality ?? {}), JSON.stringify(snapshot), text(snapshot.createdBy), text(snapshot.createdAt))
     },
     insertTrial(trial: VirtualTrial, twinModelId?: string): void {
-      insertTrialStmt.run(trial.trialId, trial.snapshotId, twinModelId ?? trial.modelId, trial.objectiveId, 'completed', JSON.stringify(trial), sha256(trial), trial.createdBy, trial.createdAt)
+      insertTrialStmt.run(text(trial.trialId), text(trial.snapshotId), text(twinModelId ?? trial.modelId), text(trial.objectiveId), 'completed', JSON.stringify(trial), sha256(trial), text(trial.createdBy), text(trial.createdAt))
     },
     insertRecommendation(certificate: RecommendationCertificate, createdBy: string, expiresAt?: string): void {
-      insertRecStmt.run(certificate.recommendationId, certificate.trialId, certificate.certificateHash, certificate.status.toLowerCase(), JSON.stringify(certificate), expiresAt ?? null, createdBy, certificate.createdAt)
+      insertRecStmt.run(text(certificate.recommendationId), text(certificate.trialId), text(certificate.certificateHash), text(certificate.status).toLowerCase(), JSON.stringify(certificate), expiresAt ?? null, text(createdBy), text(certificate.createdAt))
     },
   }
 }

package/.output/public/_nuxt/builds/meta/7827fe9f-b4c2-461f-b1fb-c0c7da7d66b8.json

@@ -1 +0,0 @@
-{"id":"7827fe9f-b4c2-461f-b1fb-c0c7da7d66b8","timestamp":1790316372736,"prerendered":[]}