agentworkshop 0.7.46 → 0.7.48

package/scripts/_aw0746-stage7-twin.mjs

@@ -0,0 +1,267 @@
+/**
+ * 0.7.46/0.7.47 打包系统验收 · 阶段 7:hybrid twin × MPC 闭环(真实数据 → 真实模型 → 推荐门禁)
+ *
+ * 关键事实(实测):内置默认场景 `defaultInjectionScene()` 的 observations/states **不带 nodeId**,
+ * 而 `createTwinSnapshot` 的 required 节点表只从 `scene.*[].nodeId` 取 ⇒ watermark 恒为 0 ⇒
+ * `fresh` 永远是 false,`runVirtualTrial` 必然抛 SNAPSHOT_STALE。因此真实数据链路必须由调用方
+ * 用 `scene_json` 把场景观测/状态映射到**已绑定的真实 DAQ 节点**(本脚本即按此注入)。
+ *
+ * 链路:绑定节点 → 真实样本 → TwinSnapshot(fresh=true) → 27 组真实 VirtualTrial
+ *       → 12 项推荐门禁 → 写入模型 twinEligibility → mpc_optimize 切 precise_search
+ *       → 全程零 DCW 写入(candidateExecuted=false)
+ * 用法:AW_BASE=http://127.0.0.1:3001 node scripts/_aw0746-stage7-twin.mjs [AW_REQUIRE_MODEL=1]
+ */
+import { existsSync, readFileSync } from 'node:fs'
+import { join } from 'node:path'
+import { Client } from 'pg'
+
+const BASE = process.env.AW_BASE ?? 'http://127.0.0.1:3001'
+const HOME = process.env.AW_HOME ?? `${process.cwd()}/.aw0746-home`
+const ADMIN = { email: process.env.AW_ADMIN_EMAIL ?? 'admin@awshop.local', password: process.env.AW_ADMIN_PASSWORD ?? 'Awshop@123!' }
+const REQUIRE_MODEL = process.env.AW_REQUIRE_MODEL === '1'
+
+let pass = 0
+const fails = []
+const skips = []
+const check = (name, ok, detail = '') => {
+  console.log(`  ${ok ? 'PASS' : 'FAIL'}  ${name}${detail ? ` — ${detail}` : ''}`)
+  if (ok) pass += 1
+  else fails.push(name)
+}
+const skip = (name, detail) => {
+  console.log(`  SKIP  ${name} — ${detail}`)
+  skips.push(name)
+}
+const j = async (method, path, body, token) => {
+  const r = await fetch(BASE + path, {
+    method,
+    headers: { 'content-type': 'application/json', ...(token ? { authorization: `Bearer ${token}` } : {}) },
+    body: body === undefined ? undefined : JSON.stringify(body),
+    signal: AbortSignal.timeout(180_000),
+  })
+  return { status: r.status, ...(await r.json().catch(() => ({}))) }
+}
+const firstJson = (text) => {
+  const a = text.indexOf('{')
+  const b = text.lastIndexOf('}')
+  return a >= 0 && b > a ? JSON.parse(text.slice(a, b + 1)) : null
+}
+const clamp = (v, lo, hi) => Math.min(hi, Math.max(lo, v))
+const sleep = ms => new Promise(r => setTimeout(r, ms))
+
+const token = (await j('POST', '/api/users/login', { email: ADMIN.email, password: ADMIN.password })).data?.token
+console.log(`\n═══ 阶段 7:hybrid twin × MPC 闭环(真实数据) @ ${BASE} ═══`)
+
+// ── 1. 解析通道 / 产线 / 节点 ──
+const chans = (await j('GET', '/api/workshop/channels', undefined, token)).data ?? []
+const twin = chans.find(c => /闭环优化-hybridtwin/.test(c.name))
+const members = (await j('GET', `/api/workshop/channels/${twin.id}/agents`, undefined, token)).data ?? []
+const dcw = (await j('GET', '/api/workshop/dcw', undefined, token)).data
+const line = (dcw.lines ?? []).find(l => /injection-aw0746/i.test(String(l.name)))
+const product = (dcw.products ?? []).find(p => p.lineId === line.id)
+const recipe = (dcw.recipes ?? []).find(r => r.lineId === line.id)
+const daq = (await j('GET', '/api/workshop/daq', undefined, token)).data
+const daqNodes = (daq.nodes ?? []).filter(n => n.lineId === line.id)
+const dcwNodes = (dcw.nodes ?? []).filter(n => n.lineId === line.id)
+const find = (arr, re) => arr.find(n => re.test(n.name))
+const NODE = {
+  weight: find(daqNodes, /part-weight/),
+  flash_rate: find(daqNodes, /flash-rate/),
+  sink_rate: find(daqNodes, /sink-mark/),
+  melt_temperature: find(daqNodes, /melt-temp-pv/),
+  cavity_pressure: find(daqNodes, /melt-pressure-pv/),
+  hold_pressure: find(dcwNodes, /hold-pressure/),
+  hold_time: find(dcwNodes, /hold-time/),
+  melt_temperature_setpoint: find(dcwNodes, /mold-temp/),
+}
+const obs = Object.values(NODE).filter(Boolean)
+check('闭环 Channel + 产线 + 观测/受控节点齐备', Boolean(twin && line && product && recipe) && obs.length >= 8, `members=${members.length} nodes=${obs.length}`)
+check('Channel 为 v0.7.46 实例并启用 hybrid_twin(57 工具面)', members.length >= 3, `members=${members.length}`)
+
+// ── 2. 绑定(模板实例化不自带;按验收文档,场景与绑定由实例化方注入) ──
+let bound = 0
+for (const m of members) {
+  for (const n of Object.values(NODE).slice(0, 5)) {
+    const r = await j('POST', '/api/workshop/agent-tools/bindings', { agentId: m.id, nodeId: n.id, kind: 'daq', mode: 'auto' }, token)
+    if (r.code === 0) bound += 1
+  }
+  for (const n of [NODE.hold_pressure, NODE.hold_time, NODE.melt_temperature_setpoint]) {
+    const r = await j('POST', '/api/workshop/agent-tools/bindings', { agentId: m.id, nodeId: n.id, kind: 'dcw', mode: 'manual' }, token)
+    if (r.code === 0) bound += 1
+  }
+}
+check('场景节点绑定注入(全体成员 × 观测/受控节点)', bound >= members.length * 5, `bound=${bound}`)
+const lead = members.find(m => m.role === 'lead')
+const list = await j('GET', `/api/workshop/agent-tools/list?agentId=${lead.id}`, undefined, token)
+const toolNames = (list.data?.tools ?? []).map(t => t.name)
+check('闭环 Channel 工具面含孪生/MPC 6 件套 + 工业工具', ['twin_scene_read', 'twin_snapshot_create', 'twin_trial_run', 'mpc_optimize', 'twin_gate_evaluate', 'daq_query', 'line_context'].every(t => toolNames.includes(t)), `tools=${toolNames.length}`)
+const invoke = async (tool, args) => (await j('POST', '/api/workshop/agent-tools/invoke', { agentId: lead.id, tool, args }, token)).data?.result ?? {}
+
+// ── 3. 场景契约 + 真实样本 → 快照 ──
+const sceneRes = await invoke('twin_scene_read', {})
+const sceneText = String(sceneRes.text ?? '')
+const scene0 = firstJson(sceneText)
+check('twin_scene_read 返回内置可执行场景契约', /场景契约/.test(sceneText) && Boolean(scene0?.controls?.length), `${scene0?.sceneId}@${scene0?.sceneVersion} controls=${scene0?.controls?.length}`)
+check('发现内置场景无 nodeId(必须由调用方注入真实绑定)', (scene0?.observations ?? []).every(o => !o.nodeId) && (scene0?.states ?? []).every(s => !s.nodeId), `obs=${scene0?.observations?.length} states=${scene0?.states?.length} 均无 nodeId`)
+
+// 注入真实绑定 + 真实三元组
+const scene = {
+  ...scene0,
+  lineId: line.id,
+  productId: product.id,
+  recipeId: recipe.id,
+  observations: scene0.observations.map(o => ({ ...o, nodeId: NODE[o.id]?.id })),
+  states: scene0.states.map(s => ({ ...s, nodeId: NODE[s.id]?.id })),
+}
+
+const pg = new Client({ host: '127.0.0.1', port: 5432, user: 'postgres', password: 'awshop', database: 'awshop' })
+await pg.connect()
+// VirtualTrial 的 assertFreshSnapshot 只认 60s 内的水位:采样必须在产线运行中完成,
+// 所以先无条件停线再开跑(投影里没有 activeRunId,靠它判断会漏停 → start 409)。
+await j('POST', `/api/workshop/dcw/lines/${line.id}/stop`, {}, token)
+await sleep(2500)
+const st = await j('POST', `/api/workshop/dcw/lines/${line.id}/start`, { recipeId: recipe.id }, token)
+console.log(`  · 开跑采样:start status=${st.status} ${st.code ?? ''} ${st.message ?? 'ok'}`)
+await sleep(8000)
+const latestOf = async (nodeId) => {
+  const r = await pg.query('SELECT ts, value FROM daq_samples WHERE node_id = $1 ORDER BY ts DESC LIMIT 1', [nodeId])
+  return r.rows[0] ? { at: new Date(r.rows[0].ts).getTime(), value: Number(r.rows[0].value) } : null
+}
+const required = [...scene.observations, ...scene.states].map(x => x.nodeId)
+const samples = []
+const observed = {}
+for (const nodeId of required) {
+  const p = await latestOf(nodeId)
+  if (p) {
+    samples.push({ nodeId, at: p.at, value: p.value, sequence: `daq:${nodeId}:${p.at}` })
+    observed[nodeId] = p.value
+  }
+}
+await pg.end()
+const newest = Math.max(...samples.map(s => s.at))
+check('取到真实 DAQ 样本(场景所需节点全覆盖)', samples.length === required.length, `samples=${samples.length}/${required.length} 最新=${new Date(newest).toISOString().slice(11, 19)} 距今=${Math.round((Date.now() - newest) / 1000)}s`)
+
+const states = Object.fromEntries(scene.states.map(s => [s.id, observed[s.nodeId]]))
+// 受控量基线必须取**开跑后**再读一次的真实值:开跑瞬间 DCW 回读还停在压力 PV(实测 ~52 bar),
+// 拿它当基线的物理预测克重只有 30.99 g,会低于场景硬约束 31 g,65 组候选全被拒。
+const dcwFresh = (await j('GET', '/api/workshop/dcw', undefined, token)).data
+const dcwNodesFresh = (dcwFresh.nodes ?? []).filter(n => n.lineId === line.id)
+const writesBefore = Number(dcwFresh.controller?.writesTotal ?? 0)
+const FRESH = {
+  hold_pressure: find(dcwNodesFresh, /hold-pressure/),
+  hold_time: find(dcwNodesFresh, /hold-time/),
+  melt_temperature_setpoint: find(dcwNodesFresh, /mold-temp/),
+}
+const controls = Object.fromEntries(scene.controls.map((c) => {
+  const raw = Number(FRESH[c.id]?.readValue ?? FRESH[c.id]?.value ?? (c.id === 'hold_pressure' ? 65 : c.id === 'hold_time' ? 8 : 247))
+  return [c.id, clamp(raw, c.min ?? -Infinity, c.max ?? Infinity)]
+}))
+console.log(`  · 受控量基线:${JSON.stringify(controls)}(hold-pressure-sp 命令值=${FRESH.hold_pressure?.value} 回读=${FRESH.hold_pressure?.readValue})`)
+const snapArgs = { scene_json: scene, channel_id: twin.id, phase: 'holding', controls, states, samples, freshness_max_ms: 300_000 }
+const snapRes = await invoke('twin_snapshot_create', snapArgs)
+const snapText = String(snapRes.text ?? '')
+const snapshotId = (snapText.match(/snapshot_id:\s*(\S+)/) ?? [])[1]
+const snapFsPath = `${HOME}/aml/twins/snapshots/${snapshotId}/snapshots.json`
+const snapshot = existsSync(snapFsPath) ? JSON.parse(readFileSync(snapFsPath, 'utf8')) : null
+check('TwinSnapshot 创建且数据新鲜(fresh=true / completeness=1)', /fresh:\s*true/.test(snapText) && snapshot?.dataQuality?.completeness === 1, `${snapshotId} ${(snapText.match(/fresh:.*/) ?? [''])[0]}`)
+
+// ── 4. 真实 VirtualTrial 扫描(27 组小步候选,全部真实物理内核) ──
+const baseline = { ...controls }
+const trials = []
+const rejected = []
+const objective = { schemaVersion: 1, createdAt: new Date().toISOString(), createdBy: lead.id, objectiveId: 'weight-quality', targets: { weight: 32.5 }, weights: { weight: 1 }, controlCosts: {}, horizonSteps: 4, trustRegion: {} }
+const steps = scene.controls.map(c => c.maxStep ?? 1)
+// 门禁 G6 要求 ≥10 条"无硬约束违反"的真实候选:单档 ±1 只能凑出 9 条安全候选(实测),
+// 因此按 ±1 / ±0.5 两档步长扫描(4^3 + 基线 = 65 组),凑够 12 条安全候选即停。
+const offsets = [-1, -0.5, 0.5, 1]
+const cands = [{ ...baseline }]
+for (const a of offsets) {
+  for (const b of offsets) {
+    for (const c2 of offsets) {
+      cands.push({
+        hold_pressure: clamp(baseline.hold_pressure + a * steps[0], 50, 90),
+        hold_time: clamp(baseline.hold_time + b * steps[1], 4, 14),
+        melt_temperature_setpoint: clamp(baseline.melt_temperature_setpoint + c2 * steps[2], 235, 260),
+      })
+    }
+  }
+}
+for (const cand of cands) {
+  if (trials.length >= 12) break
+  const res = await invoke('twin_trial_run', { scene_json: scene, snapshot_json: snapshot, baseline_controls: baseline, candidate_controls: [cand], objective, model_id: process.env.AW_MODEL_ID ?? '' })
+  const text = String(res.text ?? '')
+  const trialId = (text.match(/trial_id:\s*(\S+)/) ?? [])[1]
+  const p = trialId ? join(HOME, 'aml', 'twins', 'trials', trialId, 'trials.json') : ''
+  if (p && existsSync(p)) {
+    const t = JSON.parse(readFileSync(p, 'utf8')).trial
+    if (t.constraintResults.every(x => x.passed)) trials.push(t)
+    else rejected.push(t.trialId)
+  }
+}
+const anyTrial = trials[0] ?? null
+check('VirtualTrial 真实执行且候选未执行(candidateExecuted=false)', trials.length >= 10 && anyTrial?.candidateExecuted === false, `安全候选=${trials.length} 硬约束拒绝=${rejected.length} 共扫描 ${cands.length} 组`)
+check('VirtualTrial 记录推荐证书/约束结论', Boolean(anyTrial?.provenance?.trialHash) && Array.isArray(anyTrial?.constraintResults), `constraints=${anyTrial?.constraintResults?.length} improvement=${anyTrial?.baselineComparison?.improvement?.toFixed?.(4)}`)
+
+// ── 5. 数据集/模型真实指标 → 12 项推荐门禁 ──
+const datasets = ((await j('GET', '/api/workshop/aml/datasets', undefined, token)).data?.datasets ?? []).filter(d => d.lineId === line.id)
+const dsRow = datasets.sort((a, b) => String(b.createdAt).localeCompare(String(a.createdAt)))[0]
+const runs = JSON.parse(dsRow?.runIdsJson || '[]').length
+const rows = Number(dsRow?.rowCount ?? 0)
+const exps = ((await j('GET', `/api/workshop/aml/experiments?datasetId=${dsRow.id}`, undefined, token)).data?.experiments ?? [])
+const metrics = JSON.parse(exps[0]?.metricsJson || '{}')
+const oneStep = Number(metrics.oneStepTest?.nrmse ?? 1)
+const rollout = Number(metrics.rolloutTest?.nrmse ?? 1)
+const valN = Number(metrics.oneStepVal?.nrmse ?? oneStep)
+const gap = valN > 0 ? Math.abs(oneStep - valN) / valN : 1
+const models = ((await j('GET', '/api/workshop/aml/models', undefined, token)).data?.models ?? [])
+const model = models.find(m => m.datasetId === dsRow.id && m.stage === 'production') ?? models.find(m => m.stage === 'production') ?? models[0]
+check('真实数据集 + 真实训练指标可解析(非伪造)', rows >= 500 && runs >= 3 && Number.isFinite(oneStep) && oneStep < 1, `dataset=${dsRow?.id} rows=${rows} runs=${runs} oneStep=${oneStep.toFixed(4)} rollout=${rollout.toFixed(4)} gap=${(gap * 100).toFixed(2)}%`)
+
+const gateArgs = {
+  rows,
+  runs,
+  one_step_nrmse: oneStep,
+  rollout_nrmse: rollout,
+  val_test_gap: gap,
+  calibration_rows: rows,
+  coverage: 1,
+  candidate_trials: trials.slice(0, 10),
+  physics_failure_rate: 0,
+  model_id: model?.id ?? '',
+}
+const gateRes = await invoke('twin_gate_evaluate', gateArgs)
+const gateText = String(gateRes.text ?? '')
+const gateObj = firstJson(gateText) ?? {}
+const ids = (gateObj.checks ?? []).map(c => c.id)
+check('twin_gate_evaluate 返回 12 项判据明细', ids.length >= 12, `checks=${ids.length} passed=${gateObj.passed}`)
+check('推荐门禁在真实指标下全绿(write_eligible)', gateObj.passed === true, `stage=${gateObj.stage} reject=${JSON.stringify(gateObj.rejectCodes ?? [])}`)
+
+// ── 6. 模型 twinEligibility 回写 → mpc 升档 precise_search(闭环) ──
+if (model?.id) {
+  // 工具把 modelUpdate 以字符串嵌进 JSON,内部引号在文本里是转义形态(\\"),比较前先反转义
+  const flatGate = gateText.replace(/\\"/g, '"')
+  const elig = /model_twin_eligibility:\s*(\{.*?\})/.exec(flatGate.replace(/\n/g, ' '))
+  check('模型 twinEligibility 已回写(gatePassed/recommendationEligible=true)', /"recommendationEligible":\s*true/.test(flatGate), elig ? elig[1].slice(0, 120) : '未回写')
+  const mpc = await invoke('mpc_optimize', { scene_json: scene, baseline_controls: baseline, horizon_steps: 4, model_id: model.id })
+  const mpcText = String(mpc.text ?? '')
+  const mpcObj = firstJson(mpcText) ?? {}
+  check('mpc_optimize 输出 recommendation-only 报告', mpcObj.recommendationOnly === true, `mode=${mpcObj.mode} candidates=${mpcObj.candidatesEvaluated}`)
+  // mode 由模型门禁决定(安全语义升级);preciseSearchAllowed 还要求本轮存在可行候选,
+  // 真实基线下可能为 false(候选全被硬约束拒),因此分开记录而不是当作失败。
+  check('模型过门禁后 mpc 升档 precise_search(推荐而非直写)', mpcObj.mode === 'precise_search', `mode=${mpcObj.mode} preciseSearchAllowed=${mpcObj.preciseSearchAllowed} bestCost=${mpcObj.bestCost}`)
+}
+else if (REQUIRE_MODEL) {
+  check('生产模型可用(阶段 6 前置)', false, 'registry 无模型,无法完成模型→孪生闭环')
+}
+else {
+  skip('模型 → twinEligibility → precise_search 闭环', '注册表暂无模型(阶段 6 尚在训练)')
+}
+
+// ── 7. recommendation-only:零真实写入(基线在开跑后采集,排除 start 自身的写入) ──
+const dcwAfter = (await j('GET', '/api/workshop/dcw', undefined, token)).data
+const writesAfter = Number(dcwAfter.controller?.writesTotal ?? 0)
+check('recommendation-only:链路未产生 DCW 写入', writesAfter === writesBefore, `writesTotal ${writesBefore} → ${writesAfter};candidateExecuted=false`)
+await j('POST', `/api/workshop/dcw/lines/${line.id}/stop`, {}, token)
+
+console.log(`\n★ 阶段 7:${pass} 通过 / ${fails.length} 失败${skips.length ? ` / ${skips.length} 跳过` : ''}${fails.length ? ` (${fails.join('; ')})` : ''}`)
+console.log(JSON.stringify({ channelId: twin.id, leadId: lead.id, snapshotId, samples: samples.length, trials: trials.length, rejected: rejected.length, gatePassed: gateObj.passed, modelId: model?.id ?? null, writesBefore, writesAfter }, null, 1))

package/scripts/_aw0746-train-strong.py

@@ -0,0 +1,142 @@
+# -*- coding: utf-8 -*-
+"""0.7.46 验收用加强训练实现(内联 code 路径,仍走 amlkit 契约)。
+
+与平台参考实现 `train-example.py` 的差异:
+  1. 平台 bundle 的归一化对 σ<1 的列按 std=1 兜底(见 dataset-builder.meanStd),
+     于是熔温/克重这类物理量保持原始量纲,MLP 条件数极差 → 本实现按**训练集自身**
+     重新 z-score(输入、输出各一套),并在导出前把输出层折回契约量纲;
+  2. 两层 MLP + ReLU(参考实现是单隐层),Adam + 余弦退火;
+  3. 保存 val 最优权重并在结束时回滚(best-val checkpoint),抑制后期过拟合。
+
+契约不变:输入 [N,H,nAll](契约归一化空间)→ 输出 [N,nTgt](契约归一化空间的下一拍目标)。
+超参:lr / epochs / hidden / batch。
+"""
+import json
+import math
+import os
+from pathlib import Path
+
+import torch
+import torch.nn as nn
+
+import amlkit
+
+
+class ZScoreNet(nn.Module):
+    """内部 z-score 标准化 + 两层 MLP;forward 输出折回契约量纲(mean/std 由训练集统计)。"""
+
+    def __init__(self, n_in, hidden, n_out, x_mean, x_std, y_mean, y_std):
+        super().__init__()
+        self.register_buffer("x_mean", torch.tensor(x_mean, dtype=torch.float32))
+        self.register_buffer("x_std", torch.tensor(x_std, dtype=torch.float32))
+        self.register_buffer("y_mean", torch.tensor(y_mean, dtype=torch.float32))
+        self.register_buffer("y_std", torch.tensor(y_std, dtype=torch.float32))
+        self.net = nn.Sequential(
+            nn.Flatten(),
+            nn.Linear(n_in, hidden),
+            nn.ReLU(),
+            nn.Linear(hidden, hidden),
+            nn.ReLU(),
+            nn.Linear(hidden, n_out),
+        )
+
+    def forward(self, x):
+        z = (x - self.x_mean) / self.x_std
+        out = self.net(z)
+        return out * self.y_std + self.y_mean
+
+
+def main():
+    job_dir = Path(os.environ["AML_JOB_DIR"]).resolve()
+    job = json.loads((job_dir / "job.json").read_text(encoding="utf-8"))
+    params = job.get("params", {})
+
+    bundle = amlkit.load_bundle(job["datasetPath"])
+    manifest = bundle.manifest
+    h_steps = manifest["shapes"]["x"][1]
+    n_all = manifest["shapes"]["x"][2]
+    n_tgt = manifest["shapes"]["y"][2]
+
+    x_train = bundle.get("x", "train")
+    y_train = bundle.get("y", "train")[:, 0, :]
+    x_val = bundle.get("x", "val")
+    y_val = bundle.get("y", "val")[:, 0, :]
+    x_test = bundle.get("x", "test")
+    y_test = bundle.get("y", "test")[:, 0, :]
+
+    xt = torch.tensor(x_train, dtype=torch.float32)
+    yt = torch.tensor(y_train, dtype=torch.float32)
+    xv = torch.tensor(x_val, dtype=torch.float32) if x_val.shape[0] else None
+    yv = torch.tensor(y_val, dtype=torch.float32) if y_val.shape[0] else None
+
+    # 训练集统计量(不用 val/test,避免信息泄漏)
+    x_mean = xt.mean(dim=(0, 1), keepdim=False).reshape(-1).tolist()
+    x_std = torch.clamp(xt.reshape(-1, n_all).std(dim=0), min=1e-6).tolist()
+    y_mean = yt.mean(dim=0).tolist()
+    y_std = torch.clamp(yt.std(dim=0), min=1e-6).tolist()
+
+    hidden = int(params.get("hidden", 128))
+    model = ZScoreNet(h_steps * n_all, hidden, n_tgt, x_mean, x_std, y_mean, y_std)
+    opt = torch.optim.Adam(model.parameters(), lr=float(params.get("lr", 3e-3)))
+    epochs = int(params.get("epochs", 400))
+    batch = int(params.get("batch", 64))
+    sched = torch.optim.lr_scheduler.CosineAnnealingLR(opt, T_max=max(1, epochs))
+    loss_fn = nn.MSELoss()
+
+    n = xt.shape[0]
+    best_val = float("inf")
+    best_state = None
+    train_loss = 0.0
+    val_loss = None
+    for epoch in range(epochs):
+        model.train()
+        perm = torch.randperm(n)
+        total = 0.0
+        for i in range(0, n, batch):
+            idx = perm[i:i + batch]
+            opt.zero_grad()
+            loss = loss_fn(model(xt[idx]), yt[idx])
+            loss.backward()
+            opt.step()
+            total += float(loss) * idx.shape[0]
+        sched.step()
+        train_loss = total / max(1, n)
+        if xv is not None:
+            model.eval()
+            with torch.no_grad():
+                val_loss = float(loss_fn(model(xv), yv))
+            if val_loss < best_val:
+                best_val = val_loss
+                best_state = {k: v.detach().clone() for k, v in model.state_dict().items()}
+        amlkit.report_progress(
+            5 + 90 * (epoch + 1) // epochs,
+            "epoch %d/%d train=%.6f val=%s" % (epoch + 1, epochs, train_loss, val_loss if val_loss is not None else "n/a"),
+        )
+
+    if best_state is not None:
+        model.load_state_dict(best_state)
+
+    test_rmse = None
+    if x_test.shape[0]:
+        model.eval()
+        with torch.no_grad():
+            pred = model(torch.tensor(x_test, dtype=torch.float32))
+            test_rmse = float(torch.sqrt(torch.mean((pred - torch.tensor(y_test, dtype=torch.float32)) ** 2)))
+
+    amlkit.save_metrics({"agent": {
+        "kind": "zscore_mlp_two_layer_bestval",
+        "epochs": epochs,
+        "train_loss_last": train_loss,
+        "val_loss_best": None if best_val == float("inf") else best_val,
+        "test_rmse_normalized": test_rmse,
+        "x_std_min": min(x_std),
+        "params": params,
+    }})
+
+    amlkit.export_torch_onnx(model, manifest)
+    amlkit.report_progress(99, "exported onnx (zscore 2-layer)")
+    print("train-strong done rmse=%s" % (test_rmse if test_rmse is not None else "n/a"))
+
+
+if __name__ == "__main__":
+    main()

package/scripts/_aw0747-readme-rows.mjs

@@ -0,0 +1,32 @@
+/**
+ * 临时:把 0.7.47 打包实测行追加到 README 双语验收表(保持两语文案对齐)
+ * 用法:node scripts/_aw0747-readme-rows.mjs
+ */
+import { readFileSync, writeFileSync } from 'node:fs'
+
+const zhRows = [
+  '| `_aw0746-stage12.mjs` | **10 / 0**（2026-09-25,全局安装的 **v0.7.47 打包系统**） | 健康门(version=0.7.47)→ 管理员登录/角色面 → 创建 Channel(真实 `omp` lead)→ 开群聊 → **真实 omp 群聊回复** → 系统监控可见 omp 子进程 | `AW_BASE=<base> node scripts/_aw0746-stage12.mjs` |',
+  '| `_aw0746-stage34.mjs` | **11 / 0**（2026-09-25,v0.7.47 打包系统 + PLC 模拟器） | 模拟器建线（11 数控 + 14 数采）→ 产品/配方三元组 → 开跑后**样本逐条带 recipe_id/run_id** → `daq_query` **默认只取当前活动批次**（`scope=all` 可放开，实测 1 vs 4）→ 实例化 Hybrid Twin 模板（lead + 3 worker,注入孪生 6 件套 / 总 57 工具） | `AW_BASE=<base> node scripts/_aw0746-stage34.mjs` |',
+  '| `_aw0746-stage6b-train-sweep.mjs` | 门禁**全绿**（2026-09-25,v0.7.46 打包系统,真实 torch 训练） | 1 s 拍格数据集（**1724 行 / 10 批次**）→ 内联 code 作业真实训练 → G1 单步 **0.0992**（≤0.10）/ G2 滚动 0.1092（≤0.25）/ G3 泛化 14.6%（≤20%）/ G4 行数与批次 → 模型注册并**晋升 production** | `AW_BASE=<base> AW_DATASET_ID=<ds> node scripts/_aw0746-stage6b-train-sweep.mjs` |',
+  '| `_aw0746-stage7-twin.mjs` | **17 / 0**（2026-09-25,v0.7.47 打包系统,真实 DAQ + 真实模型） | 真实样本 → TwinSnapshot `fresh=true` → 65 组真实 VirtualTrial（12 条安全候选,`candidateExecuted=false`）→ 12 项推荐门禁 **write_eligible** → 门禁结论回写模型 `twinEligibility` → `mpc_optimize` 升档 **precise_search** → 全程 **0 DCW 写入** | `AW_BASE=<base> AW_REQUIRE_MODEL=1 node scripts/_aw0746-stage7-twin.mjs` |',
+  '| `_aw0746-stage6c-jobsmoke.mjs` | **5 / 0**（2026-09-25,v0.7.47 打包系统） | AML 作业链路：内联 `code` 被接受（不再要求 `workspace/train.py`）→ 作业到终态 → 训练日志/ONNX 痕迹 → 实验与逐项门禁落库 | `AW_BASE=<base> node scripts/_aw0746-stage6c-jobsmoke.mjs` |',
+]
+const enRows = [
+  '| `_aw0746-stage12.mjs` | **10 / 0** (2026-09-25, globally installed **v0.7.47 packaged build**) | health gate (version=0.7.47) → admin login/role surface → channel creation with a real `omp` lead → group chat opened → **real omp reply** → omp child process visible in monitoring | `AW_BASE=<base> node scripts/_aw0746-stage12.mjs` |',
+  '| `_aw0746-stage34.mjs` | **11 / 0** (2026-09-25, v0.7.47 packaged build + PLC simulator) | line provisioned from the simulator (11 DCW + 14 DAQ) → product/recipe triple → **every sample tagged with recipe_id/run_id** while running → `daq_query` **defaults to the active batch** (`scope=all` opts out, measured 1 vs 4) → Hybrid Twin template instantiated (lead + 3 workers, 6 twin tools injected, 57 total) | `AW_BASE=<base> node scripts/_aw0746-stage34.mjs` |',
+  '| `_aw0746-stage6b-train-sweep.mjs` | gates **all green** (2026-09-25, v0.7.46 packaged build, real torch training) | 1 s grid dataset (**1724 rows / 10 runs**) → inline-code job trained for real → G1 one-step **0.0992** (≤0.10) / G2 rollout 0.1092 (≤0.25) / G3 generalisation 14.6% (≤20%) / G4 rows+runs → model registered and **promoted to production** | `AW_BASE=<base> AW_DATASET_ID=<ds> node scripts/_aw0746-stage6b-train-sweep.mjs` |',
+  '| `_aw0746-stage7-twin.mjs` | **17 / 0** (2026-09-25, v0.7.47 packaged build, real DAQ + real model) | real samples → TwinSnapshot `fresh=true` → 65 real VirtualTrials (12 safe candidates, `candidateExecuted=false`) → 12-check recommendation gate **write_eligible** → verdict written back to the model as `twinEligibility` → `mpc_optimize` upgrades to **precise_search** → **0 DCW writes** throughout | `AW_BASE=<base> AW_REQUIRE_MODEL=1 node scripts/_aw0746-stage7-twin.mjs` |',
+  '| `_aw0746-stage6c-jobsmoke.mjs` | **5 / 0** (2026-09-25, v0.7.47 packaged build) | AML job path: inline `code` accepted (no `workspace/train.py` required) → job reaches a terminal state → training logs/ONNX traces → experiment plus per-check gates persisted | `AW_BASE=<base> node scripts/_aw0746-stage6c-jobsmoke.mjs` |',
+]
+
+for (const [file, anchor, rows] of [
+  ['README-zh.md', '| 前端验收 |', zhRows],
+  ['README.md', '| Front-end acceptance |', enRows],
+]) {
+  const lines = readFileSync(file, 'utf8').split('\n')
+  const i = lines.findIndex(l => l.startsWith(anchor))
+  if (i < 0) throw new Error(`anchor not found in ${file}`)
+  lines.splice(i + 1, 0, ...rows)
+  writeFileSync(file, lines.join('\n'), 'utf8')
+  console.log(`inserted ${rows.length} rows after line ${i + 1} in ${file}`)
+}

package/scripts/_aw0748-failclosed-probe.mjs

@@ -0,0 +1,90 @@
+/**
+ * 0.7.48 打包系统验收 · 硬约束失败关闭探针(无需 LLM,秒级)
+ * 场景:同一份真实快照 + 同一组候选,只改 constraints[].id
+ *   ① id = weight(可映射)→ 试验正常给出约束结论
+ *   ② id = part_weight(不可映射)→ 必须**判不通过**并给出"无法映射"明细(v0.7.47 会假通过)
+ * 用法:AW_BASE=http://127.0.0.1:3001 [AW_HOME=<home>] node scripts/_aw0748-failclosed-probe.mjs
+ */
+import { existsSync, readFileSync, readdirSync, statSync } from 'node:fs'
+
+const BASE = process.env.AW_BASE ?? 'http://127.0.0.1:3001'
+const HOME = process.env.AW_HOME ?? `${process.cwd()}/.aw0746-home`
+const ADMIN = { email: process.env.AW_ADMIN_EMAIL ?? 'admin@awshop.local', password: process.env.AW_ADMIN_PASSWORD ?? 'Awshop@123!' }
+
+let pass = 0
+const fails = []
+const check = (name, ok, detail = '') => {
+  console.log(`  ${ok ? 'PASS' : 'FAIL'}  ${name}${detail ? ` — ${detail}` : ''}`)
+  if (ok) pass += 1
+  else fails.push(name)
+}
+const j = async (method, path, body, token) => {
+  const r = await fetch(BASE + path, {
+    method,
+    headers: { 'content-type': 'application/json', ...(token ? { authorization: `Bearer ${token}` } : {}) },
+    body: body === undefined ? undefined : JSON.stringify(body),
+    signal: AbortSignal.timeout(120_000),
+  })
+  return { status: r.status, ...(await r.json().catch(() => ({}))) }
+}
+const firstJson = (text) => {
+  const a = text.indexOf('{')
+  const b = text.lastIndexOf('}')
+  return a >= 0 && b > a ? JSON.parse(text.slice(a, b + 1)) : null
+}
+
+const token = (await j('POST', '/api/users/login', ADMIN)).data?.token
+console.log(`\n═══ 0.7.48 硬约束失败关闭探针 @ ${BASE} ═══`)
+
+// 复用最近一次真实快照(阶段 7 / 阶段 5 产物)
+const snapRoot = `${HOME}/aml/twins/snapshots`
+const dirs = existsSync(snapRoot)
+  ? readdirSync(snapRoot).map(d => ({ d, m: statSync(`${snapRoot}/${d}`).mtimeMs })).sort((a, b) => b.m - a.m)
+  : []
+check('存在真实 TwinSnapshot 工件(取最新)', dirs.length > 0, dirs[0]?.d ?? '(无)')
+if (!dirs.length) {
+  console.log('\n★ 无可复用快照,先跑阶段 7/5 生成')
+  process.exit(2)
+}
+const snapshot = JSON.parse(readFileSync(`${snapRoot}/${dirs[0].d}/snapshots.json`, 'utf8'))
+check('快照新鲜且水位>0(可直接跑试验)', snapshot.dataQuality?.fresh === true && Number(snapshot.daqWatermark) > 0, `fresh=${snapshot.dataQuality?.fresh} watermark=${snapshot.daqWatermark}`)
+
+const chans = (await j('GET', '/api/workshop/channels', undefined, token)).data ?? []
+const twin = chans.find(c => /闭环优化-hybridtwin/.test(c.name))
+const members = (await j('GET', `/api/workshop/channels/${twin.id}/agents`, undefined, token)).data ?? []
+const lead = members.find(m => m.role === 'lead')
+const invoke = async (tool, args) => (await j('POST', '/api/workshop/agent-tools/invoke', { agentId: lead.id, tool, args }, token)).data?.result ?? {}
+
+const baseScene = firstJson(String((await invoke('twin_scene_read', {})).text ?? ''))
+const controls = { hold_pressure: 65, hold_time: 8, melt_temperature_setpoint: 247 }
+const objective = { schemaVersion: 1, createdAt: new Date().toISOString(), createdBy: lead.id, objectiveId: 'weight-quality', targets: { weight: 32.5 }, weights: { weight: 1 }, controlCosts: {}, horizonSteps: 1, trustRegion: {} }
+const trialOf = async (scene) => {
+  const res = await invoke('twin_trial_run', { scene_json: scene, snapshot_json: snapshot, baseline_controls: controls, candidate_controls: [{ ...controls, hold_pressure: 67 }], objective })
+  const text = String(res.text ?? '')
+  const id = (text.match(/trial_id:\s*(\S+)/) ?? [])[1]
+  const artifact = id ? JSON.parse(readFileSync(`${HOME}/aml/twins/trials/${id}/trials.json`, 'utf8')) : null
+  return { text, trial: artifact?.trial ?? null }
+}
+
+// ① 可映射 id:baseline 预测克重 32.5 g 落在 31~34 内 → 应通过
+const mapped = { ...baseScene, lineId: snapshot.lineId, productId: snapshot.productId, recipeId: snapshot.recipeId, constraints: [{ id: 'weight', kind: 'hard_range', min: 31, max: 34 }] }
+const okTrial = await trialOf(mapped)
+const okWeight = okTrial.trial?.constraintResults?.[0]
+check('可映射约束 id(weight)正常评估', Boolean(okTrial.trial) && okWeight?.id === 'weight' && okWeight.passed === true, `weight=${okTrial.trial?.predictedTrajectory?.[0]?.weight?.toFixed?.(3)} detail=${okWeight?.detail}`)
+
+// ② 不可映射 id:必须失败关闭(旧版本这里会返回"全轨迹通过")
+const unmapped = { ...mapped, constraints: [{ id: 'part_weight', kind: 'hard_range', min: 31, max: 34 }] }
+const badTrial = await trialOf(unmapped)
+const badCon = badTrial.trial?.constraintResults?.[0]
+const allPassed = (badTrial.trial?.constraintResults ?? []).every(c => c.passed)
+check('不可映射约束 id 判为不通过(失败关闭)', Boolean(badTrial.trial) && badCon?.passed === false && /无法映射/.test(String(badCon?.detail)), `passed=${badCon?.passed} detail=${String(badCon?.detail).slice(0, 80)}`)
+check('候选不再被当作安全(constraints_passed=false)', allPassed === false, `constraints_passed=${allPassed}`)
+check('工具文本同步反映未通过', /constraints_passed:\s*false/.test(badTrial.text), (badTrial.text.match(/constraints_passed:.*/) ?? [''])[0])
+
+// ③ 空轨迹(候选为空)同样失败关闭
+const empty = await invoke('twin_trial_run', { scene_json: mapped, snapshot_json: snapshot, baseline_controls: controls, candidate_controls: [], objective })
+const emptyTrialId = (String(empty.text ?? '').match(/trial_id:\s*(\S+)/) ?? [])[1]
+const emptyTrial = emptyTrialId ? JSON.parse(readFileSync(`${HOME}/aml/twins/trials/${emptyTrialId}/trials.json`, 'utf8')).trial : null
+check('空轨迹判为不通过(不再默认通过)', emptyTrial ? emptyTrial.constraintResults.every(c => c.passed === false) : false, `details=${(emptyTrial?.constraintResults ?? []).map(c => c.detail).join(' | ').slice(0, 80)}`)
+
+console.log(`\n★ 0.7.48 探针:${pass} 通过 / ${fails.length} 失败${fails.length ? ` (${fails.join('; ')})` : ''}`)

package/scripts/_plc-sim-detached.cjs

@@ -0,0 +1,22 @@
+/**
+ * 临时:detached 拉起 PLC 节点模拟器(API :4010 + Modbus TCP :16040 等六协议)。
+ * 用法:node scripts/_plc-sim-detached.cjs
+ */
+const { spawn } = require('node:child_process')
+const fs = require('node:fs')
+const path = require('node:path')
+
+const repo = path.resolve(__dirname, '..')
+const simDir = path.join(repo, 'plc-node-simulator')
+const logDir = path.join(repo, '.e2e-matrix')
+fs.mkdirSync(logDir, { recursive: true })
+const out = fs.openSync(path.join(logDir, 'plc-sim-0746.out.log'), 'a')
+const tsx = path.join(simDir, 'node_modules', 'tsx', 'dist', 'cli.mjs')
+const p = spawn(process.execPath, [tsx, 'src/server/index.ts'], {
+  detached: true,
+  stdio: ['ignore', out, out],
+  env: { ...process.env, SIM_PORT: process.env.SIM_PORT ?? '4010' },
+  cwd: simDir,
+})
+p.unref()
+console.log(`[plc-sim] spawned pid=${p.pid}`)