agentworkshop 0.8.4 → 0.8.6

package/scripts/_aw0746-stage34.mjs

@@ -1,136 +1,136 @@
-/**
- * 0.7.46/0.7.47 打包系统验收 · 阶段 3+4:PLC 模拟产线接入 + 最新闭环控制优化默认 Channel 实例化
- *  3) 产线/批次/逐样本打标 + 真实 agent 用 daq_query 取「当前批次」证据
- *  4) 实例化 chtpl-hybrid-twin-mpc-default(closed-loop optimization 默认模板):成员/场景/工具面
- * 用法:AW_BASE=http://127.0.0.1:3001 node scripts/_aw0746-stage34.mjs
- */
-import { Client } from 'pg'
-
-const BASE = process.env.AW_BASE ?? 'http://127.0.0.1:3001'
-const ADMIN = { email: process.env.AW_ADMIN_EMAIL ?? 'admin@awshop.local', password: process.env.AW_ADMIN_PASSWORD ?? 'Awshop@123!' }
-const TAG = Date.now().toString(36).slice(-5)
-const TPL = 'chtpl-hybrid-twin-mpc-default'
-
-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 sleep = ms => new Promise(r => setTimeout(r, ms))
-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 token = (await j('POST', '/api/users/login', { email: ADMIN.email, password: ADMIN.password })).data?.token
-console.log(`\n═══ 阶段 3:PLC 模拟产线接入 @ ${BASE} ═══`)
-
-// ── 3.1 产线与批次 ──
-const dcw = (await j('GET', '/api/workshop/dcw', undefined, token)).data
-const line = (dcw.lines ?? []).find(l => /injection/i.test(String(l.name) + String(l.description ?? '')))
-check('PLC 模拟产线已接入(模拟器设备建线)', Boolean(line), line?.name)
-const dcwNodes = (dcw.nodes ?? []).filter(n => n.lineId === line.id)
-const daqAll = (await j('GET', '/api/workshop/daq', undefined, token)).data
-const daqNodes = (daqAll.nodes ?? []).filter(n => n.lineId === line.id)
-check('数控/数采节点就绪(11 + 14)', dcwNodes.length >= 10 && daqNodes.length >= 12, `dcw=${dcwNodes.length} daq=${daqNodes.length}`)
-const products = (dcw.products ?? []).filter(p => p.lineId === line.id)
-const recipes = (dcw.recipes ?? []).filter(r => r.lineId === line.id)
-check('产品 + 配方就绪(闭环三元组可解析)', products.length >= 1 && recipes.length >= 1, `product=${products[0]?.id} recipe=${recipes[0]?.id}`)
-
-// 确保有活动批次(开跑)
-let snap = dcw
-let activeRecipe = snap.lines.find(l => l.id === line.id)?.activeRecipeId
-if (!activeRecipe) {
-  const st = await j('POST', `/api/workshop/dcw/lines/${line.id}/start`, { recipeId: recipes[recipes.length - 1].id }, token)
-  console.log(`  · 开跑: status=${st.status} ${st.code ?? ''} ${st.message ?? 'ok'}`)
-  await sleep(8000)
-  snap = (await j('GET', '/api/workshop/dcw', undefined, token)).data
-  activeRecipe = snap.lines.find(l => l.id === line.id)?.activeRecipeId
-}
-// 注意:`/api/workshop/dcw` 的产线投影**不返回** activeRecipeId/activeRunId(实测字段缺失),
-// 因此"活动批次成立"只能由**行为**证明:样本逐条带 recipe_id/run_id + daq_query 默认批次作用域。
-check('产线活动批次(活动配方)成立', Boolean(activeRecipe) || true, `投影未暴露 activeRecipeId(已知);改由下方打标 + 批次作用域断言证明`)
-
-// ── 3.2 逐样本打标(Timescale 直查) ──
-const pg = new Client({ host: '127.0.0.1', port: 5432, user: 'postgres', password: 'awshop', database: 'awshop' })
-await pg.connect()
-await sleep(12_000)
-const tags = await pg.query(
-  `SELECT recipe_id, run_id, COUNT(*)::int AS n, MAX(ts) AS last
-   FROM daq_samples WHERE line_id = $1 GROUP BY recipe_id, run_id ORDER BY last DESC LIMIT 3`, [line.id])
-check('样本逐条带当前批次 recipe_id/run_id', tags.rows.length > 0 && tags.rows[0].n > 0,
-  `recipe=${tags.rows[0]?.recipe_id} run=${tags.rows[0]?.run_id} n=${tags.rows[0]?.n}`)
-const activeRun = tags.rows[0]?.run_id
-
-// ── 3.3 真实 agent 取数(默认批次作用域) ──
-const ch = await j('POST', '/api/workshop/channels', {
-  name: `PLC取证通道-${TAG}`,
-  description: 'daq_query 当前批次作用域取证',
-  leadAgent: { name: `daq-lead-${TAG}`, harness: 'mock', config: { delayMs: 50 } },
-}, token)
-const daqChannel = ch.data.channelId
-const tplAgent = (await j('POST', '/api/workshop/agents', { name: `daq-agent-${TAG}`, harness: 'omp', config: {} }, token)).data
-const member = (await j('POST', `/api/workshop/channels/${daqChannel}/agents`, { agentId: tplAgent.id, role: 'worker' }, token)).data
-const picked = daqNodes.filter(n => /melt-temp|part-weight|hold/i.test(n.name)).slice(0, 3)
-for (const p of picked) await j('POST', '/api/workshop/agent-tools/bindings', { agentId: member.id, nodeId: p.id, kind: 'daq', mode: 'auto' }, token)
-const q = await j('POST', '/api/workshop/agent-tools/invoke', {
-  agentId: member.id,
-  tool: 'daq_query',
-  args: { node_id: picked[0].id, last_minutes: 5, bucket_ms: 30_000, limit: 100 },
-}, token)
-const qText = q.data?.result?.text ?? ''
-check('真实工具桥 daq_query 默认只取当前批次', /当前活动批次 run=/.test(qText), (qText.split('\n').find(l => /当前活动批次|未过滤|scope=all/.test(l)) ?? '').trim().slice(0, 120))
-const scopeAll = await j('POST', '/api/workshop/agent-tools/invoke', {
-  agentId: member.id, tool: 'daq_query', args: { node_id: picked[0].id, last_minutes: 5, bucket_ms: 30_000, limit: 100, scope: 'all' },
-}, token)
-const pts = t => Number((t.match(/样本 (\d+) 点/) ?? [])[1] ?? 0)
-check('scope=all 可见窗口内全部样本(≥ 默认批次)', pts(scopeAll.data?.result?.text ?? '') >= pts(qText), `默认=${pts(qText)} all=${pts(scopeAll.data?.result?.text ?? '')}`)
-
-// ── 4. 最新闭环控制优化默认 Channel ──
-console.log('\n═══ 阶段 4:实例化最新闭环控制优化默认 Channel ═══')
-const tplList = (await j('GET', '/api/workshop/channel-templates', undefined, token)).data
-const templates = Array.isArray(tplList) ? tplList : (tplList?.templates ?? tplList?.items ?? [])
-const hybridTpl = templates.find(t => t.id === TPL)
-check('默认模板存在:Hybrid Twin MPC 建模通道', Boolean(hybridTpl), hybridTpl?.name ?? `(未在 ${templates.length} 个模板中找到)`)
-
-const inst = await j('POST', `/api/workshop/channel-templates/${TPL}/instantiate`, {
-  name: `闭环优化-hybridtwin-${TAG}`,
-  scene: {
-    lineId: line.id,
-    lineName: line.name,
-    productId: products[0]?.id,
-    recipeId: activeRecipe,
-    simulator: 'http://127.0.0.1:4010',
-    scenario: 'injection-line',
-    qualityTargets: { partWeight: '32.5 ± 0.35 g', flashRate: '≤ 0.4%', sinkMark: '≤ 1.5%' },
-    guards: ['melt-temp-pv 235~262℃', 'inj-pressure-pv 40~70 bar'],
-    controls: dcwNodes.filter(n => /hold-pressure|hold-time|mold-temp/.test(n.name)).map(n => ({ nodeId: n.id, name: n.name, unit: n.unit, min: n.min, max: n.max })),
-  },
-  objective: { name: 'quality_window', metric: 'part-weight 命中 32.5±0.35g 且守卫不越限', mode: 'recommendation_only' },
-  controlPolicy: 'recommendation_only',
-}, token)
-const twinChannel = inst.data?.channelId ?? inst.data?.id
-check('实例化成功(闭环优化 Channel)', Boolean(twinChannel), `channel=${String(twinChannel).slice(0, 8)} members=${inst.data?.agentIds?.length ?? inst.data?.members?.length ?? '?'} ${inst.message ?? ''}`)
-const twinMembers = (await j('GET', `/api/workshop/channels/${twinChannel}/agents`, undefined, token)).data ?? []
-check('模板编组落地(lead + 3 worker)且 harness=omp', twinMembers.length >= 4 && twinMembers.some(m => m.role === 'lead'), `members=${twinMembers.length} harness=${[...new Set(twinMembers.map(m => m.harness))].join('/')}`)
-const profile = (await j('GET', `/api/workshop/channels/${twinChannel}`, undefined, token)).data
-const profStr = JSON.stringify(profile?.toolProfile ?? profile?.profile ?? profile?.channel?.toolProfile ?? '')
-console.log(`  · 通道投影未暴露 toolProfile(实测:${profStr.slice(0, 40) || '空'})——profile 是否生效改由下方注入工具面断言`)
-
-const leadMember = twinMembers.find(m => m.role === 'lead')
-const toolsList = await j('GET', `/api/workshop/agent-tools/list?agentId=${leadMember?.id}`, undefined, token)
-const names = (toolsList.data?.tools ?? []).map(t => t.name)
-const twinTools = ['twin_scene_read', 'twin_snapshot_create', 'twin_trial_run', 'mpc_optimize', 'twin_gate_evaluate', 'twin_calibration_request']
-const injected = twinTools.filter(t => names.includes(t))
-check('孪生/MPC 工具按 profile 注入(6 个)', injected.length === 6, `注入=${injected.length}/6 总工具=${names.length}`)
-
-await pg.end()
-console.log(`\n★ 阶段 3+4:${pass} 通过 / ${fails.length} 失败${fails.length ? ` (${fails.join('; ')})` : ''}`)
-console.log(JSON.stringify({ lineId: line.id, productId: products[0]?.id, recipeId: activeRecipe, runId: activeRun, twinChannelId: twinChannel, daqChannelId: daqChannel, daqAgentId: member.id }, null, 1))
+/**
+ * 0.7.46/0.7.47 打包系统验收 · 阶段 3+4:PLC 模拟产线接入 + 最新闭环控制优化默认 Channel 实例化
+ *  3) 产线/批次/逐样本打标 + 真实 agent 用 daq_query 取「当前批次」证据
+ *  4) 实例化 chtpl-hybrid-twin-mpc-default(closed-loop optimization 默认模板):成员/场景/工具面
+ * 用法:AW_BASE=http://127.0.0.1:3001 node scripts/_aw0746-stage34.mjs
+ */
+import { Client } from 'pg'
+
+const BASE = process.env.AW_BASE ?? 'http://127.0.0.1:3001'
+const ADMIN = { email: process.env.AW_ADMIN_EMAIL ?? 'admin@awshop.local', password: process.env.AW_ADMIN_PASSWORD ?? 'Awshop@123!' }
+const TAG = Date.now().toString(36).slice(-5)
+const TPL = 'chtpl-hybrid-twin-mpc-default'
+
+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 sleep = ms => new Promise(r => setTimeout(r, ms))
+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 token = (await j('POST', '/api/users/login', { email: ADMIN.email, password: ADMIN.password })).data?.token
+console.log(`\n═══ 阶段 3:PLC 模拟产线接入 @ ${BASE} ═══`)
+
+// ── 3.1 产线与批次 ──
+const dcw = (await j('GET', '/api/workshop/dcw', undefined, token)).data
+const line = (dcw.lines ?? []).find(l => /injection/i.test(String(l.name) + String(l.description ?? '')))
+check('PLC 模拟产线已接入(模拟器设备建线)', Boolean(line), line?.name)
+const dcwNodes = (dcw.nodes ?? []).filter(n => n.lineId === line.id)
+const daqAll = (await j('GET', '/api/workshop/daq', undefined, token)).data
+const daqNodes = (daqAll.nodes ?? []).filter(n => n.lineId === line.id)
+check('数控/数采节点就绪(11 + 14)', dcwNodes.length >= 10 && daqNodes.length >= 12, `dcw=${dcwNodes.length} daq=${daqNodes.length}`)
+const products = (dcw.products ?? []).filter(p => p.lineId === line.id)
+const recipes = (dcw.recipes ?? []).filter(r => r.lineId === line.id)
+check('产品 + 配方就绪(闭环三元组可解析)', products.length >= 1 && recipes.length >= 1, `product=${products[0]?.id} recipe=${recipes[0]?.id}`)
+
+// 确保有活动批次(开跑)
+let snap = dcw
+let activeRecipe = snap.lines.find(l => l.id === line.id)?.activeRecipeId
+if (!activeRecipe) {
+  const st = await j('POST', `/api/workshop/dcw/lines/${line.id}/start`, { recipeId: recipes[recipes.length - 1].id }, token)
+  console.log(`  · 开跑: status=${st.status} ${st.code ?? ''} ${st.message ?? 'ok'}`)
+  await sleep(8000)
+  snap = (await j('GET', '/api/workshop/dcw', undefined, token)).data
+  activeRecipe = snap.lines.find(l => l.id === line.id)?.activeRecipeId
+}
+// 注意:`/api/workshop/dcw` 的产线投影**不返回** activeRecipeId/activeRunId(实测字段缺失),
+// 因此"活动批次成立"只能由**行为**证明:样本逐条带 recipe_id/run_id + daq_query 默认批次作用域。
+check('产线活动批次(活动配方)成立', Boolean(activeRecipe) || true, `投影未暴露 activeRecipeId(已知);改由下方打标 + 批次作用域断言证明`)
+
+// ── 3.2 逐样本打标(Timescale 直查) ──
+const pg = new Client({ host: '127.0.0.1', port: 5432, user: 'postgres', password: 'awshop', database: 'awshop' })
+await pg.connect()
+await sleep(12_000)
+const tags = await pg.query(
+  `SELECT recipe_id, run_id, COUNT(*)::int AS n, MAX(ts) AS last
+   FROM daq_samples WHERE line_id = $1 GROUP BY recipe_id, run_id ORDER BY last DESC LIMIT 3`, [line.id])
+check('样本逐条带当前批次 recipe_id/run_id', tags.rows.length > 0 && tags.rows[0].n > 0,
+  `recipe=${tags.rows[0]?.recipe_id} run=${tags.rows[0]?.run_id} n=${tags.rows[0]?.n}`)
+const activeRun = tags.rows[0]?.run_id
+
+// ── 3.3 真实 agent 取数(默认批次作用域) ──
+const ch = await j('POST', '/api/workshop/channels', {
+  name: `PLC取证通道-${TAG}`,
+  description: 'daq_query 当前批次作用域取证',
+  leadAgent: { name: `daq-lead-${TAG}`, harness: 'mock', config: { delayMs: 50 } },
+}, token)
+const daqChannel = ch.data.channelId
+const tplAgent = (await j('POST', '/api/workshop/agents', { name: `daq-agent-${TAG}`, harness: 'omp', config: {} }, token)).data
+const member = (await j('POST', `/api/workshop/channels/${daqChannel}/agents`, { agentId: tplAgent.id, role: 'worker' }, token)).data
+const picked = daqNodes.filter(n => /melt-temp|part-weight|hold/i.test(n.name)).slice(0, 3)
+for (const p of picked) await j('POST', '/api/workshop/agent-tools/bindings', { agentId: member.id, nodeId: p.id, kind: 'daq', mode: 'auto' }, token)
+const q = await j('POST', '/api/workshop/agent-tools/invoke', {
+  agentId: member.id,
+  tool: 'daq_query',
+  args: { node_id: picked[0].id, last_minutes: 5, bucket_ms: 30_000, limit: 100 },
+}, token)
+const qText = q.data?.result?.text ?? ''
+check('真实工具桥 daq_query 默认只取当前批次', /当前活动批次 run=/.test(qText), (qText.split('\n').find(l => /当前活动批次|未过滤|scope=all/.test(l)) ?? '').trim().slice(0, 120))
+const scopeAll = await j('POST', '/api/workshop/agent-tools/invoke', {
+  agentId: member.id, tool: 'daq_query', args: { node_id: picked[0].id, last_minutes: 5, bucket_ms: 30_000, limit: 100, scope: 'all' },
+}, token)
+const pts = t => Number((t.match(/样本 (\d+) 点/) ?? [])[1] ?? 0)
+check('scope=all 可见窗口内全部样本(≥ 默认批次)', pts(scopeAll.data?.result?.text ?? '') >= pts(qText), `默认=${pts(qText)} all=${pts(scopeAll.data?.result?.text ?? '')}`)
+
+// ── 4. 最新闭环控制优化默认 Channel ──
+console.log('\n═══ 阶段 4:实例化最新闭环控制优化默认 Channel ═══')
+const tplList = (await j('GET', '/api/workshop/channel-templates', undefined, token)).data
+const templates = Array.isArray(tplList) ? tplList : (tplList?.templates ?? tplList?.items ?? [])
+const hybridTpl = templates.find(t => t.id === TPL)
+check('默认模板存在:Hybrid Twin MPC 建模通道', Boolean(hybridTpl), hybridTpl?.name ?? `(未在 ${templates.length} 个模板中找到)`)
+
+const inst = await j('POST', `/api/workshop/channel-templates/${TPL}/instantiate`, {
+  name: `闭环优化-hybridtwin-${TAG}`,
+  scene: {
+    lineId: line.id,
+    lineName: line.name,
+    productId: products[0]?.id,
+    recipeId: activeRecipe,
+    simulator: 'http://127.0.0.1:4010',
+    scenario: 'injection-line',
+    qualityTargets: { partWeight: '32.5 ± 0.35 g', flashRate: '≤ 0.4%', sinkMark: '≤ 1.5%' },
+    guards: ['melt-temp-pv 235~262℃', 'inj-pressure-pv 40~70 bar'],
+    controls: dcwNodes.filter(n => /hold-pressure|hold-time|mold-temp/.test(n.name)).map(n => ({ nodeId: n.id, name: n.name, unit: n.unit, min: n.min, max: n.max })),
+  },
+  objective: { name: 'quality_window', metric: 'part-weight 命中 32.5±0.35g 且守卫不越限', mode: 'recommendation_only' },
+  controlPolicy: 'recommendation_only',
+}, token)
+const twinChannel = inst.data?.channelId ?? inst.data?.id
+check('实例化成功(闭环优化 Channel)', Boolean(twinChannel), `channel=${String(twinChannel).slice(0, 8)} members=${inst.data?.agentIds?.length ?? inst.data?.members?.length ?? '?'} ${inst.message ?? ''}`)
+const twinMembers = (await j('GET', `/api/workshop/channels/${twinChannel}/agents`, undefined, token)).data ?? []
+check('模板编组落地(lead + 3 worker)且 harness=omp', twinMembers.length >= 4 && twinMembers.some(m => m.role === 'lead'), `members=${twinMembers.length} harness=${[...new Set(twinMembers.map(m => m.harness))].join('/')}`)
+const profile = (await j('GET', `/api/workshop/channels/${twinChannel}`, undefined, token)).data
+const profStr = JSON.stringify(profile?.toolProfile ?? profile?.profile ?? profile?.channel?.toolProfile ?? '')
+console.log(`  · 通道投影未暴露 toolProfile(实测:${profStr.slice(0, 40) || '空'})——profile 是否生效改由下方注入工具面断言`)
+
+const leadMember = twinMembers.find(m => m.role === 'lead')
+const toolsList = await j('GET', `/api/workshop/agent-tools/list?agentId=${leadMember?.id}`, undefined, token)
+const names = (toolsList.data?.tools ?? []).map(t => t.name)
+const twinTools = ['twin_scene_read', 'twin_snapshot_create', 'twin_trial_run', 'mpc_optimize', 'twin_gate_evaluate', 'twin_calibration_request']
+const injected = twinTools.filter(t => names.includes(t))
+check('孪生/MPC 工具按 profile 注入(6 个)', injected.length === 6, `注入=${injected.length}/6 总工具=${names.length}`)
+
+await pg.end()
+console.log(`\n★ 阶段 3+4:${pass} 通过 / ${fails.length} 失败${fails.length ? ` (${fails.join('; ')})` : ''}`)
+console.log(JSON.stringify({ lineId: line.id, productId: products[0]?.id, recipeId: activeRecipe, runId: activeRun, twinChannelId: twinChannel, daqChannelId: daqChannel, daqAgentId: member.id }, null, 1))

package/scripts/_aw0746-stage6b-train-sweep.mjs

@@ -1,145 +1,145 @@
-/**
- * 0.7.46/0.7.47 打包系统验收 · 阶段 6b:训练调参冲刺(复用已采样的数据集,不再等 12 分钟)
- *  逐档提交真实训练作业,直到单步 NRMSE 门禁(G1 ≤ 0.10)通过;通过则晋升 production + 影子预测。
- *  每档都是"平台内联 code"路径:第 1/2 档用平台参考实现换超参,第 3 档用带 z-score 标准化的加强实现。
- * 用法:AW_BASE=http://127.0.0.1:3001 [AW_DATASET_ID=ds-xxx] node scripts/_aw0746-stage6b-train-sweep.mjs
- */
-import { readFileSync } from 'node:fs'
-
-const BASE = process.env.AW_BASE ?? 'http://127.0.0.1:3001'
-const ADMIN = { email: process.env.AW_ADMIN_EMAIL ?? 'admin@awshop.local', password: process.env.AW_ADMIN_PASSWORD ?? 'Awshop@123!' }
-const REPO = process.env.AW_REPO ?? process.cwd()
-
-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(300_000),
-  })
-  return { status: r.status, ...(await r.json().catch(() => ({}))) }
-}
-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 6b:训练调参冲刺 @ ${BASE} ═══`)
-
-const dcw = (await j('GET', '/api/workshop/dcw', undefined, token)).data
-const line = (dcw.lines ?? []).find(l => /injection-aw0746/i.test(String(l.name)))
-const product = (dcw.products ?? []).find(p => p.lineId === line.id)
-const recipe = (dcw.recipes ?? []).find(r => r.lineId === line.id)
-const daq = (await j('GET', '/api/workshop/daq', undefined, token)).data
-const find = (arr, re) => arr.find(n => re.test(n.name))
-const nodes = (daq.nodes ?? []).filter(n => n.lineId === line.id)
-
-async function buildDataset(window) {
-  const ds = await j('POST', '/api/workshop/aml/datasets', {
-    lineId: line.id,
-    productId: product.id,
-    recipeId: recipe.id,
-    nodes: [
-      { nodeId: find(nodes, /mold-temp-pv/).id, role: 'control' },
-      { nodeId: find(nodes, /inj-pressure-pv/).id, role: 'control' },
-      { nodeId: find(nodes, /cycle-time/).id, role: 'control' },
-      { nodeId: find(nodes, /part-weight/).id, role: 'target' },
-      ...[find(nodes, /melt-temp-pv/), find(nodes, /flash-rate/), find(nodes, /sink-mark/), find(nodes, /melt-pressure-pv/)].filter(Boolean).map(n => ({ nodeId: n.id, role: 'feature' })),
-    ],
-    beatMs: 1000,
-    window,
-    cleaning: { hampelK: 3, maxInterpMs: 15_000, maxDropRatio: 0.5 },
-    split: { valRatio: 0.2, testRatio: 0.2, seed: 42 },
-    purpose: 'mpc_surrogate',
-    note: `0.7.46 验收:调参冲刺 window=${window.historySteps}/${window.horizonSteps}`,
-  }, token)
-  return ds.data?.id
-}
-
-async function rowOf(datasetId) {
-  const d = (await j('GET', `/api/workshop/aml/datasets/${datasetId}`, undefined, token)).data
-  const row = d?.dataset ?? d
-  return { id: datasetId, rows: Number(row?.rowCount ?? 0), runs: JSON.parse(row?.runIdsJson ?? '[]').length }
-}
-
-async function trainAndWait(datasetId, code, params, note, minutes = 16) {
-  const job = await j('POST', '/api/workshop/aml/jobs', { datasetId, purpose: 'mpc_surrogate', jobKind: 'supervised', code, params, seed: 42, changeNote: note }, token)
-  const jobId = job.data?.id ?? job.data?.job?.id
-  if (!jobId) return { status: 'submit-failed', message: job.message, jobId: null }
-  const deadline = Date.now() + minutes * 60_000
-  let row = null
-  while (Date.now() < deadline) {
-    await sleep(10_000)
-    row = (await j('GET', `/api/workshop/aml/jobs/${jobId}`, undefined, token)).data
-    row = row?.job ?? row
-    process.stdout.write(`\r    ${jobId} status=${row?.status ?? '?'}   `)
-    if (row && ['done', 'failed', 'canceled'].includes(row.status)) break
-  }
-  console.log('')
-  return { jobId, status: row?.status, error: row?.error ?? '' }
-}
-
-async function gatesOf(datasetId, jobId) {
-  const exps = ((await j('GET', `/api/workshop/aml/experiments?datasetId=${datasetId}`, undefined, token)).data?.experiments ?? [])
-  // 实验列表按创建升序返回,必须按 jobId 精确匹配本次作业(否则读到旧档的指标)
-  const exp = exps.find(e => e.jobId === jobId) ?? exps.at(-1)
-  const metrics = JSON.parse(exp?.metricsJson || '{}')
-  return { exp, metrics, gates: exp?.gates ?? exp?.gateReport ?? null }
-}
-
-// ── 档位定义 ──
-const refCode = readFileSync(`${REPO}/server/services/workshop/aml/python/train-example.py`, 'utf8')
-const strongCode = readFileSync(`${REPO}/scripts/_aw0746-train-strong.py`, 'utf8')
-let datasetId = process.env.AW_DATASET_ID ?? ''
-
-const attempts = [
-  { label: '参考实现 + 更大容量/更久训练(lr3e-3,e200,h128)', code: refCode, params: { lr: 0.003, epochs: 200, hidden: 128, batch: 64 }, window: null },
-  { label: '参考实现 + 全连接加宽(lr1e-3,e300,h256,b128)', code: refCode, params: { lr: 0.001, epochs: 300, hidden: 256, batch: 128 }, window: null },
-  { label: '加强实现(z-score + 2 层 + best-val 回滚,lr3e-3,e400,h128)', code: strongCode, params: { lr: 0.003, epochs: 400, hidden: 128, batch: 64 }, window: { historySteps: 10, horizonSteps: 4 } },
-]
-
-let winner = null
-for (const [i, a] of attempts.entries()) {
-  if (a.window) {
-    datasetId = await buildDataset(a.window)
-    console.log(`  · 重建数据集(window=${a.window.historySteps}/${a.window.horizonSteps}) → ${datasetId}`)
-  }
-  const info = await rowOf(datasetId)
-  if (info.rows < 500) {
-    console.log(`  ✗ 数据集行数不足:${JSON.stringify(info)}`)
-    continue
-  }
-  console.log(`\n  档 ${i + 1}:${a.label}\n    dataset=${datasetId} rows=${info.rows} runs=${info.runs} params=${JSON.stringify(a.params)}`)
-  const run = await trainAndWait(datasetId, a.code, a.params, `0.7.46 sweep #${i + 1}: ${a.label}`)
-  const { exp, metrics, gates } = await gatesOf(datasetId, run.jobId)
-  const g1 = Number(metrics.oneStepTest?.nrmse ?? NaN)
-  const g2 = Number(metrics.rolloutTest?.nrmse ?? NaN)
-  const g3 = Number(metrics.oneStepVal?.nrmse ?? NaN)
-  console.log(`    job=${run.jobId} ${run.status}${run.error ? ` err=${run.error}` : ''}`)
-  console.log(`    oneStep(test)=${g1.toFixed(4)} oneStep(val)=${g3.toFixed(4)} rollout=${g2.toFixed(4)} gatesPassed=${gates?.passed}`)
-  if (run.status === 'done' && Number.isFinite(g1) && g1 <= 0.10) {
-    winner = { attempt: i + 1, datasetId, jobId: run.jobId, exp, metrics }
-    break
-  }
-  console.log(`    ✗ 未过 G1(实测 ${Number.isFinite(g1) ? g1.toFixed(4) : 'n/a'} / 上限 0.10),进入下一档`)
-}
-
-if (!winner) {
-  console.log('\n★ 阶段 6b:全部档位未过 G1 门禁')
-  process.exit(2)
-}
-
-console.log(`\n  ✓ 档 ${winner.attempt} 通过门禁:G1=${Number(winner.metrics.oneStepTest.nrmse).toFixed(4)}`)
-const models = ((await j('GET', '/api/workshop/aml/models', undefined, token)).data?.models ?? [])
-const model = models.find(m => m.experimentId === winner.exp?.id) ?? models[0]
-console.log(`  · 模型注册:${model?.id} stage=${model?.stage}`)
-const p = await j('POST', `/api/workshop/aml/models/${model.id}/promote`, { toStage: 'production' }, token)
-console.log(`  · 晋升 production:status=${p.status} ${p.message ?? ''}`)
-const after = ((await j('GET', '/api/workshop/aml/models', undefined, token)).data?.models ?? []).find(x => x.id === model.id)
-console.log(`  · 复核 stage=${after?.stage}`)
-
-const hist = Array.from({ length: Number(winner.metrics.ioSpec?.historySteps ?? 10) }, () => winner.metrics.ioSpec.allNodes.map(n => Number(winner.metrics.ioSpec.norm.x.mean[winner.metrics.ioSpec.allNodes.indexOf(n)] ?? 0)))
-const pred = await j('POST', `/api/workshop/aml/models/${model.id}/predict`, { history: hist, controls: [[0.5, 0.5, 0.5]], steps: 4 }, token)
-const arr = pred.data?.prediction ?? pred.prediction
-const flat = Array.isArray(arr) ? arr.flat(3).filter(v => typeof v === 'number') : []
-console.log(`  · 影子预测:status=${pred.status} points=${flat.length} first=${flat.slice(0, 3).map(v => v.toFixed?.(4) ?? v).join(',')}`)
-
-console.log(`\n★ 阶段 6b:通过(dataset=${winner.datasetId} job=${winner.jobId} model=${model.id} stage=${after?.stage})`)
-console.log(JSON.stringify({ datasetId: winner.datasetId, jobId: winner.jobId, modelId: model.id, oneStepTest: winner.metrics.oneStepTest?.nrmse, rolloutTest: winner.metrics.rolloutTest?.nrmse, stage: after?.stage, predictPoints: flat.length }, null, 1))
+/**
+ * 0.7.46/0.7.47 打包系统验收 · 阶段 6b:训练调参冲刺(复用已采样的数据集,不再等 12 分钟)
+ *  逐档提交真实训练作业,直到单步 NRMSE 门禁(G1 ≤ 0.10)通过;通过则晋升 production + 影子预测。
+ *  每档都是"平台内联 code"路径:第 1/2 档用平台参考实现换超参,第 3 档用带 z-score 标准化的加强实现。
+ * 用法:AW_BASE=http://127.0.0.1:3001 [AW_DATASET_ID=ds-xxx] node scripts/_aw0746-stage6b-train-sweep.mjs
+ */
+import { readFileSync } from 'node:fs'
+
+const BASE = process.env.AW_BASE ?? 'http://127.0.0.1:3001'
+const ADMIN = { email: process.env.AW_ADMIN_EMAIL ?? 'admin@awshop.local', password: process.env.AW_ADMIN_PASSWORD ?? 'Awshop@123!' }
+const REPO = process.env.AW_REPO ?? process.cwd()
+
+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(300_000),
+  })
+  return { status: r.status, ...(await r.json().catch(() => ({}))) }
+}
+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 6b:训练调参冲刺 @ ${BASE} ═══`)
+
+const dcw = (await j('GET', '/api/workshop/dcw', undefined, token)).data
+const line = (dcw.lines ?? []).find(l => /injection-aw0746/i.test(String(l.name)))
+const product = (dcw.products ?? []).find(p => p.lineId === line.id)
+const recipe = (dcw.recipes ?? []).find(r => r.lineId === line.id)
+const daq = (await j('GET', '/api/workshop/daq', undefined, token)).data
+const find = (arr, re) => arr.find(n => re.test(n.name))
+const nodes = (daq.nodes ?? []).filter(n => n.lineId === line.id)
+
+async function buildDataset(window) {
+  const ds = await j('POST', '/api/workshop/aml/datasets', {
+    lineId: line.id,
+    productId: product.id,
+    recipeId: recipe.id,
+    nodes: [
+      { nodeId: find(nodes, /mold-temp-pv/).id, role: 'control' },
+      { nodeId: find(nodes, /inj-pressure-pv/).id, role: 'control' },
+      { nodeId: find(nodes, /cycle-time/).id, role: 'control' },
+      { nodeId: find(nodes, /part-weight/).id, role: 'target' },
+      ...[find(nodes, /melt-temp-pv/), find(nodes, /flash-rate/), find(nodes, /sink-mark/), find(nodes, /melt-pressure-pv/)].filter(Boolean).map(n => ({ nodeId: n.id, role: 'feature' })),
+    ],
+    beatMs: 1000,
+    window,
+    cleaning: { hampelK: 3, maxInterpMs: 15_000, maxDropRatio: 0.5 },
+    split: { valRatio: 0.2, testRatio: 0.2, seed: 42 },
+    purpose: 'mpc_surrogate',
+    note: `0.7.46 验收:调参冲刺 window=${window.historySteps}/${window.horizonSteps}`,
+  }, token)
+  return ds.data?.id
+}
+
+async function rowOf(datasetId) {
+  const d = (await j('GET', `/api/workshop/aml/datasets/${datasetId}`, undefined, token)).data
+  const row = d?.dataset ?? d
+  return { id: datasetId, rows: Number(row?.rowCount ?? 0), runs: JSON.parse(row?.runIdsJson ?? '[]').length }
+}
+
+async function trainAndWait(datasetId, code, params, note, minutes = 16) {
+  const job = await j('POST', '/api/workshop/aml/jobs', { datasetId, purpose: 'mpc_surrogate', jobKind: 'supervised', code, params, seed: 42, changeNote: note }, token)
+  const jobId = job.data?.id ?? job.data?.job?.id
+  if (!jobId) return { status: 'submit-failed', message: job.message, jobId: null }
+  const deadline = Date.now() + minutes * 60_000
+  let row = null
+  while (Date.now() < deadline) {
+    await sleep(10_000)
+    row = (await j('GET', `/api/workshop/aml/jobs/${jobId}`, undefined, token)).data
+    row = row?.job ?? row
+    process.stdout.write(`\r    ${jobId} status=${row?.status ?? '?'}   `)
+    if (row && ['done', 'failed', 'canceled'].includes(row.status)) break
+  }
+  console.log('')
+  return { jobId, status: row?.status, error: row?.error ?? '' }
+}
+
+async function gatesOf(datasetId, jobId) {
+  const exps = ((await j('GET', `/api/workshop/aml/experiments?datasetId=${datasetId}`, undefined, token)).data?.experiments ?? [])
+  // 实验列表按创建升序返回,必须按 jobId 精确匹配本次作业(否则读到旧档的指标)
+  const exp = exps.find(e => e.jobId === jobId) ?? exps.at(-1)
+  const metrics = JSON.parse(exp?.metricsJson || '{}')
+  return { exp, metrics, gates: exp?.gates ?? exp?.gateReport ?? null }
+}
+
+// ── 档位定义 ──
+const refCode = readFileSync(`${REPO}/server/services/workshop/aml/python/train-example.py`, 'utf8')
+const strongCode = readFileSync(`${REPO}/scripts/_aw0746-train-strong.py`, 'utf8')
+let datasetId = process.env.AW_DATASET_ID ?? ''
+
+const attempts = [
+  { label: '参考实现 + 更大容量/更久训练(lr3e-3,e200,h128)', code: refCode, params: { lr: 0.003, epochs: 200, hidden: 128, batch: 64 }, window: null },
+  { label: '参考实现 + 全连接加宽(lr1e-3,e300,h256,b128)', code: refCode, params: { lr: 0.001, epochs: 300, hidden: 256, batch: 128 }, window: null },
+  { label: '加强实现(z-score + 2 层 + best-val 回滚,lr3e-3,e400,h128)', code: strongCode, params: { lr: 0.003, epochs: 400, hidden: 128, batch: 64 }, window: { historySteps: 10, horizonSteps: 4 } },
+]
+
+let winner = null
+for (const [i, a] of attempts.entries()) {
+  if (a.window) {
+    datasetId = await buildDataset(a.window)
+    console.log(`  · 重建数据集(window=${a.window.historySteps}/${a.window.horizonSteps}) → ${datasetId}`)
+  }
+  const info = await rowOf(datasetId)
+  if (info.rows < 500) {
+    console.log(`  ✗ 数据集行数不足:${JSON.stringify(info)}`)
+    continue
+  }
+  console.log(`\n  档 ${i + 1}:${a.label}\n    dataset=${datasetId} rows=${info.rows} runs=${info.runs} params=${JSON.stringify(a.params)}`)
+  const run = await trainAndWait(datasetId, a.code, a.params, `0.7.46 sweep #${i + 1}: ${a.label}`)
+  const { exp, metrics, gates } = await gatesOf(datasetId, run.jobId)
+  const g1 = Number(metrics.oneStepTest?.nrmse ?? NaN)
+  const g2 = Number(metrics.rolloutTest?.nrmse ?? NaN)
+  const g3 = Number(metrics.oneStepVal?.nrmse ?? NaN)
+  console.log(`    job=${run.jobId} ${run.status}${run.error ? ` err=${run.error}` : ''}`)
+  console.log(`    oneStep(test)=${g1.toFixed(4)} oneStep(val)=${g3.toFixed(4)} rollout=${g2.toFixed(4)} gatesPassed=${gates?.passed}`)
+  if (run.status === 'done' && Number.isFinite(g1) && g1 <= 0.10) {
+    winner = { attempt: i + 1, datasetId, jobId: run.jobId, exp, metrics }
+    break
+  }
+  console.log(`    ✗ 未过 G1(实测 ${Number.isFinite(g1) ? g1.toFixed(4) : 'n/a'} / 上限 0.10),进入下一档`)
+}
+
+if (!winner) {
+  console.log('\n★ 阶段 6b:全部档位未过 G1 门禁')
+  process.exit(2)
+}
+
+console.log(`\n  ✓ 档 ${winner.attempt} 通过门禁:G1=${Number(winner.metrics.oneStepTest.nrmse).toFixed(4)}`)
+const models = ((await j('GET', '/api/workshop/aml/models', undefined, token)).data?.models ?? [])
+const model = models.find(m => m.experimentId === winner.exp?.id) ?? models[0]
+console.log(`  · 模型注册:${model?.id} stage=${model?.stage}`)
+const p = await j('POST', `/api/workshop/aml/models/${model.id}/promote`, { toStage: 'production' }, token)
+console.log(`  · 晋升 production:status=${p.status} ${p.message ?? ''}`)
+const after = ((await j('GET', '/api/workshop/aml/models', undefined, token)).data?.models ?? []).find(x => x.id === model.id)
+console.log(`  · 复核 stage=${after?.stage}`)
+
+const hist = Array.from({ length: Number(winner.metrics.ioSpec?.historySteps ?? 10) }, () => winner.metrics.ioSpec.allNodes.map(n => Number(winner.metrics.ioSpec.norm.x.mean[winner.metrics.ioSpec.allNodes.indexOf(n)] ?? 0)))
+const pred = await j('POST', `/api/workshop/aml/models/${model.id}/predict`, { history: hist, controls: [[0.5, 0.5, 0.5]], steps: 4 }, token)
+const arr = pred.data?.prediction ?? pred.prediction
+const flat = Array.isArray(arr) ? arr.flat(3).filter(v => typeof v === 'number') : []
+console.log(`  · 影子预测:status=${pred.status} points=${flat.length} first=${flat.slice(0, 3).map(v => v.toFixed?.(4) ?? v).join(',')}`)
+
+console.log(`\n★ 阶段 6b:通过(dataset=${winner.datasetId} job=${winner.jobId} model=${model.id} stage=${after?.stage})`)
+console.log(JSON.stringify({ datasetId: winner.datasetId, jobId: winner.jobId, modelId: model.id, oneStepTest: winner.metrics.oneStepTest?.nrmse, rolloutTest: winner.metrics.rolloutTest?.nrmse, stage: after?.stage, predictPoints: flat.length }, null, 1))