agentworkshop 0.7.46 → 0.7.47

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/_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}`)

package/scripts/_probe-density.cjs

@@ -0,0 +1,15 @@
+/** 临时:探测 daq_samples 采样密度(评估 AML beatMs 可下调空间) */
+const { Client } = require('pg')
+;(async () => {
+  const c = new Client({ host: '127.0.0.1', port: 5432, user: 'postgres', password: 'awshop', database: 'awshop' })
+  await c.connect()
+  const q = (sql) => c.query(sql).then(r => r.rows)
+  const rows = await q(`select node_id, count(*)::int n, min(ts) a, max(ts) b,
+      round(extract(epoch from (max(ts)-min(ts)))/greatest(count(*)-1,1),2) avg_gap_s
+    from daq_samples where ts > now() - interval '45 minutes'
+    group by node_id order by n desc limit 12`)
+  console.table(rows)
+  const t = await q(`select count(*)::int total, count(distinct node_id)::int nodes from daq_samples where ts > now() - interval '45 minutes'`)
+  console.log('total', t)
+  await c.end()
+})().catch(e => { console.error(e.message); process.exit(1) })

package/scripts/_probe-gate-model.mjs

@@ -0,0 +1,34 @@
+/** 临时:看 twin_gate_evaluate 的 modelUpdate 尾部 + 最近作业来源 */
+import { DatabaseSync } from 'node:sqlite'
+const BASE = 'http://127.0.0.1:3001'
+const ADMIN = { email: 'admin@awshop.local', password: 'Awshop@123!' }
+
+const j = async (method, path, body, token) => {
+  const r = await fetch(BASE + path, { method, headers: { 'content-type': 'application/json', ...(token ? { authorization: `Bearer ${token}` } : {}) }, body: body === undefined ? undefined : JSON.stringify(body) })
+  return { status: r.status, ...(await r.json().catch(() => ({}))) }
+}
+const token = (await j('POST', '/api/users/login', ADMIN)).data?.token
+const chans = (await j('GET', '/api/workshop/channels', undefined, token)).data ?? []
+const twin = chans.find(c => /闭环优化-hybridtwin/.test(c.name))
+const members = (await j('GET', `/api/workshop/channels/${twin.id}/agents`, undefined, token)).data ?? []
+const lead = members.find(m => m.role === 'lead')
+const models = ((await j('GET', '/api/workshop/aml/models', undefined, token)).data?.models ?? [])
+const prod = models.find(m => m.stage === 'production')
+console.log('production model =', prod?.id, 'ds =', prod?.datasetId)
+
+const res = await j('POST', '/api/workshop/agent-tools/invoke', {
+  agentId: lead.id,
+  tool: 'twin_gate_evaluate',
+  args: { rows: 1724, runs: 10, one_step_nrmse: 0.0763, rollout_nrmse: 0.0964, val_test_gap: 0.1072, calibration_rows: 1724, coverage: 1, candidate_trials: [], physics_failure_rate: 0, model_id: prod?.id ?? '' },
+}, token)
+const text = String(res.data?.result?.text ?? '')
+const tail = text.slice(-700)
+console.log('--- gate tail ---')
+console.log(tail)
+console.log('--- parsed keys ---', Object.keys(JSON.parse(text)).join(','))
+
+const db = new DatabaseSync('D:/codes/ABO/AgentWorkShop/.aw0746-home/data/workshop.sqlite', { readOnly: true })
+const cols = db.prepare('pragma table_info(aml_jobs)').all().map(c => c.name)
+console.log('\naml_jobs columns:', cols.join(','))
+const pick = ['id', 'status', 'kind', 'created_by', 'change_note', 'created_at'].filter(c => cols.includes(c))
+for (const r of db.prepare(`select ${pick.join(',')} from aml_jobs order by created_at desc limit 6`).all()) console.log('  ', JSON.stringify(r).slice(0, 220))

package/scripts/_probe-model-eligibility.cjs

@@ -0,0 +1,13 @@
+/** 临时:检查 AML 模型 metrics_json 里的 twinEligibility 是否回写 */
+const { DatabaseSync } = require('node:sqlite')
+const db = new DatabaseSync('D:/codes/ABO/AgentWorkShop/.aw0746-home/data/workshop.sqlite', { readOnly: true })
+const rows = db.prepare('select id, stage, dataset_id, experiment_id, metrics_json from aml_models order by created_at desc limit 4').all()
+for (const r of rows) {
+  let m = {}
+  try { m = JSON.parse(r.metrics_json || '{}') } catch { m = { parse: 'fail' } }
+  console.log(`${r.id} stage=${r.stage} ds=${r.dataset_id} exp=${r.experiment_id}`)
+  console.log(`   twinEligibility=${JSON.stringify(m.twinEligibility ?? null)} oneStep=${m.oneStepTest?.nrmse}`)
+}
+const jobs = db.prepare('select id, status, created_by_kind, created_by, substr(change_note,1,60) note, created_at from aml_jobs order by created_at desc limit 6').all()
+console.log('\njobs:')
+for (const j of jobs) console.log(`  ${j.id} ${j.status} by=${j.created_by_kind}/${j.created_by} ${j.created_at} ${j.note}`)

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)
+      }
+    }
+  }
+}