ci: 负载下的启动耗时证据 (#81/#144/#40)

.github/workflows/conpty-compat.yml

@@ -297,3 +297,64 @@ jobs:
           retention-days: 14
           if-no-files-found: warn
           path: startup-latency.json
+
+  windows-2025-startup-under-load:
+    # Bounded follow-up for #81 / #144 / #40: does local CPU contention on the
+    # runner inflate PowerShell start-up enough to reach the fixture deadlines.
+    runs-on: windows-2025
+    timeout-minutes: 20
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+      - uses: actions/setup-python@6d4fd22b20465ccc0f7354476588e8df7c9ca0e0 # v7.1.0
+        with:
+          python-version: '3.12'
+      - name: Measure start-up latency under CPU load
+        shell: pwsh
+        run: |
+          @'
+          import json, multiprocessing, platform, subprocess, time
+          def burn(deadline):
+              while time.monotonic() < deadline:
+                  sum(i * i for i in range(20000))
+          def samples(command, count, timeout):
+              rows = []
+              for _ in range(count):
+                  started = time.monotonic()
+                  try:
+                      result = subprocess.run(command, capture_output=True, timeout=timeout)
+                      rows.append({"elapsed": round(time.monotonic() - started, 4), "returncode": result.returncode, "timeout": False})
+                  except subprocess.TimeoutExpired:
+                      rows.append({"elapsed": round(time.monotonic() - started, 4), "returncode": None, "timeout": True})
+              return rows
+          def summary(rows):
+              times = sorted(row["elapsed"] for row in rows)
+              def at(fraction):
+                  return times[min(len(times) - 1, int(len(times) * fraction))]
+              return {"count": len(rows), "min": times[0], "p50": at(0.5), "p90": at(0.9), "max": times[-1],
+                      "over_1s": sum(1 for t in times if t >= 1.0),
+                      "over_5s": sum(1 for t in times if t >= 5.0),
+                      "timeouts": sum(1 for row in rows if row["timeout"])}
+          command = ["powershell.exe", "-NoLogo", "-NoProfile", "-NonInteractive", "-Command", "exit 0"]
+          deadline = time.monotonic() + 120
+          workers = [multiprocessing.Process(target=burn, args=(deadline,)) for _ in range(8)]
+          for worker in workers:
+              worker.start()
+          time.sleep(2)
+          loaded = samples(command, 20, 15)
+          for worker in workers:
+              worker.terminate()
+          report = {"schema": "termforge-startup-under-load-v1", "platform": platform.platform(),
+                    "workers": 8, "loaded": summary(loaded), "samples": loaded}
+          print(json.dumps({k: v for k, v in report.items() if k != "samples"}))
+          with open("startup-under-load.json", "w", encoding="utf-8") as stream:
+              json.dump(report, stream, indent=2)
+          '@ | Out-File -Encoding utf8 startup-under-load.py
+          python startup-under-load.py
+      - name: Preserve latency evidence
+        if: always()
+        uses: actions/upload-artifact@cf430e030ddbb5b0abf93d22962f4752f3646cd9 # v7.0.2
+        with:
+          name: startup-under-load-r2
+          retention-days: 14
+          if-no-files-found: warn
+          path: startup-under-load.json

research/ui/results/2026-10-30-hosted-startup-latency/README.md

@@ -23,9 +23,19 @@ dispatch-only 作业 `windows-2025-startup-latency`（[工作流](../../../../.g
 - 因此 `ps51` 在 CI 里出现 **10219 ms** 的 `timed_out`（#81）不能归因于启动成本；Node 夹具的 10 秒超时（#144）同理；
 - 三处（ps51 10.2 s、纯 PowerShell 12 s、Node 10 s）都属于"纯启动/纯脚本、几乎无输出、刚好越过期限"，与本次测得的正常分布相差 1–2 个数量级，指向**宿主级停顿**（安全扫描、VM 调度、存储阻塞）。
 
+## 附加：本地 CPU 负载下的启动耗时
+
+为区分「启动成本」与「宿主停顿」，另跑一次 `windows-2025-startup-under-load`（8 个 CPU 忙循环同时运行，测 20 次 PowerShell 5.1 启动）：
+
+| 项 | 次数 | min | p50 | p90 | max | ≥1s | ≥5s | 超时 |
+|---|---|---|---|---|---|---|---|---|
+| PowerShell 5.1（8 路 CPU 负载） | 20 | 0.156 s | 0.172 s | 0.172 s | 1.921 s | 1 | 0 | 0 |
+
+对照空闲时的 p50 0.110 s：负载只把中位数抬到 0.172 s，最慢 1.921 s，**仍与 10 秒期限相差一个数量级以上**。
+
 ## 边界
 
 - 只测空闲镜像；未在完整 CTest 同时段、安全软件扫描中或并行负载下测量；
 - 未修改任何期限、夹具或准备流程；
 - 不能据此断定停顿的具体来源，只排除"启动成本"这一解释；
-- 与 #144 的 Node 夹具计时（warm 63 ms、冷启动 1.47 s）互相印证。
+- 与 #144 的 Node 夹具计时（warm 63 ms、冷启动 1.47 s）互相印证。

research/ui/results/2026-10-30-hosted-startup-latency/validation.json

@@ -50,11 +50,27 @@
       "timeouts": 0
     }
   },
-  "interpretation": "Pure start-up cost is 1-2 orders of magnitude below the fixed fixture deadlines, so the observed 10-12 s timeouts cannot be explained by start-up cost.",
+  "interpretation": "Pure start-up cost is 1-2 orders of magnitude below the fixed fixture deadlines, so the observed 10-12 s timeouts cannot be explained by start-up cost. Local CPU contention raises the median only to 0.172 s, so it does not explain the 10 s stalls either.",
   "limitations": [
     "idle image only; not under full CTest concurrency or AV scanning",
     "no deadline or fixture change",
-    "does not identify the stall source, only excludes start-up cost"
+    "does not identify the stall source, only excludes start-up cost",
+    "loaded run uses eight busy loops, not the full CTest or an AV scan"
   ],
-  "performance_claim": false
+  "performance_claim": false,
+  "loaded_run": {
+    "run": 40945977134,
+    "job": "windows-2025-startup-under-load",
+    "workers": 8,
+    "powershell": {
+      "count": 20,
+      "min": 0.156,
+      "p50": 0.172,
+      "p90": 0.172,
+      "max": 1.921,
+      "over_1s": 1,
+      "over_5s": 0,
+      "timeouts": 0
+    }
+  }
 }