test(agent): observe actual CPU inference thread configuration

Author: chaxus

docs/evaluations/2026-10-07-cpu-thread-probe-bindings.json

@@ -0,0 +1,9 @@
+{
+  "sha256": {
+    "2026-10-07-cpu-thread-probe.json": "54a4d78c08de1425c3d4779be1aabf66c443791584431bdfd58a613eda9ce75c",
+    "2026-10-07-cpu-thread-probe.mjs": "6552db335f8d675a54373fef7037b92b934d32b6116f7029f03131c1759c433d",
+    "2026-10-07-cpu-thread-probe.md": "9807d51b5fde6be7a9ec8def6c3b6252b62b1bce3d4f5b0f29dd64346d700302"
+  },
+  "scope": "Current load observation only; no inference or prior-run certification",
+  "build": "core1791345713/vendorb6864850e7b3"
+}

docs/evaluations/2026-10-07-cpu-thread-probe.json

@@ -0,0 +1,25 @@
+{
+  "scope": "Gemma native CPU load thread observation only; no inference; cannot retroactively certify prior runs",
+  "errors": [],
+  "browserVersion": "151.0.7922.34",
+  "environment": {
+    "crossOriginIsolated": true,
+    "hardwareConcurrency": 10,
+    "sharedArrayBuffer": "function"
+  },
+  "loads": [
+    {
+      "requested": {
+        "n_threads": 4,
+        "n_ctx": 2048,
+        "n_gpu_layers": 0,
+        "reasoning": false
+      },
+      "actualThreads": 4,
+      "multithread": true,
+      "loaded": true
+    }
+  ],
+  "note": "模型已加载，可以开始对话。",
+  "closed": true
+}

docs/evaluations/2026-10-07-cpu-thread-probe.md

@@ -0,0 +1,5 @@
+# Current CPU thread observation
+
+A fresh Chromium context loaded the local Gemma candidate through the real application CPU provider. The SDK load hook recorded requested n_threads=4, n_ctx=2048, n_gpu_layers=0 and reasoning=false; after loading, getNumThreads() returned 4 and isMultithread() returned true. The page reported crossOriginIsolated=true, hardwareConcurrency=10 and SharedArrayBuffer available. No inference was requested; browser closed without page errors.
+
+This rules out a missing multithread configuration for this observed load only. Prior sampling runs did not capture actual threads, so this probe cannot certify their configuration retrospectively or explain their latency. It does not test performance, output quality or physical device coverage. No production setting is changed. Further performance comparisons must capture threads and native timing in the same run and control concurrent host activity.

docs/evaluations/2026-10-07-cpu-thread-probe.mjs

@@ -0,0 +1,19 @@
+import {chromium} from '@playwright/test';
+import fs from 'node:fs/promises';
+const r={scope:'Gemma native CPU load thread observation only; no inference; cannot retroactively certify prior runs',errors:[]};
+const b=await chromium.launch({channel:'chromium'});r.browserVersion=b.version();const c=await b.newContext({serviceWorkers:'block'});
+await c.addInitScript(()=>{if(window===top){localStorage.setItem('agent-panel-provider','wllama');localStorage.removeItem('agent-panel-gguf-url');}});
+try{
+const p=await c.newPage();p.on('pageerror',e=>r.errors.push(e.message));
+await p.goto('http://127.0.0.1:5193/editor?new=docx&agent=1&locale=zh-CN');
+await p.waitForFunction(()=>{const e=document.querySelector('#app iframe')?.contentWindow.editor;return e?.isDocumentLoadComplete&&e?.isLoadFullApi;},null,{timeout:90000});
+r.environment=await p.evaluate(async()=>{
+const {Wllama}=await import('/assets/client-D5_UYdz1.js');const original=Wllama.prototype.loadModel;
+window.__loads=[];
+Wllama.prototype.loadModel=async function(files,options){const x={requested:{n_threads:options?.n_threads,n_ctx:options?.n_ctx,n_gpu_layers:options?.n_gpu_layers,reasoning:options?.reasoning}};window.__loads.push(x);const result=await original.call(this,files,options);x.actualThreads=this.getNumThreads();x.multithread=this.isMultithread();x.loaded=true;return result;};
+return {crossOriginIsolated,hardwareConcurrency:navigator.hardwareConcurrency,sharedArrayBuffer:typeof SharedArrayBuffer};});
+await p.evaluate(()=>window.__toggleAgentPanel());await p.locator('.agent-panel').waitFor();await p.locator('.agent-panel-settings-toggle').click();await p.locator('.agent-panel-gguf-files').setInputFiles('.scratch/node_modules/ai-models/gemma-4-E2B-it-Q4_0.gguf');await p.locator('.agent-panel-gguf-load').click();
+await p.waitForFunction(()=>document.querySelector('.agent-panel-note')?.textContent?.includes('模型已加载'),null,{timeout:180000});
+r.loads=await p.evaluate(()=>window.__loads);r.note=await p.locator('.agent-panel-note').textContent();
+}catch(e){r.error=String(e);}finally{await c.close();await b.close();r.closed=true;await fs.writeFile('.scratch/2026-10-07-cpu-thread-probe.json',JSON.stringify(r,null,2));console.log(JSON.stringify(r));}
+if(r.error)process.exitCode=1;