research: make the Qwen3.8 RTX 3090 benchmark reproducible on Windows

.github/workflows/research-qwen38-benchmark.yml

@@ -0,0 +1,35 @@
+name: Qwen3.8 research benchmark harness
+
+on:
+  push:
+    branches:
+      - research/qwen38-3090-benchmark-ready
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  benchmark-harness:
+    runs-on: windows-latest
+    timeout-minutes: 10
+    steps:
+      - name: Check out source
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Install test runner
+        run: python -m pip install pytest
+
+      - name: Compile benchmark and tests
+        run: >-
+          python -m py_compile
+          benchmarks/bench_decode_moe.py
+          tests/benchmarks/test_bench_decode_moe.py
+
+      - name: Run focused tests
+        run: python -m pytest -q tests/benchmarks/test_bench_decode_moe.py