merge triage/cpu-arm: keep ROCm-ignored-on-win/mac inside resolveTorchVariant; docs reflect CPU build on AMD-only Windows

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

backend/api/routers/setup/wizard.py

@@ -108,8 +108,8 @@ def _amd_unusable_note() -> str:
     """Why an AMD card sits idle, per OS - and only options that really exist."""
     if sys.platform == "win32":
         return (
-            "AMD GPU detected, but this install's PyTorch is the NVIDIA CUDA "
-            "build, which cannot drive it, and no ROCm build of the PyTorch "
+            "AMD GPU detected, but this install's PyTorch (the NVIDIA CUDA or "
+            "CPU-only build) cannot drive it, and no ROCm build of the PyTorch "
             "version VoiceStudio ships exists for Windows. PyTorch "
             "engines run on the CPU; audio.cpp (Vulkan) can use the GPU - see "
             "Settings > Performance for the per-engine list."

docs/install/troubleshooting.md

@@ -666,8 +666,8 @@ Apple-Silicon install skips this index entirely.
 busy and the Radeon idle, and **Settings → About → Run self-check** says the
 compute device is `cpu`.
 
-**Cause:** the default install ships the NVIDIA CUDA build of PyTorch, which
-cannot drive AMD GPUs. This is not a driver problem on your side. Open
+**Cause:** the default install has the NVIDIA CUDA build of PyTorch (or the
+CPU-only build when no NVIDIA driver is present), and neither can drive AMD GPUs. This is not a driver problem on your side. Open
 **Settings → Performance → GPU acceleration**: it names your card, the installed
 PyTorch build and, for every engine, whether it uses the GPU.
 

docs/install/windows.md

@@ -115,9 +115,9 @@ ships the CUDA build of PyTorch; with an NVIDIA GPU and a regular NVIDIA
 driver it's picked up automatically (no CUDA Toolkit install needed).
 
 **AMD GPUs — including Ryzen / Ryzen AI integrated Radeon graphics — do not
-accelerate PyTorch engines on Windows.** The installer ships the NVIDIA CUDA
-build of PyTorch 2.8, which cannot drive a Radeon card, so those engines run on
-the CPU. pytorch.org publishes no Windows ROCm wheels, and VoiceStudio's ROCm
+accelerate PyTorch engines on Windows.** The runtime has PyTorch 2.8 as either
+the NVIDIA CUDA build or (with no NVIDIA driver present) the CPU-only build;
+neither can drive a Radeon card, so those engines run on the CPU. pytorch.org publishes no Windows ROCm wheels, and VoiceStudio's ROCm
 option (`OMNIVOICE_TORCH_VARIANT=rocm`) is Linux-only — it is ignored on Windows
 rather than failing setup. (The Ryzen AI NPU is likewise not used.) Everything
 still works on CPU, just slower. What you can do today:

electron/src/main/runtime-project.test.ts

@@ -255,9 +255,10 @@ describe('packaged runtime setup', () => {
     );
     await installRuntime(bundle, project, 'uv', run, new AbortController().signal);
     // No Linux-only index lookup that would fail the whole bootstrap...
-    expect(run.mock.calls.find(([, args]) => args[0] === 'pip' && args[1] === 'install')).toBe(
-      undefined,
+    const rocmLookups = run.mock.calls.filter(([, args]) =>
+      args.some((arg) => arg === ROCM_TORCH_INDEX || ROCM_TORCH_PINS.some((pin) => arg === pin)),
     );
+    expect(rocmLookups).toEqual([]);
     // ...and the default runtime stays reusable instead of being rebuilt forever.
     expect(await runtimeReady(bundle, project)).toBe(true);
   });