Windows: GPU never re-detected after swapping CPU build for NVIDIA build in the same data dir #316

Summary

On Windows desktop, a user with an NVIDIA GPU who upgrades/switches from the CPU portable build to the NVIDIA build — reusing the same %LOCALAPPDATA%\StemDeck data dir — stays pinned to CPU. Stem separation runs on CPU and Settings shows CUDA (NVIDIA) — not available, despite a working GPU + driver and CUDA torch bundled in the NVIDIA build. The only recovery today is manually deleting %LOCALAPPDATA%\StemDeck and relaunching.

Root cause

The setup gate in desktop/ui/setup.js treats a persisted cpu-only-package device reason as permanently settled:

const torchDeviceSettled =
  runtime.torchDevice === "cuda" || runtime.torchDevice === "mps" ||
  (runtime.torchDevice === "cpu" && runtime.torchDeviceReason === "cpu-only-package");

When torchDeviceSettled is true (and version matches, which it always does on Windows — no runtime pack), setup early-returns and skips ensure_torch_device entirely. But ensure_torch_device is exactly the function whose clear_stale_cpu_marker + GPU re-probe (#247) would heal the swap. So the self-heal built for this case never gets to run.

The cpu-only-package reason is only valid while the current install is still the CPU-only build. After an in-place swap to the NVIDIA (CUDA) build, that assumption is stale but never re-checked.

Fix

Cross-check the persisted cpu-only-package reason against the current install (is_cpu_only_package(root)). If the persisted reason says CPU-only but the current package is not cpu-only, invalidate the reason so the device is treated as unsettled and setup re-runs ensure_torch_device (which clears the stale marker, detects the GPU, and persists cuda/verified). This reuses the existing #247 "unsettled reason self-heals" path — one heal, then it records the correct device and subsequent launches match.

Scope

• Windows desktop: affected — this fix.
• macOS desktop: not affected (uses versioned runtime pack; mps/mps-unavailable reasons are not treated as permanently settled).
• Docker / self-hosted: not affected (no Tauri app / no persisted-device machinery; image upgrades are clean replacements; Linux torch is CUDA).

Separately tracked follow-up: make the Windows NVIDIA build install cu124 torch explicitly in make-portable.ps1 so its GPU support is reproducible rather than an implicit property of the CI runner's PyPI resolution.