Fixed documentation

docs/TROUBLESHOOTING.md

@@ -63,7 +63,7 @@ GPU device.
 
 ```bash
 scripts/install-rocm.sh
-uv run stt-daemon --check
+.venv/bin/stt-daemon --check    # not `uv run`: it syncs the venv and reverts the wheels
 ```
 
 Re-run `scripts/install-rocm.sh` after any later `uv sync`, or after a plain `uv run` (it syncs

scripts/install-rocm.sh

@@ -26,7 +26,7 @@ Example:
   scripts/install-rocm.sh
   scripts/install-rocm.sh --ct2-version 4.8.4 --rocm-index https://download.pytorch.org/whl/rocm7.2
 
-Note: any later `uv sync` reverts these wheels, so re-run this script afterwards.
+Note: any later `uv sync` or plain `uv run` reverts these wheels, so re-run this script afterwards.
 EOF
 }
 
@@ -105,9 +105,13 @@ main() {
     "$venv_python" -c 'import ctranslate2, torch; print(torch.__version__); print(torch.version.hip); print(ctranslate2.get_cuda_device_count()); import sys; sys.exit(1 if torch.version.hip is None or ctranslate2.get_cuda_device_count() == 0 else 0)'
 
     cat << 'EOF'
-ROCm stack installed. Next step: run `uv run stt-daemon --check` to confirm
-(the transcriber/CLI now report the resolved device).
-Note: any later `uv sync` will revert these wheels, so re-run this script afterwards.
+ROCm stack installed. Next step: confirm the resolved device with the venv binary,
+not `uv run` (which syncs the venv and puts the CUDA wheels straight back):
+
+    .venv/bin/stt-daemon --check          # or: uv run --no-sync stt-daemon --check
+
+Expect `device=cuda` / `compute_type=float16`; `device=cpu` means a sync reverted the wheels.
+Note: any later `uv sync` or plain `uv run` reverts these wheels, so re-run this script afterwards.
 EOF
 }