feat(transcription): drive engine choice from AccelPath resolver (P0)

load_transcriber now resolves the accel path instead of assuming a GPU
BackendId has a working offload. Only requests whisper GPU offload for
WhisperCuda/WhisperVulkan; NPU goes to the ONNX engine; everything else decodes
on CPU and reports the backend actually used.

src-tauri/src/commands.rs

@@ -119,8 +119,15 @@ fn load_transcriber(
     backend: BackendId,
     whisper_model: &Path,
 ) -> Result<(Box<dyn Transcriber>, BackendId), crate::transcription::TrxError> {
+    use crate::hardware::{resolve_accel, AccelPath};
+
+    // Resolve how this backend is actually served in *this* build (CUDA/Vulkan
+    // baked in? NPU/DirectML runtime present?) rather than assuming a GPU
+    // backend has a working accel path just because the hardware exists.
+    let path = resolve_accel(backend);
+
     #[cfg(feature = "npu")]
-    if backend == BackendId::Npu {
+    if path == AccelPath::OnnxOpenVino {
         use crate::transcription::{onnx_models, OnnxNpuTranscriber};
         if onnx_models::is_installed(onnx_models::DEFAULT_ONNX_MODEL) {
             let dir = onnx_models::model_dir(onnx_models::DEFAULT_ONNX_MODEL);
@@ -134,10 +141,18 @@ fn load_transcriber(
             );
         }
     }
-    match WhisperTranscriber::load(whisper_model, backend) {
-        Ok(t) => Ok((Box::new(t), backend)),
-        Err(e) if backend != BackendId::Cpu => {
-            tracing::warn!("backend {backend:?} failed to load ({e}); falling back to CPU");
+
+    // whisper.cpp path: only ask for GPU offload when the resolver picked a
+    // whisper GPU backend that's compiled in — otherwise a bogus `use_gpu` for a
+    // vendor with no accel path just no-ops. Anything else decodes on the CPU.
+    let whisper_backend = match path {
+        AccelPath::WhisperCuda | AccelPath::WhisperVulkan => backend,
+        _ => BackendId::Cpu,
+    };
+    match WhisperTranscriber::load(whisper_model, whisper_backend) {
+        Ok(t) => Ok((Box::new(t), whisper_backend)),
+        Err(e) if whisper_backend != BackendId::Cpu => {
+            tracing::warn!("backend {whisper_backend:?} failed to load ({e}); falling back to CPU");
             WhisperTranscriber::load(whisper_model, BackendId::Cpu)
                 .map(|t| (Box::new(t) as Box<dyn Transcriber>, BackendId::Cpu))
         }