feat(transcription): wire OnnxNpuTranscriber behind npu feature (T3.4)

src-tauri/src/transcription/mod.rs

@@ -1,8 +1,9 @@
 //! Transcription engine (Phase 1 CPU; Phase 3 acceleration). FR-TRX-*.
 //!
 //! Primary engine: whisper-rs (whisper.cpp) for CPU/Vulkan/CUDA. The NPU path is
-//! a second `Transcriber` impl using ONNX Runtime (ort + DirectML) — same trait,
-//! same `TranscriptSegment` output (ADR-0003/0004). Callers never branch on engine.
+//! a second `Transcriber` impl using ONNX Runtime (ort + OpenVINO EP on Intel
+//! NPUs) — same trait, same `TranscriptSegment` output (ADR-0003/0004). Callers
+//! never branch on engine.
 
 use crate::models::{BackendId, TranscriptSegment};
 use std::path::Path;
@@ -198,35 +199,17 @@ fn audio_ctx_for_window(n_samples: usize) -> i32 {
     ((window_secs / 30.0) * 1500.0).ceil().clamp(64.0, 1500.0) as i32
 }
 
-/// ONNX Runtime + DirectML transcriber for the NPU tier (T3.4).
-///
-/// ponytail: the encoder/decoder inference loop (ORT session + DirectML
-/// execution provider, greedy decode over a Whisper ONNX export) is a
-/// multi-day integration on its own and untestable without NPU hardware in
-/// this environment — ships as a real `Transcriber` impl that fails loudly
-/// instead of a silent no-op. `hardware::WinHardwareDetector` reports NPU as
-/// unavailable (see hardware/mod.rs) so nothing routes here until it's built.
-#[cfg(feature = "directml")]
-pub struct OnnxNpuTranscriber;
-
-#[cfg(feature = "directml")]
-impl Transcriber for OnnxNpuTranscriber {
-    fn load(_model: &Path, _backend: BackendId) -> Result<Self, TrxError> {
-        Err(TrxError::Load(
-            "NPU (ONNX Runtime + DirectML) transcription is not yet implemented (T3.4)".into(),
-        ))
-    }
-    fn transcribe_stream(&self, _audio: AudioWindow, _out: SegmentSink) -> Result<(), TrxError> {
-        Err(TrxError::Inference(
-            "NPU backend not yet implemented".into(),
-        ))
-    }
-    fn transcribe_file(&self, _wav: &Path) -> Result<Vec<TranscriptSegment>, TrxError> {
-        Err(TrxError::Inference(
-            "NPU backend not yet implemented".into(),
-        ))
-    }
-}
+// NPU tier (T3.4): Whisper ONNX on the Intel NPU via ONNX Runtime + OpenVINO.
+// A real `OnnxNpuTranscriber` implementing the same `Transcriber` trait; see
+// `npu.rs`. Gated on the `npu` feature so the CPU-only build never pulls `ort`.
+#[cfg(feature = "npu")]
+pub mod mel;
+#[cfg(feature = "npu")]
+pub mod npu;
+#[cfg(feature = "npu")]
+pub mod onnx_models;
+#[cfg(feature = "npu")]
+pub use npu::OnnxNpuTranscriber;
 
 /// Streaming window worker (Phase 1, T1.5/T1.6): accumulates raw 16kHz-mono
 /// chunks from the `audio` service into fixed-size, **non-overlapping** windows