fix: Improve Whisper model status verification and PyInstaller builds

- Add robust model cache verification (check model.bin + config.json)
- Add new status messages: model_cached, incomplete_cache, model_error
- Forward model status events to frontend for better UI feedback
- Add clean_build_cache() to remove stale spec files before build
- Add --clean flag to PyInstaller commands
- Change sidecar from --onefile to --onedir for faster startup
- Add missing hidden imports: onnxruntime, wave, huggingface_hub.utils

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

sidecar/transcriber.py

@@ -60,12 +60,31 @@ def check_and_download_whisper_model(model_size: str) -> bool:
     repo_cache_name = f"models--Systran--faster-whisper-{model_size}"
     model_cache_path = cache_dir / repo_cache_name
 
-    # Check if model files exist
+    # Check if model files exist - verify essential files are present
     if model_cache_path.exists():
         snapshots_dir = model_cache_path / "snapshots"
-        if snapshots_dir.exists() and any(snapshots_dir.iterdir()):
-            # Model is cached, no download needed
-            return True
+        if snapshots_dir.exists():
+            # Check for actual model files, not just any file
+            for snapshot in snapshots_dir.iterdir():
+                if snapshot.is_dir():
+                    # Essential faster-whisper model files
+                    required_files = ["model.bin", "config.json"]
+                    has_all_files = all(
+                        (snapshot / f).exists() for f in required_files
+                    )
+                    if has_all_files:
+                        print(json.dumps({
+                            "status": "model_cached",
+                            "model": model_size,
+                            "path": str(snapshot)
+                        }), flush=True)
+                        return True
+            # Snapshots exist but no valid model found
+            print(json.dumps({
+                "status": "incomplete_cache",
+                "model": model_size,
+                "message": "Model cache incomplete, will re-download"
+            }), flush=True)
 
     # Model not cached, need to download
     print(json.dumps({
@@ -491,20 +510,29 @@ class Transcriber:
 
         try:
             # Check if model needs to be downloaded (with progress reporting)
-            check_and_download_whisper_model(model_size)
+            download_ok = check_and_download_whisper_model(model_size)
+            if not download_ok:
+                print(json.dumps({
+                    "status": "model_error",
+                    "error": "Failed to download model"
+                }), flush=True)
+                raise RuntimeError("Failed to download Whisper model")
 
             # Now load the model
             print(json.dumps({"status": "loading_model", "model": model_size}), flush=True)
             self.model = WhisperModel(model_size, device=device, compute_type=compute_type)
             self.converter = opencc.OpenCC("s2twp")
-            print(json.dumps({"status": "model_loaded"}), flush=True)
+            print(json.dumps({"status": "model_loaded", "model": model_size}), flush=True)
 
             # Pre-load VAD model at startup (not when streaming starts)
             if ONNX_AVAILABLE:
                 self.vad_model = SileroVAD()
 
         except Exception as e:
-            print(json.dumps({"error": f"Failed to load model: {e}"}), flush=True)
+            print(json.dumps({
+                "status": "model_error",
+                "error": f"Failed to load model: {e}"
+            }), flush=True)
             raise
 
     def transcribe_file(self, audio_path: str, add_punctuation: bool = False) -> str: