feat: add lightweight real-time native voice preview - #1

What changed

• adds a minimal Windows WPF microphone app under native-app/
• streams 16 kHz PCM chunks every 40 ms into sherpa-onnx OnlineRecognizer and renders partial results while the user speaks
• uses a Chinese Zipformer CTC INT8 model (25.1 MiB) with endpoint reset and final right-context flushing
• supplies initial left context and removes long near-duplicate text at endpoint boundaries
• keeps inference off the UI and NAudio callback threads, validates model files by size and SHA-256, and supports resumable downloads
• adds a streaming smoke benchmark, package script, architecture notes, and Windows Actions artifact

Why

This establishes a smaller, deterministic real-time path before adapting voice input to QuickerPc/Quicker. The preview intentionally excludes global hotkeys and automatic text injection so streaming stability, partial latency, recognition quality, and package size can be evaluated independently.

Measured baseline

• 25.1 MiB streaming model versus 228.45 MiB SenseVoice model (about 89% smaller)
• 5.00 s Chinese WAV produced 9 transcript updates; first partial at about 0.57 s with real-time chunk playback
• 272 s continuous sample completed with 306 updates; endpoint overlap was merged without duplicate phrases
• model-loaded WPF process: about 280 MiB working set / 231 MiB private memory
• framework-dependent publish: 21.4 MB directory / 8.1 MB zip

Validation

• dotnet build native-app/VoxType.Native.slnx -c Release --no-restore — 0 warnings, 0 errors
• scripts/validate-models-catalog.ps1 — native streaming catalog valid
• repeated 5 s real-time WAV streaming smoke — stable final text across chunk timing
• 272 s accelerated streaming smoke — 0.032 RTF with endpoint resets and no boundary duplicate
• published WPF startup — model loaded, controls exposed, window responsive