fix: install required CUDA and cuDNN runtime dependencies

README.md

@@ -107,7 +107,8 @@ cargo test
 - **Update checks** run only on demand or when explicitly enabled in settings.
 
 ## Recently added
-
+
+- **NVIDIA CUDA GPU Acceleration** — offload local Parakeet model inference to NVIDIA GPUs with on-demand runtime downloading and auto CPU fallback. CUDA setup reuses a complete CUDA 11/cuDNN 8 runtime available through `PATH`; otherwise it downloads pinned private copies. Setup downloads up to 1.52 GiB of archives and installs up to 2.33 GiB of runtime files, including proprietary NVIDIA components covered by the terms in [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md).
 - **Audio Device Selection & VU Meter** — select physical microphone devices directly in Settings with a real-time input level meter and Silero VAD detection.
 - **Post-Release Audio Grace Buffer (Tail Hold)** — 160ms tail buffer and expanded VAD margin to completely prevent trailing word and syllable clipping.
 - **Context Editing & Safe Clipboard Handoff** — AI sparkle indicator for selection editing; if target window focus shifts during transcription, text is safely copied to the clipboard with an overlay notification.

THIRD_PARTY_NOTICES.md

@@ -0,0 +1,28 @@
+# Third-Party Notices
+
+Aura is licensed under AGPL-3.0-only. Optional CUDA acceleration downloads
+additional proprietary NVIDIA runtime components after installation. Those
+components are not part of Aura and are not licensed under the AGPL.
+
+## NVIDIA CUDA Toolkit 11.8
+
+When compatible system copies are unavailable, Aura downloads the CUDA runtime
+libraries cuBLAS 11.11.3.6 and cuFFT 10.9.0.58 directly from
+`developer.download.nvidia.com`. Use and distribution are governed by the NVIDIA
+CUDA Toolkit license:
+
+https://docs.nvidia.com/cuda/archive/11.8.0/eula/index.html
+
+## NVIDIA cuDNN 8.5.0
+
+When a compatible system copy is unavailable, Aura downloads the cuDNN 8.5.0.96
+inference runtime directly from `developer.download.nvidia.com`. Use and
+distribution are governed by the NVIDIA cuDNN Software License Agreement:
+
+https://docs.nvidia.com/deeplearning/cudnn/archives/cudnn-850/sla/index.html
+
+The CUDA setup flow presents both agreements before installing optional CUDA
+support. Compatible libraries already available through the process `PATH` are
+reused; downloaded runtime files are installed in Aura's private application
+data directory. Both layouts are used only by Aura's local transcription
+processes.

src-tauri/capabilities/main.json

@@ -19,6 +19,7 @@
     "allow-delete-gpu-binaries",
     "allow-cancel-gpu-download",
     "allow-check-gpu-downloaded",
+    "allow-check-nvidia-runtime-on-path",
     "allow-get-history",
     "allow-clear-history",
     "allow-copy-to-clipboard",

src-tauri/permissions/gpu_commands.toml

@@ -12,3 +12,8 @@ commands.allow = ["delete_gpu_binaries"]
 identifier = "allow-check-gpu-downloaded"
 description = "Enables the check_gpu_downloaded command."
 commands.allow = ["check_gpu_downloaded"]
+
+[[permission]]
+identifier = "allow-check-nvidia-runtime-on-path"
+description = "Enables checking whether the required NVIDIA runtime is available on PATH."
+commands.allow = ["check_nvidia_runtime_on_path"]