[Bug] Whisper CUDA/cuBLAS transcription fails on Windows after v1.7.5 upgrade; CPU works

Describe the bug

In Memo AI v1.7.5, local Whisper transcription fails when CUDA/cuBLAS is selected. The same media file, model, and language settings work correctly when the accelerator is changed to CPU.

When CUDA is used, Windows Security reports that part of the application was blocked because the publisher of a DLL loaded by whisper-cli.exe could not be verified. Memo then produces no transcription text and shows that no transcription content was found.

This started after upgrading to Memo AI v1.7.5. The official release notes state that v1.7.5 upgraded Whisper CLI to 1.8.6:
https://memo.ac/zh_tw/releases

To Reproduce

1. Open Memo AI v1.7.5 on Windows.
2. Import a local MP4 file.
3. Select Local Model → Local Whisper.
4. Select source language Traditional Chinese.
5. Select model quality Medium (ggml-medium.bin).
6. Select CUDA as the accelerator.
7. Disable automatic segmentation, background music separation, and voice detection.
8. Start transcription.
9. Windows Security blocks a component used by whisper-cli.exe; Memo produces no subtitle text and reports that no transcription content was found.
10. Repeat the same transcription using CPU; subtitle text is generated successfully.

Expected behavior

CUDA/cuBLAS transcription should load the bundled Whisper components and generate subtitle text, as CPU mode does.

Actual behavior

• CUDA/cuBLAS starts but produces no transcription text.
• Windows Security blocks a DLL used by whisper-cli.exe because its publisher cannot be verified.
• CPU mode with the same Medium model works correctly.

Relevant sanitized log excerpt

[error] nvcc --version error: Command failed: nvcc --version
Start transcription original parameters <cuBLAS>
model: 'ggml-medium.bin'
language: 'zh_t'
localEngine: 'whisper'

Desktop

• OS: Windows x64 (exact edition/build not yet confirmed)
• Model: ggml-medium.bin
• Memo AI version: v1.7.5
• Whisper CLI: 1.8.6 according to the v1.7.5 release notes
• Accelerator: CUDA/cuBLAS
• GPU model and driver: not yet confirmed

Additional context

• Automatic segmentation, background music separation, and voice detection were disabled during the retry.
• A related older issue is #294, but that report concerned Memo v1.3.5 and missing dependencies. This report concerns Memo v1.7.5 and an explicit Windows Security publisher-verification block after the Whisper CLI 1.8.6 upgrade.
• A privacy-redacted full log archive is available. Local paths, media filenames, IP addresses, and task identifiers have been removed while preserving the CUDA/cuBLAS and Whisper diagnostic messages.
• Redacted log SHA-256: fa220ed305bf2b18105ddb7433b2b15177eb36fe695ab53d48a79208bc67f942

Start by reproducing the Windows v1.7.5 workflow with the ggml-medium.bin model, comparing CUDA/cuBLAS against CPU, and inspect the sanitized log around whisper-cli.exe and the nvcc --version error. Check the v1.7.5 release notes for the Whisper CLI 1.8.6 upgrade. Done means CUDA loads its bundled components without the Windows Security block and produces subtitle text like CPU mode.