[Bug] Sherpa ONNX/Qwen3 downloads successfully but is absent from the local engine selector on macOS v1.7.5

Describe the bug

On Memo v1.7.5 for macOS (Apple Silicon), the Qwen3-ASR-0.6B-Int8 model can be downloaded successfully from Settings → Transcription Settings → Sherpa ONNX, but Sherpa ONNX is not available in the Home screen's “Local Model → Select Engine” dropdown.

As a result, the downloaded Qwen3 model cannot be selected or used for video/audio transcription.

Steps to reproduce

1. Open Memo v1.7.5 on macOS.
2. Navigate to Settings → Transcription Settings → Sherpa ONNX.
3. Download Qwen3-ASR-0.6B-Int8.
4. Confirm that the download completes.
5. Return to the Home screen and select a local video or audio file.
6. Open Local Model → Select Engine.
7. Observe that the menu contains:
   • Local Whisper
   • Parakeet (not installed)
   • FunASR (not installed)
8. Observe that Sherpa ONNX is missing entirely.
9. Quit and relaunch Memo; the behavior remains unchanged.

Expected behavior

After a supported Sherpa ONNX model has been downloaded:

• Sherpa ONNX should appear in the local engine selector.
• Selecting it should expose the installed Sherpa models.
• Qwen3-ASR-0.6B-Int8 should be selectable for transcription.

Actual behavior

• The model catalog reports that Qwen3 has been downloaded.
• The model files are present on disk.
• Sherpa ONNX does not appear in the local engine selector.
• Memo therefore falls back to Local Whisper, and the Qwen3 model is unusable from the UI.

Download verification

The following directory is present and occupies approximately 954 MB:

~/Library/Application Support/Memo/models/
  sherpa-onnx-qwen3-asr-0.6B-int8-2026-03-25/

It contains the expected model assets, including:

conv_frontend.onnx
encoder.int8.onnx
decoder.int8.onnx
tokenizer/

This indicates that the issue is not an incomplete or failed model download.

Additional diagnostic evidence

A read-only inspection of the v1.7.5 application bundle shows that it includes:

• The sherpa-onnx-darwin-arm64 native runtime.
• A backend configuration for sherpa-onnx-qwen3-asr-0.6B-int8-2026-03-25.

However, the renderer bundle appears to contain logic that resets localEngine === "sherpa-onnx" back to "whisper", while the dedicated Sherpa item in the local-engine selector is not rendered.

This suggests a UI feature-flag/build-condition regression or an incomplete Sherpa rollout, rather than a missing runtime or model asset.

Environment

• Memo version: 1.7.5
• OS: macOS 26.5.2 (Build 25F84)
• Architecture: arm64
• Device: Mac mini (Mac16,11)
• Chip: Apple M4 Pro
• Memory: 64 GB
• Model: Qwen3-ASR-0.6B-Int8
• Model directory: Default Memo model directory

Impact

This is particularly confusing because the Settings UI presents Qwen3 as a supported, downloadable model and consumes approximately 1 GB of disk space, while the Home UI provides no way to use it.

Suggested resolution

Please consider one of the following:

1. Restore/enable the Sherpa ONNX entry in the local-engine selector and preserve the selected sherpa-onnx engine state; or
2. If Sherpa transcription is intentionally disabled in this release, hide or disable its model-download controls and explain the limitation in the UI.

Thank you for looking into this.

Start in the renderer bundle's local-engine selector and trace the localEngine === "sherpa-onnx" reset, then compare it with the Settings model catalog and Sherpa ONNX backend configuration for Qwen3. Reproduce with the downloaded model and verify that Sherpa ONNX is rendered, remains selected, exposes Qwen3, and can start transcription.