macOS watchdog kernel panic when local LLM polishing begins (M4 Air, 16 GB, v1.7.5)

[bananakiu]

Describe the bug

OpenWhispr's local transcription completes successfully, but when local LLM polishing/cleanup begins, the entire Mac becomes unresponsive and eventually requires or undergoes a reboot. This happened with two very different local models, so it does not appear to be caused solely by model size:

• gpt-oss-20b-mxfp4
• Llama-3.2-3B-Instruct-Q4_K_M.gguf (approximately 2.02 GB)

The Llama model is correctly Q4_K_M quantized. The same failure occurs at the same pipeline stage with both models.

macOS generated a kernel panic after the latest incident:

watchdog timeout: no checkins from watchdogd in 91 seconds

The panic report said:

• Compressor: 18% of compressed-pages limit (OK)
• Segments: 16% of limit (OK)
• 9 swapfiles with swap space reported as OK
• Panicked task: kernel_task

This suggests the incident was not ordinary memory exhaustion. I did not find an explicit GPU Restart or Metal fault string in the panic report.

macOS also generated repeated resource diagnostics specifically for OpenWhispr:

• 2,147.54 MB of file-backed writes over 1,800 seconds
• 8,590.00 MB over 338 seconds (about 25.41 MB/s)
• 2,147.50 MB over 994 seconds

The 8.59 GB event's heaviest OpenWhispr stack was repeatedly blocked in write() through Electron/Node. OpenWhispr's own recorded footprint in that report was only approximately 113-153 MB. The Llama GGUF finished being written shortly before the later freeze.

To reproduce

1. Run OpenWhispr 1.7.5 on an M4 MacBook Air with 16 GB unified memory.
2. Configure local Whisper base for transcription.
3. Configure local dictation cleanup/polishing with either:
   • gpt-oss-20b-mxfp4, or
   • Llama-3.2-3B-Instruct-Q4_K_M.gguf.
4. Dictate a short passage.
5. Observe that transcription finishes successfully.
6. When local polishing begins, the entire computer freezes.
7. After the latest occurrence, macOS recorded a watchdog kernel panic.

Expected behavior

The quantized 3B cleanup model should polish the transcript without making macOS unresponsive. If local inference cannot start safely, OpenWhispr should fail gracefully, stop the model process, and show an error.

Desktop

• Device: MacBook Air (Mac16,12)
• Chip: Apple M4, 10 cores
• Memory: 16 GB unified memory
• OS: macOS 26.1 (25B78)
• OpenWhispr: 1.7.5 (arm64)
• Local transcription: Whisper base
• Local cleanup model: Llama-3.2-3B-Instruct-Q4_K_M.gguf
• Llama model size: 2,019,377,696 bytes

Additional context

The issue is tightly correlated with the local polishing stage:

• Opening OpenWhispr does not cause the freeze.
• Local transcription works and produces text.
• The freeze begins only when cleanup/polishing starts.
• Changing from a 20B model to a correctly quantized 3B model does not prevent it.
• Plain local transcription without local polishing is stable.

Relevant local reports are available under /Library/Logs/DiagnosticReports/:

• OpenWhispr_2026-07-16-135347_*.diag
• OpenWhispr_2026-07-16-135926_*.diag
• OpenWhispr_2026-07-16-230643_*.diag
• panic-full-2026-07-16-232329.0002.panic

I have not attached the raw reports yet because they may contain private machine identifiers. I can provide redacted excerpts or reports if needed.

Please let me know which additional logs or debug-mode reproduction data would be most useful. I am avoiding another local-polishing reproduction for now because it caused a system-level watchdog panic.