[Bug] Voice mode crashes CLI with ONNX Runtime assertion in Nemotron ASR on Linux

从提供的 systemd-coredump 跟踪以及随附的 Microsoft.AI.Foundry.Local.Core.so 和 libonnxruntime-genai.so 工件开始。跟踪 /voice 录音路径经过 NemotronSpeechState::StepToken 和 OrtSession::Run 的过程，然后使用列出的模型和 runtime 在 Linux x64 上复现。完成的标准是：兼容性故障已得到分析，且 CLI 在转录期间不再中止，或者问题已缩小到随附的 native stack。

Describe the bug

Enabling and using voice input causes GitHub Copilot CLI to abort with signal 6 (SIGABRT) and dump core on Linux x64. The crash occurs while the local Nemotron speech model is processing audio.

Copilot CLI version

1.0.83

Environment

• OS: Manjaro Linux
• Kernel: 7.1.13-2-MANJARO
• Architecture: x86_64
• glibc: 2.44
• Node.js runtime shown by CLI logs: v24.20.0
• Microphone: HD-Audio Generic / ALC257 Analog
• Voice model: nemotron-3.5-asr-streaming-0.6b-generic-cpu:3

Steps to reproduce

1. Start Copilot CLI 1.0.83 on Linux x64.
2. Enable voice input with /voice.
3. Start recording and speak into the default microphone.
4. Stop recording / allow transcription to start.
5. The Copilot CLI process aborts and the system records a core dump.

Expected behavior

The spoken input is transcribed and inserted into the prompt without terminating the CLI.

Actual behavior

The CLI terminates with Signal: 6 (ABRT). The crash stack is in ONNX Runtime GenAI and Foundry Local Core:

__assert_fail (libc.so.6)
libonnxruntime.so
OrtSession::Run (libonnxruntime-genai.so)
Generators::NemotronEncoderSubState::Run
Generators::NemotronSpeechState::StepToken
OgaGenerator_GenerateNextToken
Microsoft.AI.Foundry.Local.Core.so

The process command line was:

/home/soenke/.nvm/versions/node/v24.13.0/lib/node_modules/@github/copilot/node_modules/@github/copilot-linux-x64/copilot --prefer-version 1.0.83

The complete systemd-coredump stack trace is attached in the report supplied with this issue.

Additional context

• The microphone is detected by ALSA and appears as a capture device.
• The voice server starts successfully and accepts a client connection.
• A previous voice-server log entry also reported PvRecorder failed to read audio data frame, but the fatal crash reported here happens later in the Nemotron/ONNX Runtime inference path.
• Updating with copilot update confirms that 1.0.83 is currently the latest available version.
• The crash appears to be in the bundled native voice stack (Microsoft.AI.Foundry.Local.Core.so / libonnxruntime-genai.so), rather than in repository code or the prompt UI.
• Related issue: #4024 (voice transcription routing failure for nemotron_speech); this report is different because the CLI aborts instead of returning an empty transcript.

Logs

A full stack trace is available from the reporter and can be provided if needed. The key frames are:

#4  __assert_fail (libc.so.6)
#5  ... (libonnxruntime.so)
#20 OrtSession::RunEP ... (libonnxruntime-genai.so)
#21 Generators::NemotronEncoderSubState::Run ...
#22 Generators::NemotronSpeechState::StepToken ...
#23 Generators::Generator::GenerateNextToken ...
#24 OgaGenerator_GenerateNextToken ...
#25+ Microsoft.AI.Foundry.Local.Core.so

Could the voice inference path handle this model/runtime error without aborting the entire CLI, and could the Linux CPU Nemotron model compatibility be investigated?