Bug Description

Sending a voice note through the Hermes gateway on Apple Silicon causes the gateway Python process to hard-crash with a macOS "Python quit unexpectedly" dialog instead of transcribing the audio.

This appears to be the local STT path using faster-whisper / ctranslate2 with device="auto". On this machine, ctranslate2 aborts inside Intel OpenMP while the Python process is running x86_64 under Rosetta.

Steps to Reproduce

1. Run Hermes gateway as a launchd service on macOS Apple Silicon.
2. Configure/use local STT via faster-whisper.
3. Send a Telegram voice note to Hermes.
4. Observe that the voice note is cached, then the gateway/Python process crashes.

Expected Behavior

Voice notes should either:

• Transcribe successfully, or
• Fail gracefully and fall back to another STT provider/return an error.

The gateway should not terminate.

Actual Behavior

macOS shows a crash dialog for Python. Hermes gateway restarts, but every subsequent voice note reproduces the crash.

Crash signature from ~/Library/Logs/DiagnosticReports/Python-2026-05-19-091521.ips:

procName: Python
coalitionName: ai.hermes.gateway
translated: true
exception: EXC_CRASH / SIGABRT
termination: Abort trap: 6
faulting stack includes: __kmp_abort_process

Earlier reports on the same machine show the same signature:

Python-2026-05-19-084427.ips
Python-2026-05-19-084836.ips
Python-2026-05-19-085950.ips
Python-2026-05-19-091144.ips
Python-2026-05-19-091521.ips

Environment

• Hermes gateway process: hermes_cli.main gateway run --replace
• OS: macOS 26.3.1, build 25D2128
• Hardware: Apple Silicon Mac, model Mac16,10
• Python: 3.11.15 from /usr/local/Cellar/[email protected]/
• Python process architecture: x86_64 translated under Rosetta (translated: true in crash report)
• STT path: local faster-whisper

Suspected Root Cause

tools/transcription_tools.py::_load_local_whisper_model() currently tries:

WhisperModel(model_name, device="auto", compute_type="auto")

On this Apple Silicon/Rosetta setup, the device="auto" path can trigger a native abort in the backend before Python can catch an exception. Existing CUDA fallback logic only handles Python exceptions, so it cannot recover from SIGABRT.

Proposed Fix

Detect Apple Silicon/macOS Rosetta before loading faster-whisper and use CPU directly:

WhisperModel(model_name, device="cpu", compute_type="int8")

This avoids the GPU/Metal/OpenMP autodetection path that is crashing.