Bug: Voice transcription crashes with CUBLAS_STATUS_NOT_SUPPORTED on RTX 5090 (Blackwell, sm_120). The _CUDA_LIB_ERROR_MARKERS fallback list in tools/transcription_tools.py doesn't include this Blackwell-specific cuBLAS error, so the load fails outright instead of falling back to CPU as designed.

Environment:

• Hermes Agent v0.11.0 (v2026.4.23)
• Ubuntu 24.04, Python 3.11.15
• RTX 5090 (32 GB, sm_120 / Blackwell)
• faster-whisper bundled in voice extra (CTranslate2 backend)

Repro:

1. Fresh Hermes install on a host with a Blackwell GPU
2. Send any voice message via Telegram → Hermes caches the .ogg
3. Transcription tries default device="auto" → loads on CUDA → at first transcribe call, CTranslate2 raises RuntimeError: cuBLAS failed with status CUBLAS_STATUS_NOT_SUPPORTED
4. The mid-transcribe fallback in _transcribe_local calls _looks_like_cuda_lib_error() to decide whether to fall back to CPU, but CUBLAS_STATUS_NOT_SUPPORTED is not in _CUDA_LIB_ERROR_MARKERS, so the exception is re-raised and the voice message dies silently from the user's perspective.
5. Manually running WhisperModel("base", device="cpu", compute_type="int8") works fine and transcribes correctly.

Stack trace excerpt:

ERROR tools.transcription_tools: Local transcription failed: cuBLAS failed with status CUBLAS_STATUS_NOT_SUPPORTED
  File ".../faster_whisper/transcribe.py", line 1400, in encode
    return self.model.encode(features, to_cpu=to_cpu)
RuntimeError: cuBLAS failed with status CUBLAS_STATUS_NOT_SUPPORTED

Root cause: tools/transcription_tools.py:325-334 defines _CUDA_LIB_ERROR_MARKERS to catch dlopen-style failures and missing-runtime errors (libcublas, libcudnn, libcudart, cannot be loaded, etc.) but doesn't catch architecture-mismatch errors. On Blackwell (sm_120) the bundled CTranslate2 wheel's cuBLAS calls return CUBLAS_STATUS_NOT_SUPPORTED because the GEMM kernels weren't compiled with sm_120 support. The library loads fine, but every actual GPU op fails.

Suggested fix: Add CUBLAS_STATUS_NOT_SUPPORTED (or the broader CUBLAS_STATUS_) to _CUDA_LIB_ERROR_MARKERS. The mid-transcribe fallback at _transcribe_local() will then evict the broken model and reload on CPU, identical to how it currently handles dlopen failures.

_CUDA_LIB_ERROR_MARKERS = (
    "libcublas",
    "libcudnn",
    "libcudart",
    "cannot be loaded",
    "cannot open shared object",
    "no kernel image is available",
    "no CUDA-capable device",
    "CUDA driver version is insufficient",
    "CUBLAS_STATUS_NOT_SUPPORTED",   # NEW: Blackwell sm_120
    "CUBLAS_STATUS_",                 # NEW: catch other arch-mismatch variants
)

Confirmation: Patched locally; voice message that previously crashed now transcribes correctly with the documented warning:

faster-whisper CUDA runtime failed mid-transcribe (cuBLAS failed with status CUBLAS_STATUS_NOT_SUPPORTED) — evicting cached model and retrying on CPU (int8).

Affected users: anyone with a Blackwell GPU (RTX 5090, B100, B200, etc.) and the bundled CTranslate2 wheel. Will become more common as 5090 adoption increases.