perf(voice): run STT+TTS on the GPU by default with CPU fallback

Both voice backends defaulted to CPU. Fix the "CUDA unavailable" gaps so
everything that benefits from the RTX 5050 uses it:

- MeloTTS venv had CPU-only torch (2.12.0+cpu) -> installed Blackwell-capable
  torch/torchaudio 2.11.0+cu128 (sm_120 verified with a real GPU matmul).
- faster-whisper CUDA loaded but transcribe() died with "libcublas.so.12 not
  found": installed nvidia-cublas-cu12 + nvidia-cudnn-cu12 into the whisper
  venv and inject those nvidia/*/lib dirs into the worker's LD_LIBRARY_PATH at
  spawn (the loader only honours it at exec).
- WSAI_WHISPER_DEVICE / WSAI_MELO_DEVICE now default to "auto": pick CUDA when
  present, else CPU, and each worker falls back to CPU if a CUDA load fails so
  the voice loop never dies on a GPU-less host.

Verified end-to-end through the real backend classes: both workers report
"ready on cuda"; steady-state STT ~170ms (was ~1350ms CPU), TTS ~4s first call
vs ~23s CPU. All 12 tests pass.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

wsai/backends/whisper.py

@@ -19,7 +19,8 @@ Env:
     WSAI_WHISPER_PYTHON    interpreter with faster-whisper installed
                            (default: /home/claude/jarvis-stt/whisper312/bin/python)
     WSAI_WHISPER_MODEL     model size/name (default: small)
-    WSAI_WHISPER_DEVICE    cpu | cuda   (default cpu; cuda needs GPU approval)
+    WSAI_WHISPER_DEVICE    cpu | cuda | auto  (default auto: GPU if present,
+                           else CPU; the worker falls back to CPU if CUDA fails)
     WSAI_WHISPER_LANGUAGE  forced language, e.g. ko (default ko; "" = autodetect)
 """
 
@@ -41,6 +42,20 @@ log = logging.getLogger("wsai.stt.whisper")
 _DEFAULT_PYTHON = "/home/claude/jarvis-stt/whisper312/bin/python"
 
 
+def _cuda_lib_dirs(python_exe: str) -> list[str]:
+    """nvidia/*/lib dirs of the worker venv (cublas, cudnn, ...), for
+    LD_LIBRARY_PATH so ctranslate2 can dlopen the CUDA runtime. Empty if the
+    interpreter has no such packages (CPU-only install)."""
+    import glob
+
+    # Use the literal path, NOT .resolve(): the venv's bin/python is a symlink
+    # into the uv-managed interpreter, and resolving it would jump out of the
+    # venv and miss its site-packages/nvidia libs.
+    venv = Path(python_exe).parent.parent  # .../bin/python -> venv root
+    dirs = glob.glob(str(venv / "lib" / "python*" / "site-packages" / "nvidia" / "*" / "lib"))
+    return sorted(set(dirs))
+
+
 class WhisperSTT:
     def __init__(
         self,
@@ -53,7 +68,7 @@ class WhisperSTT:
     ) -> None:
         self.python = python or os.environ.get("WSAI_WHISPER_PYTHON", _DEFAULT_PYTHON)
         self.model = model or os.environ.get("WSAI_WHISPER_MODEL", "small")
-        self.device = device or os.environ.get("WSAI_WHISPER_DEVICE", "cpu")
+        self.device = device or os.environ.get("WSAI_WHISPER_DEVICE", "auto")
         # "" means autodetect; a real code like "ko" forces the language.
         env_lang = os.environ.get("WSAI_WHISPER_LANGUAGE", "ko")
         self.language = language if language is not None else (env_lang or None)
@@ -97,6 +112,15 @@ class WhisperSTT:
             "WSAI_WHISPER_MODEL": self.model,
             "WSAI_WHISPER_DEVICE": self.device,
         }
+        # ctranslate2 dlopens libcublas/libcudnn from the whisper venv's nvidia
+        # pip packages; the dynamic loader only honours LD_LIBRARY_PATH captured
+        # at exec, so inject those lib dirs into the child env here (harmless on
+        # CPU). Without this the CUDA model loads but transcribe() dies with
+        # "Library libcublas.so.12 is not found".
+        lib_dirs = _cuda_lib_dirs(self.python)
+        if lib_dirs:
+            prev = env.get("LD_LIBRARY_PATH", "")
+            env["LD_LIBRARY_PATH"] = ":".join(lib_dirs + ([prev] if prev else []))
         if self.language:
             env["WSAI_WHISPER_LANGUAGE"] = self.language
         repo_root = str(Path(__file__).resolve().parents[2])