fix: force SentenceTransformer to CPU (ZeroGPU)

SentenceTransformer's default auto-detects CUDA at __init__ and
moves the model there. Under ZeroGPU's CUDA emulation, that
initializes CUDA in the parent process, which then prevents
worker_init from binding a real GPU ("No CUDA GPUs are available"
at torch.init(nvidia_uuid)).

Matches the architecture invariant in README ("Embedding pre-filter
runs on CPU before the @spaces.GPU function is invoked").

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

candidate_filter.py

@@ -36,7 +36,11 @@ def load_embedder() -> SentenceTransformer:
     global _embedder
     if _embedder is None:
         print(f"[candidate_filter] Loading {EMBEDDER_NAME}…")
-        _embedder = SentenceTransformer(EMBEDDER_NAME)
+        # device="cpu" is required, not just intentional: SentenceTransformer's
+        # default auto-detects CUDA and moves the model there at __init__,
+        # which initializes CUDA in the parent process — fatal on ZeroGPU
+        # (worker_init then fails with "No CUDA GPUs are available").
+        _embedder = SentenceTransformer(EMBEDDER_NAME, device="cpu")
     return _embedder