fix: Use a prebuilt CPU wheel for llama-cpp-python, not a source build

The first Space rebuild after adding llama-cpp-python failed with
BUILD_ERROR / "Job timeout": pip was compiling llama.cpp's C++ core from
source, which routinely exceeds HF Spaces' build-job time limit. Pointing
at abetlen's prebuilt wheel index resolves directly to a 24MB manylinux2014
x86_64 wheel instead (verified locally with pip download --only-binary),
skipping the compile entirely.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

requirements.txt

@@ -13,7 +13,14 @@ control>=0.9.4
 cvxpy>=1.4.0
 rich>=13.7.0
 rank-bm25>=0.2.2
-llama-cpp-python>=0.3.0
+# Pulls a prebuilt CPU wheel instead of compiling llama.cpp's C++ core from
+# source -- a source build routinely exceeds HF Spaces' build-job timeout
+# ("Job timeout" BUILD_ERROR, observed directly). This backend is CPU-only by
+# design (see the GGUF branch in orchestrator.py): llama.cpp's raw CUDA calls
+# aren't visible to ZeroGPU's torch-based interception, so a GPU wheel
+# wouldn't get real GPU time here anyway.
+--extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu
+llama-cpp-python==0.3.35
 pypdf>=3.17.0
 pymupdf>=1.23.0
 scikit-learn>=1.3.0