fix: build on llama.cpp ubuntu24 base, resolves glibc mismatch

Dockerfile

@@ -1,46 +1,47 @@
 # Paper Decoder — HF Space (Docker SDK)
-# llama.cpp server (official image binaries) + Tesseract OCR (ukr/rus) + Gradio 6
+# Base: official llama.cpp server image (Ubuntu 24.04). Building directly on it
+# avoids the glibc mismatch hit when copying its binaries into Debian bookworm
+# (binaries need GLIBC_2.38; bookworm ships 2.36).
+# llama-server binary and its libs live in /app.
 
-# Stage 1: take llama-server + shared libs from the official CPU server image.
-# Layout verified 2026-06-10 against .devops/cpu.Dockerfile: binary and all
-# *.so* live flat in /app; runtime needs libgomp1 and curl.
-FROM ghcr.io/ggml-org/llama.cpp:server AS llama
+FROM ghcr.io/ggml-org/llama.cpp:server
 
-FROM python:3.11-slim-bookworm
+ENV DEBIAN_FRONTEND=noninteractive \
+    LD_LIBRARY_PATH=/app
 
-# System deps: OCR (Ukrainian + Russian traineddata), llama.cpp runtime libs
+# Python + OCR (Ukrainian + Russian traineddata). curl is already in the base.
 RUN apt-get update && apt-get install -y --no-install-recommends \
+        python3 \
+        python3-pip \
         tesseract-ocr \
         tesseract-ocr-ukr \
         tesseract-ocr-rus \
-        libgomp1 \
-        curl \
     && rm -rf /var/lib/apt/lists/*
 
-COPY --from=llama /app /opt/llama
-ENV LD_LIBRARY_PATH=/opt/llama
-
-# Bake the model into the image at build time: faster Space restarts,
-# no runtime download. Exact file pin recorded after first build.
-RUN pip install --no-cache-dir "huggingface_hub>=1.0,<2.0" && \
+# Bake the model into the image at build time.
+# Pin confirmed from first build log: Qwen3-4B-Instruct-2507-Q4_K_M.gguf
+# Ubuntu 24.04 pip is PEP 668 managed -> --break-system-packages.
+RUN pip3 install --no-cache-dir --break-system-packages "huggingface_hub>=1.0,<2.0" && \
     hf download unsloth/Qwen3-4B-Instruct-2507-GGUF \
-        --include "*Q4_K_M*.gguf" --local-dir /opt/models && \
-    find /opt/models -name "*.gguf" -exec ls -lh {} \;
+        --include "Qwen3-4B-Instruct-2507-Q4_K_M.gguf" --local-dir /opt/models && \
+    ls -lh /opt/models
 
 COPY requirements.txt /tmp/requirements.txt
-RUN pip install --no-cache-dir -r /tmp/requirements.txt
+RUN pip3 install --no-cache-dir --break-system-packages -r /tmp/requirements.txt
 
-# HF Spaces convention: run as non-root user with UID 1000
-RUN useradd -m -u 1000 user
-USER user
-ENV HOME=/home/user \
-    PATH=/home/user/.local/bin:$PATH \
+# Ubuntu 24.04 already ships a UID-1000 user named 'ubuntu' — reuse it
+# (useradd -u 1000 would fail as non-unique).
+USER ubuntu
+ENV HOME=/home/ubuntu \
+    PATH=/home/ubuntu/.local/bin:$PATH \
     GRADIO_SERVER_NAME=0.0.0.0 \
     GRADIO_SERVER_PORT=7860
 
-WORKDIR /home/user/app
-COPY --chown=user app.py start.sh ./
+WORKDIR /home/ubuntu/app
+COPY --chown=ubuntu app.py start.sh ./
 RUN chmod +x start.sh
 
 EXPOSE 7860
+# Parent image's ENTRYPOINT is /app/llama-server — reset it.
+ENTRYPOINT []
 CMD ["./start.sh"]

start.sh

@@ -8,7 +8,7 @@ if [ -z "$MODEL" ]; then
 fi
 echo ">> starting llama-server with $MODEL"
 
-/opt/llama/llama-server \
+/app/llama-server \
   -m "$MODEL" \
   --host 127.0.0.1 \
   --port 8081 \