Revamp Dockerfile.backend to a multi-stage build with an INSTALL_DOCLING build-arg: the builder stage installs dependencies and (optionally) downloads docling/EasyOCR models when INSTALL_DOCLING=1, while the final stage is a minimal runtime that copies the installed app. pyproject.toml now moves torch to an optional 'torch' extra and ensures docling-serve is managed via extras so heavy ML runtimes are not required by default. src/main.py defers and guards the torch import, logging CUDA info only when torch is present. src/tui/managers/docling_manager.py adds a runtime check for docling-serve and fails gracefully with guidance if it's not installed. These changes allow smaller images by default and make startup resilient when optional ML packages are absent.

Convert Dockerfile.backend to a multi-stage BuildKit-friendly layout (syntax header + base/builder/runtime stages). Install uv in a lightweight base image, move build-time deps to a builder stage, and make docling/model installation optional via ARG INSTALL_DOCLING=1. Use --no-install-recommends for apt, add cache mount for uv, conditionally run uv sync --extra docling and download docling/EasyOCR models into /app/.EasyOCR when requested, and clean temporary caches. The final runtime stage remains minimal and copies the prepared /app from the builder.

…cling-core-optional

…penrag into make-docling-core-optional

Capture and log docling start output and architecture in Makefile; extract endpoint from output, default to http://localhost:5001 if missing, run status checks, and emit a clear error if health checks fail. Refactor docling host detection to try well-known hostnames first, use container-specific fallbacks (gateway/bridge) and add clearer logging. Treat .md files like .txt for simple processing to skip docling in processors and document service. Also add several local documentation PDFs to .gitignore.

Introduce a configurable startup timeout for docling and make CI fail-fast with cleanup. Added a --timeout argument to scripts/docling_ctl.py and forwarded it to DoclingManager.start; DoclingManager.start now accepts a timeout and waits up to that many seconds for the service to listen (with periodic log updates). Updated Makefile test-ci and test-ci-local to pass --timeout 180, and on startup failure the pipeline now stops docling, brings down compose services, and exits with error. Also clarified some startup/failure log messages.

Simplify how the external docling serve process is started by removing shutil-based checks and fallbacks for `uv` or `python -m docling_serve`. The code now always builds a fixed `uvx ... docling-serve==1.5.0 run` command (keeping platform-specific OCR package selection). Also small whitespace added in the docstring. This reduces branching and assumes `uvx` is available in the runtime environment.

Remove the apt-get install libgl1 step from the integration workflow and add --with opencv-python-headless to the uvx install args for docling-serve==1.5.0. This switches OpenCV to a headless Python package to avoid needing system GL libraries on the CI runner.

On non-mac platforms, create a temporary uvx override file to replace opencv-python with opencv-python-headless so headless CI runners lacking libGL can run docling. Add tempfile import, build the uvx command to pass --override and --with opencv-python-headless, and ensure the override file is removed after resolution or on errors. Minor command assembly refactor and cleanup on FileNotFoundError/exception paths.