feat: add pixi task runner to replace manual setup

[maciejmajek]

feat: add pixi task runner to replace manual setup

This PR adds pixi, a package manager and task runner, to replace the previous 20-step manual setup process. DEMO_ROOT and O3DE_ROOT are now set automatically from the repo clone location — no .bashrc exports needed.

What changed

• pixi.toml — defines all build and run tasks; single source of truth for the project workflow
• pixi_activate.sh — automatically sources ROS 2 Jazzy and the project workspace on every pixi run invocation
• pixi.lock — lockfile for reproducible environments
• scripts/download_models.sh — downloads all 5 models from HuggingFace automatically (embedding and reranker now included; SharePoint links removed)
• scripts/serve_all.sh — launches all 4 inference servers in a tmux 2×2 grid via pixi run serve-all
• scripts/smoke_test.sh — health-checks all inference endpoints with green/red color output
• config.toml — removed stale manual llama-server command comments
• pyproject.toml — added tool.uv.workspace to exclude inference/ from uv workspace
• .gitignore — added models/
• All setup and running docs updated to use pixi run commands

Setup (was: ~20 manual steps)

curl -fsSL https://pixi.sh/install.sh | sh  # install pixi once
pixi run setup                               # full single-machine setup

For the 3-machine demo setup:

• Sim machine: pixi run setup
• HIL machine: pixi run setup-hil (skips O3DE — no 30–60 min build)
• HMI machine: pixi run setup-hil

Key improvements

• No more .bashrc exports — DEMO_ROOT and O3DE_ROOT are set from $PIXI_PROJECT_ROOT, so the repo works from any clone location
• No more manual source calls — ROS 2 environment is sourced automatically before every task
• HIL-optimized setup — pixi run setup-hil clones and builds only the ROS 2 workspace and Python dependencies, skipping O3DE entirely
• Local inference — pixi run build-llama / pixi run setup-hil-local replaces manual llama.cpp clone + cmake steps
• Automated model downloads — pixi run download-models fetches all 5 models from HuggingFace; embedding and reranker no longer require manual SharePoint downloads
• Run tasks — pixi run sim, pixi run ros2, pixi run orchestrator, etc. replace manual binary invocations and script calls
• Smoke test — pixi run smoke-test checks all inference server endpoints with color-coded output

Available tasks

Task	Description
pixi run setup	Full single-machine setup
pixi run setup-hil	HIL machine setup (ROS 2 + Python, no O3DE)
pixi run setup-hil-local	HIL setup + local llama.cpp inference
pixi run setup-local	Full setup + local llama.cpp inference
pixi run build-llama	Clone and build llama.cpp standalone
pixi run download-models	Download all 5 GGUF models from HuggingFace
pixi run sim	Launch O3DE simulation
pixi run ros2	Launch full ROS 2 stack
pixi run orchestrator	Launch agent orchestrator
pixi run hmi	Launch HMI
pixi run serve-all	Launch all 4 inference servers in a tmux 2×2 grid
pixi run serve-llm	Serve GPT-OSS-20B on port 8080
pixi run serve-vlm	Serve LFM2-VL-3B on port 8081
pixi run serve-embedding	Serve Qwen3-Embedding-0.6B on port 8082
pixi run serve-reranker	Serve Qwen3-Reranker-0.6B on port 8083
pixi run smoke-test	Health-check all inference servers
pixi run lint	Run all linters and formatters

[knicked]

Due to issues mentioned in #21, the project does not build in the current state.

I ported the fixes to pixi in a separate PR (#24). I will test the project with my changes and leave a comment here when it works.