ci: install PortAudio, and document it as a prerequisite

pyaudio publishes only Windows wheels, so on Linux and macOS it builds from
source against portaudio.h. It reaches us transitively through
unitree-webrtc-connect, which means every job that runs `uv sync` needs the
header -- the lint job had no apt step at all, and the test job installed only
libgl1/libgomp1.

This is not just a CI problem: a plain `pip install multiset-unitree-go2-dimos`
on a fresh Ubuntu or Jetson fails the same way, so both READMEs and
CONTRIBUTING now name the prerequisite instead of leaving it to a C compiler
error partway through the install.

Author: abhisheksonnakula

.github/workflows/ci.yml

@@ -18,6 +18,10 @@ jobs:
       - uses: astral-sh/setup-uv@v5
         with:
           enable-cache: true
+      - name: System libraries
+        # pyaudio has no Linux wheel and builds against portaudio.h. It arrives
+        # transitively via unitree-webrtc-connect, so every job that syncs needs it.
+        run: sudo apt-get update -qq && sudo apt-get install -y -qq portaudio19-dev
       # Pinned to [tool.mypy]'s target: unpinned, numpy's PEP 695 stubs are
       # unparseable under python_version 3.10 and mypy dies before our code.
       - run: uv sync --all-extras --python 3.10
@@ -59,10 +63,12 @@ jobs:
         with:
           enable-cache: true
       - name: System libraries
-        # opencv and open3d need a GL runtime that the runner image lacks.
+        # opencv and open3d need a GL runtime that the runner image lacks, and
+        # pyaudio (via unitree-webrtc-connect) has no Linux wheel -- it compiles
+        # against portaudio.h.
         run: |
           sudo apt-get update -qq
-          sudo apt-get install -y -qq libgl1 libgomp1
+          sudo apt-get install -y -qq libgl1 libgomp1 portaudio19-dev
       - name: Install the workspace
         run: uv sync --all-extras --python ${{ matrix.python-version }}
       - name: Both packages import side by side

CONTRIBUTING.md

@@ -3,12 +3,16 @@
 ## Setup
 
 ```bash
+sudo apt-get install -y portaudio19-dev   # or: brew install portaudio
 git clone https://github.com/MultiSet-AI/multiset-python
 cd multiset-python
 uv sync --all-extras
 uv run pytest
 ```
 
+PortAudio is not optional: `unitree-webrtc-connect` pulls `pyaudio`, which has
+no Linux or macOS wheel and builds from source. CI installs it too.
+
 `uv sync` installs every package in `packages/` editable into one venv. Each
 adapter's `[tool.uv.sources]` points `multiset-vps` at `packages/vps` in your
 working tree, so a change to the core is live everywhere immediately. uv strips

README.md

@@ -36,6 +36,10 @@ pip install \
 Both lines are required: pip does not follow a git URL to find the core, so the
 adapter's `multiset-vps>=0.1,<0.2` has to be named explicitly.
 
+The Go2 adapter needs PortAudio present first (`apt install portaudio19-dev`,
+or `brew install portaudio`): it depends on `pyaudio`, which has no Linux or
+macOS wheel. `multiset-vps` alone needs nothing.
+
 Only need poses from frames you already have? `multiset-vps` stands alone: three
 dependencies, no robot, no renderer.
 

packages/unitree-go2-dimos/README.md

@@ -65,6 +65,16 @@ reverse and never serves WebRTC. `find_robots` reports both and labels them, so
 
 ## Install
 
+On Linux, install PortAudio first. `unitree-webrtc-connect` depends on
+`pyaudio`, which publishes no Linux or macOS wheel and compiles against
+`portaudio.h` — without it the install fails partway through with a C
+compiler error:
+
+```bash
+sudo apt-get install -y portaudio19-dev     # Debian/Ubuntu, incl. Jetson
+brew install portaudio                      # macOS
+```
+
 Both packages, in one command:
 
 ```bash