Typed protobuf wire for the MQTT messages

[gonzalocasas]

Stacked on #25. Retarget to restructure once that merges.

This is a breaking wire change. A publisher on the new wire cannot be read by a subscriber still on JSON. Every participant in a project has to move together.

COMPAS XR owns its MQTT message contract, so it owns the .proto for it. Until now the messages travelled as COMPAS JSON with no schema, and the C# side had no counterpart at all — the two stacks could exchange geometry but not messages.

The schema

compas_xr.proto defines the coordination Header, an explicit JointTrajectory, and the six trajectory request/approval messages.

The trajectory replaces the untyped "dict of joint names and joint values". Modelling it explicitly means both languages agree on the shape without either depending on compas_fab:

message JointTrajectoryPoint {
  repeated string joint_names = 1;
  repeated double joint_values = 2;
  optional double time_from_start = 3;
}

The robot base frame travels as compas_pb.data.FrameData rather than a raw dict.

Worth a second opinion: the trajectory shape was derived from the docstring, since result.trajectory arrives from whichever planner a project plugs in. trajectory_to_pb accepts three shapes — a compas_fab-style object with .points (matched structurally, not imported), a list of mappings, and a column-major {joint: [values]} dict — and deserialization always returns the list-of-points form. If real planner output carries velocities or accelerations, the schema should grow before this ships.

Registration

Conversions register through the compas_pb.plugins entry point, the way antikythera does. compas_pb discovers and imports the module for its side effects, so nothing imports it explicitly. Verified in a process that never mentions the module.

compas_eve has no global codec hook, so the Grasshopper components pass ProtobufMessageCodec() explicitly, which keeps the wire visible at the call site rather than hidden behind a type.

Distribution

Follows the compas_pb architecture guidance: the release workflow publishes the .proto bundle and per-language bindings as release assets, so downstream runtimes pin a released version instead of scraping this repository at some ref. Generated *_pb2 files are gitignored; run_prebuild generates them in CI before anything imports compas_xr.mqtt.

Also refreshes uv.lock, which still described the pre-restructure dependency set — it listed neither compas-model nor compas-pb and resolved compas-timber to 0.7.0, a version without compas_timber.model. A uv sync against it produced an environment that could not import the package under test.

Verification

Over a live broker, and against the C# runtime in compas-net#44: all six message types round-trip in both directions with byte-identical encodings, including the absent-optionals case. 25/25 tests pass, lint clean.

🤖 Generated with Claude Code

[jckenny59]

Thank you for the major effort here in the proto conversions! 😄

After the small fixes from review comments, this is basically LGTM from my side. 👍

On the “second opinion” point about trajectory shape: I’m not sure I have a strong enough thought to give you a great second opinon. Maybe we should either (a) validate quickly with a couple of real planner outputs, or (b) discuss in the next meeting before we lock this in.

But I am good with whatever, I usually default to you in these scenarios so I am for sure open! 😄

[jckenny59]

LGTM 👍 Thanks @gonzalocasas enjoy vacation. 😄