[mbreiser]

Tooling + docs only (no Pages tree change; rig/ is the rig PCs' workspace). Companion of reiserlab/rig-nodes#8, which made the ROS 2 distribution one switch there; this PR mirrors it in rig/pixi.toml. Parent record: LAB-338 (rig-nodes extraction). The rigs' environments do not change.

What changes

• rig/pixi.toml: [feature.jazzy] (the rigs: robostack-jazzy, mutex 0.19.*, Python 3.12.*, ros-jazzy-* ros-base / rmw_zenoh / rosbag2 / demo-nodes, the two Zenoh config URIs under python3.12) and [feature.lyrical] (robostack-lyrical, mutex 0.22.*, Python 3.14.*, ros-lyrical-*, URIs under python3.14, and rig_nodes + rig_interfaces built from the rig-nodes git branch feat/ros-distro-switch, because the published channel is Jazzy-only). Workspace channels shrink to conda-forge; each feature appends its RoboStack channel with priority = -1 and the channel feature carries the rig-nodes release URL with priority = -2, so every existing environment keeps its channel order. Environments: default / typed / viz = jazzy + channel (+ typed / viz), dev = jazzy + dev, new lyrical = lyrical (own solve group).

• dev loses the ros-jazzy-* toolchain + cmake/ninja/cxx-compiler lines: pixi-build resolves each source package's build/host environment itself; the lyrical environment's build from git on this Mac lists none of them and compiles rig_interfaces fine (he8ea13f_0, python_abi 3.14.* *_cp314, mutex >=0.22.0,<0.23.0a0). The arena harness then passed 63/63 on that environment (pixi run -e lyrical test: whole graph, Studio emulator, model fly, crash phase — on Python 3.14 and rosbag2 0.33); pyzmq resolves as the abi3 build (py312… build string, imports under 3.14).

• [workspace.build-variants] ros2-distro-mutex = ["0.19.*|0.22.*"], the same line as rig-nodes' own workspace: once rig-nodes main has no manifest pin, a consumer that builds the source packages needs it or the Jazzy host solve takes mutex 0.20 (the Python 3.14 rebuild of Jazzy).

• Docs: rig/README.md (the -e lyrical line + a paragraph), a dated note in docs/development/rig-software-repos-and-runners.md (before §8), one clause in CLAUDE.md (never spell ros-jazzy-* outside a distro feature). The record itself lives in rig-nodes docs/ros-distro-switch.md.

• New CI workflow rig-graph.yml (path-filtered to rig/**): until now nothing in this repository's CI ran rig/. jazzy (blocking; ubuntu-24.04 + windows-2022) installs the committed lock's default environment and runs test-gateway, test-stimulus, test-viz and the whole-graph harness; lyrical (allowed to fail) does the same on the Lyrical environment built from git; watch (Mondays 06:17 UTC and on demand) re-solves lyrical against the live RoboStack channel in the runner only, prints which ROS packages moved, and runs the harness on it — the "keep an eye on the channel" job. The committed lock is never touched by it.

Lyrical on this Mac, beyond the harness

check-typed (8 types serialize), test-gateway 5/5, test-stimulus and test-viz pass, and the rig's own composition pixi run -e lyrical rig -- --sim --bias --bias-sim driven by the Studio emulator with --connector-cmds: every check PASS, 400 frames applied over 2 epochs, BIAS connect/startRecording/getTimestamp/stopRecording ok, fictrac p50 latency 0.52 ms (p99 2.03), clean stop on SIGINT.

Regression found on ww18 and fixed (dc2d9b2)

Moving the two ZENOH_*_CONFIG_URI lines into the distro features left the workspace-level [target.win-64.activation.env] override dead: a feature's activation env takes precedence over it, so on Windows every environment (the rigs' default included) got the literal $CONDA_PREFIX/lib/python3.x/… and rmw_zenohd could not open its config. That was the real cause of both Windows CI failures (not the 20 s router wait, which stays at 60 s anyway). Each feature now carries its own [feature.<distro>.target.win-64.activation.env]; no lock change. Found by the ww18 bench session before any merge; its Lyrical install on the rig PC took 43 s and compiled rig_interfaces with VS 2022 19.44; check-typed, test-gateway 5/5, test-stimulus 51/51, test-viz 12/12 passed there.

CI after the fix (run at dc2d9b2): jazzy windows-2022 ✓, lyrical ubuntu ✓; jazzy ubuntu 59/60 and lyrical windows 62/63, each one timing-sensitive check on a shared 2-vCPU runner (an epoch's start-frame tare one frame late; conversion fidelity |Δms| max 107 ms against a 50 ms bound). Both pass on the Mac and on ww18. The harness steps retry once; a deterministic break still fails twice. Follow-up: a --ci tolerance scale in test_ros2_poc.py / studio_emulator.py instead of the retry.

ww18, the first rig PC on Lyrical (2026-11-05, run by the ww18 session at dc2d9b2)

Install 43 s (rig_interfaces compiled by VS 2022 19.44); runtime lyrical / rmw 0.10.6 / rclpy 10.0.11 / rosbag2 0.33.4 / mutex 0.22.0 lyrical_30 / Python 3.14.10; check-typed ✓, test-gateway 5/5, test-stimulus 51/51, test-viz 12/12; arena harness 63/63 in 80 s (Jazzy baseline on the same PC 60/60 in 56 s); one real closed-loop run: FicTrac 2.1.2 on the FLIR camera at ~96 Hz (typed adapter), Production Studio, G6 on COM4 — bag 4,182 records → 6,637 converted lines identical to the live file, runlog-check 4/4, controller 0x70 = host 0x70 = 1,543, shutdown parsed 29,472 / skipped 0 / bad 0, p99 0.77 ms, Console clean; soak smokes 6/6 (simulator and real FicTrac, 35,044 frames, 0 bad, p99 0.61 ms, clean CTRL_BREAK stop); 8 h real-FicTrac soak: 48/48 emulator runs ok, graph clean exit 0 (13:04–21:04; 2,776,157 frames parsed, 0 skipped, 0 bad, fictrac p99 0.84 ms, BIAS simulator 144,261 frames 0 bad; rig.log without errors or tracebacks; adapter at 96.2–96.4 Hz to the end). Record: rigs/ww18/lyrical-bench-2026-11-05.md in arena-experiments; rig-nodes docs PR #9.

Findings adopted here: lyrical-viz environment (ros-lyrical-foxglove-bridge, same solve group as lyrical) and rig.py --foxglove choosing the bridge's environment from PIXI_ENVIRONMENT_NAME — a Jazzy foxglove_bridge cannot join a Lyrical router. Benign: Lyrical's bridge logs one Failed to retrieve parameters from node '/rig_host_monitor' on connect. Not Lyrical: 4 of 1,548 long arena_command lines are kept verbatim because the Studio truncates their head (same on Jazzy; Studio-side item).

Soak driver + lab-PC handover

• pixi run soak (rig/tests/soak_graph.py): one rig.py graph kept running for --hours (simulated FicTrac at 100 Hz or --fictrac for the real adapter, BIAS simulator + node, host monitor, gateway), a recorded run driven through the gateway by the Studio emulator every --every minutes, a rig-style stop through the process group (CTRL_BREAK on Windows), runlogs/soak-<stamp>/soak.json with every run and the gateway's accounting; refuses to start when 7447 / 8765 / 60000 are taken. Smoke on the Mac's Lyrical environment: 3/3 runs ok, 11,129 frames, 0 bad, p99 1.4 ms, clean stop.
• rig/LYRICAL-BENCH.md: the procedure for the first Lyrical session on a Windows rig PC (prerequisites incl. MSVC v143 for the source build and the C:\px build root, install, the no-hardware suites, one real Studio run, the overnight soak, where results go, cleanup).
• Bags across distros: a bag written on Lyrical (rosbag2 0.33, metadata version 9) converts with the Jazzy environment's mcap_convert (2,031 records, 1,121 frames, closed: true) and opens with Jazzy's ros2 bag info exactly as with Lyrical's. rosbag2 0.33 writes the rosbag2 MCAP metadata record (with our custom_data runtime dict) at open as well as at close, so even a crashed Lyrical bag says what it ran on. rosbag2's Reindexer keeps only ros_distro in the regenerated metadata.yaml; the MCAP record is the authority.

Lock

Diffed environment by environment against main's lock (package URLs per platform):

environment	linux-64	osx-arm64	win-64	channels
default	438 → 438, 0 removed / 0 added	399 → 399, 0 / 0	398 → 398, 0 / 0	same order
typed	440 → 440, 0 / 0	401 → 401, 0 / 0	400 → 400, 0 / 0	same order
viz	499 → 499, 0 / 0	444 → 444, 0 / 0	430 → 430, 0 / 0	same order
dev	456 → 439	433 → 400	407 → 399	rig-nodes channel no longer listed (it builds from git)
lyrical	436 (new)	399 (new)	398 (new)	conda-forge, robostack-lyrical

dev's change is the dropped toolchain shims plus conda-forge churn on a re-solve (flake8, python-dateutil, libcxx, BLAS builds); its rig_interfaces record is still cp312 + mutex >=0.19.0,<0.20.0a0 from rig-nodes main a32ee8e. pixi install --locked -e default and -e lyrical both installed from this lock on this Mac.

After rig-nodes#8 merges

Point [feature.lyrical]'s two git sources at main (and drop the note), re-lock. The dev feature already tracks main.

Tier

Tooling-only: no web tool, no Pages tree. Not a release candidate; merges on its own like the other rig/ tooling PRs (ask if it should ride the next candidate instead).

Release notes

None (nothing user-visible in the web tools).

🤖 Generated with Claude Code