[mbreiser]

Summary

This fixes the problems found when PR #252's two Foxglove dashboards were checked on real
hardware runs from the ww18 Windows rig. It merges into PR #252's branch, not main.

The test data is on branch data/ros2-poc-ww18-2026-10-15. For each of five runs, each bag was
played back at 1× through viz_node, and the outputs were recorded and moved back onto the run's
own timeline. The data branch README explains how.

• No more dropped frames. viz_node's /arena/frame subscription now uses RELIABLE_QOS
  (depth 200) like its other inputs. With depth 10 it dropped 15 of 9,969 frames (run 5.2), 17 of
  58,918 (10-min soak) and 8 of 12,226 (5c). Now it drops none.
• The display's rotation, not just its index. New /viz/arena/display_deg (DisplayAngle)
  unwraps the bridge's wrapped index: the sum of index steps, each taken the short way round the
  modulus, × frame direction × degrees per frame. The modulus comes from --frames N, or is
  learned per epoch as the largest index + 1.
  • With the 20-frame grating used on ww18, azimuth_deg stayed inside one 36° period while the
    fly turned thousands of degrees.
  • Now display − turn − bias is constant to within ±0.9° (half a frame) for 98% of samples in
    every closed-loop epoch of runs 5.2, 5.7, the soak and 5c.
  • The 3D feature bar sits at wrap180(display_deg). That is unchanged for a pattern spanning the
    arena, and for a grating the bar now turns with the pattern.
• Rates in one unit. New /viz/fictrac/fwd_dps and side_dps are the rates of FicTrac's
  integrated cols 20/21, in the °/s of speed_dps. They are computed per frame-counter step, so a
  dropped frame doesn't double them.
• arena-fictrac.json:
  • The closed-loop plot shows fly turn vs display rotation.
  • The bias angle has its own plot.
  • Kinematics shows speed, forward and side in °/s.
  • The FicTrac indicator gets a short label and a panel title.
• arena-bias.json:
  • The camera frame count and the displayed index are on separate plots.
  • The BIAS / Camera / Video indicators get short labels and are stacked full-width (the old labels were clipped).
  • The ROS log panels are titled live-only, because the bags don't record /rosout.
• tests/test_viz_node.py (pixi run test-viz): 11 checks of the new math, with no ROS graph.
• README: the Foxglove section is updated.

Testing

• tests/test_viz_node.py: 11/11 pass.
• Played five ww18 bags (runs 5.2, 5.7 crash, soak, 5c, B2) back at 1× through the new node:
  • /viz/arena/index has every frame in every run, with exact original times (|Δt| = 0) and 0 value mismatches.
  • display − turn − bias is constant to within ±0.9° (p1–p99) per closed-loop epoch.
• tools/check_layouts.py (on the data branch) on both layouts × five runs: every data panel
  resolves with data on the four FicTrac runs. The empty ones are expected: last error (nothing
  failed), the ROS log (live only), and FicTrac panels on the camera-only B2 run.
• Not run: tests/test_ros2_poc.py. It doesn't exercise viz_node, and on Windows this branch's
  pixi.toml still has the $PIXI_PROJECT_ROOT activation problem that PR LAB-303: ROS 2 bridge graph + MCAP record, ScanImage/BIAS connectors, Studio v0.97 (candidate 2026-10-20) #254 fixed.
• Foxglove (web 3.4) on run 5.2, both layouts imported: the display rotation follows the fly's turn curve for curve; bias, kinematics (°/s), camera frames and index each read on their own axis; all panel titles show. A second commit (96b6dc5) stacks the arena-bias lights full-width and makes the BIAS light read online / offline, because "reachable" and the "Camera fps" title still clipped at a 1,316 px window.

Tier

None. This targets the spike branch claude/ecstatic-noether-art3qx (PR #252), not main.

Release notes

none (nothing ships from this PR)

🤖 Generated with Claude Code