[mbreiser]

Summary

Decision doc docs/development/fictrac-columns-in-the-bag.md for LAB-303 (ROS 2 bridge, MCAP as the record). Bags will be pushed to the GitHub data repo like the .jsonl.gz is today (Michael, 2026-10-18), so the bag has to stay near the JSONL's size. The doc:

1. tables all 25 FicTrac output columns (meaning, units, which reader uses them, derivable or not), from FicTrac's doc/data_header.txt and src/Trackball.cpp;
2. records the decision (Michael, 2026-10-18): keep 10 columns (the behavior row's 1, 15, 16, 17, 22 + the error score 5, the lab-frame delta rotation 6–8 and the sequence counter 23) in one per-frame /arena/frame record, float64, one MCAP message per frame, with /fictrac/frame and /fictrac/raw no longer recorded by default (live topics unchanged);
3. measures that on the two 2026-10-14 bench bags (chunk-only index, zstd 19): 2.24× / 2.48× the .jsonl.gz, about 5 MB per 30 minute run at 50 Hz against 2.3 MB today (as recorded now 10.2× / 30.4×; all 25 columns 4.5× / 5.5×). Two levers were measured and declined for simplicity, available later: 1 s batching (1.94× / 2.14×) and int32 fixed-point at the JSONL's 5-decimal precision (≈ 1.5×);
4. lists the code changes for the implementation PR (ArenaFrame.msg, bridge_node.py, _rows.py, mcap_convert.py + full level, recorder.py/mcap_index.py tuning, tests) without implementing them;
5. ends with a proposed Linear comment for LAB-303 (text only, not applied).

Audit finding worth stating plainly: only the behavior row's 5 columns are read by anything today (bridge, kinematics.js, replay, dashboard, scripts, converter); the 25 are used only by mcap_convert --level full, whose own readers use cols 1/15–17/22 of it.

The doc references ros2-poc/ files that live on PR #254's branch (candidate 2026-10-20), not yet on main. No recorder code changes here; the sizing scripts are in the session notes and are listed for porting into ros2-poc/sizing/.

Tier

docs/skills-only.

Release notes

none

Test plan

• Measurements reproduced from one script applied identically to every table row (both bags)
• Derivability claims checked numerically on bag 1 (col 24 = Δcol 22 to 0.0; col 19 = hypot(Δx, Δy) to 1.6e-6; col 23 = col 1)
• Michael decided: 10 columns, one record per frame, float64, no batching (2026-10-18)

🤖 Generated with Claude Code