Public GLIM LiDAR-IMU shared rotations

Three alternating fresh-process pairs compare direct gyro point deskew and shared spline rotation terms with the preceding analytic-Jacobian baseline. They use one BLAS thread, the original public GLIM Livox recording's first 12 seconds, four map rounds and the same 30-iteration fit caps. GPU is not required. These measured ratios compare this cohort's two implementations; ratios from separate cohorts should not be multiplied.

Scope	Analytic-Jacobian baseline	Shared rotations	Speedup
Complete worker	142.930 s	100.615 s	1.421x
Four fits, including covariance	50.106 s	33.072 s	1.515x

Trial	Variant	Whole seconds	Fit seconds
0	prior	155.435139	50.106017
0	current	100.615110	33.071726
1	current	103.203762	33.933937
1	prior	139.053514	49.604608
2	prior	142.929572	52.508813
2	current	94.802650	31.483618

Method and source binding

• Prior: 6df5aa2526b1fb4d5bc8511b7e9a0265d1266abc.
• Current measured source: 296746c8f58e2d740c119086cac62eab8d012a7e.
• Original public livox.bag SHA-256: f57a2e85339316d4235ab693ab2876aac5e113625c7f8b0f7873b62f92e09643.
• PointCloud2 /livox/points, point-time t; IMU /livox/imu, acceleration in g.
• The 12-second clip yields roughly 11.5 reliable seconds and 118 controls; the full 61-second recording is not timed here. Supplied GT is not used.
• Whole time includes process startup, reading, initial estimates and gyro deskewed odometry, association, all fits and covariance, noise update, holdout and signed controls, result saving, hashing and verified imports. Fit time includes covariance. Parent source preparation is outside timing.

Every worker retains a schema-valid complete result.json and full per-round calibration, trajectory and covariance in measurement.json. Both exact source trees and their raw-byte manifests are archived. The original worker logs, actual imported-module digests, executed comparison launcher, tool-only schema and publisher are included. The publication index binds every file except itself. Raw recording data is not redistributed.

Quality and limits

All three result comparisons pass. Policy, block, control, separability and convergence decisions agree; estimates, std, last-round changes, noise, chi-square and correlations meet the fixed protocol tolerances. Across the three pairs, the maximum rotation difference is 2.81e-10 deg, lever-arm difference 6.23e-12 m and clock-offset difference 4.68e-13 s. The maximum whitened holdout RMS difference is 2.24e-12. The independent archive audit additionally checks full per-round covariance, trajectory and costs.

All six final results remain INCONCLUSIVE. The capture's estimated plane map and its formal covariance provide no independent absolute extrinsic truth. This is a performance/regression comparison for one clip, not an accuracy or observability claim. Speed depends on point density, control count, convergence and initial odometry. Exploratory profiles experienced changing background load and their wall-time ratios are excluded. The single formal cohort above supplies all reported speed ratios.

Audit and reproduce

python -m tools.audit_bspline_lidar_imu_benchmark docs/assets/lidar_imu_shared_rotations
python -m tools.benchmark_bspline_lidar_imu \
  --baseline-commit 6df5aa2526b1fb4d5bc8511b7e9a0265d1266abc \
  --bag data/public/glim_versatile/livox.bag \
  --max-seconds 12 --max-rounds 4 --repeats 3 \
  --output outputs/glim-shared-rotations-comparison

The output directory must be new and runtime changes committed. The archive audit uses CPU dependencies only. See the implementation and evaluation.