Complete LiDAR–wheel evaluation preparation

The complete prepared-input evaluation phase falls from 2.389020 to 0.491396 seconds (4.861703x). The whole fresh worker, including imports and audit instrumentation, falls from 5.353725 to 3.397488 seconds (1.575789x). Startup remains substantial; the larger ratio applies to the calibration API. See the benchmark.

All 2,762 original OXTS poses and proxy samples of KITTI 2011_09_30_drive_0018_sync yield 2,760 generated sensor motions in 29 blocks. The common time origin is removed to exercise finite-difference clock steps. KITTI has no wheel odometry: OXTS vf and wu are proxies. These are generated sensor motions, without raw LiDAR registration or independent accuracy evidence. The unchanged policy is warn, including a clock control that is not detected and unobservable roll. No accuracy or GPU acceleration claim is made.

This archive retains the original packets/timestamps and download provenance, prepared control and generator, both complete core source capsules/manifests, measurement code and schemas, all six fresh runs, every residual request and least-squares result, official artifacts, and validation receipts. index.json binds every publication byte other than itself.

The audit independently reconstructs all inputs, train/holdout/jackknife subsets, all 301 scan candidates per fit, selection visits and exact clock LRU work. All results, residual digests and full Jacobian digests remain exact. Only offline motion regeneration allows platform roundoff: rtol=0, 2e-7 m/s velocity and 1e-12 rad/s angular rate. Original proxy samples, timestamps, IDs and raw bytes remain exact. Generator source comparison allows CRLF/LF differences only; its archived original bytes retain a strict hash pin.

qa/cohort-selection.json records all timings and hashes from the earlier, excluded absolute-time cohort. Its clock finite differences were unresolved; the selected cohort exercises 13–15 distinct clock states per fit. Both use the same core source. The selected cohort is the sole published speed claim.

KITTI attribution: Andreas Geiger, Philip Lenz, Christoph Stiller and Raquel Urtasun. Original data and generated derivatives retain CC BY-NC-SA 3.0: dataset and license. Calibrex source snapshots and measurement code retain Apache-2.0.

python -m tools.benchmark_lidar_wheel_preparation \
  --audit-only docs/assets/lidar_wheel_evaluation_preparation