Exact batched fixed-rig point-to-plane factor

This is an incremental comparison against Draft PR #155, on the existing examples/public_datasets/livox_horizon_horizon_pcd_sample/online_accumulation_config.yaml. It runs the normal offline calibration API on the official public Livox PCD pair. The config selects native_lidar_point_to_plane, limits target points to 4,000, and permits 40 iterations. Both methods use the same 2,732 fixed correspondences and complete nine optimizer iterations (one accepted, seven rejected, then convergence). No configuration, point budget, finite differences, correspondence, probe, save or quality threshold is removed or relaxed.

Despite the config filename, these are offline PCD runs, not online rosbag replay. This workload is different from the earlier ICP/GICP/NDT comparison; its timings and ratios must not be combined with those cohorts.

Complete process and optimizer subset

Five fresh-process pairs, alternating order, no warmup and no outlier exclusion. All ten complete canonical results match, including estimated transforms, finite-difference diagnostics, optimizer history, evidence, grades and provenance.

Metric (seconds)	Baseline median	Batched median	Ratio of medians	Baseline range	Batched range
subprocess_wall_seconds	12.383584	5.723871	2.16x	11.470166–14.037344	5.029327–6.752677
pipeline_seconds	10.011195	3.421480	2.93x	9.233168–11.354264	2.890714–4.214353
optimizer_seconds	6.369221	0.524065	12.15x	6.007813–7.586272	0.426406–0.658440

Pair	Execution order	Baseline whole process (s)	Batched whole process (s)
1	baseline → batched	14.037344	5.466299
2	batched → baseline	11.470166	5.029327
3	baseline → batched	12.732146	6.752677
4	batched → baseline	12.191389	5.723871
5	baseline → batched	12.383584	5.822800

Whole-process wall includes interpreter/import startup, config, all input reads, inspection, candidate plus all 24 known-bad controls, correspondence construction, native refinement, diagnostics, input hashing, reports, results, artifacts and worker measurement JSON writes. It excludes parent audits and CLI dispatch. The API timer is a subset starting at run_calibration. The optimizer timer wraps the existing FixedTrajectorySe3ExtrinsicSolver.solve, including its initial and final factor diagnostics; subsequent adapter diagnostics are outside that subset. No additional solve is performed to obtain the timer. The evidence-call subset is also retained in the definition, with unchanged evaluation code.

The baseline loads only the archived factor module from d017dfa03c6275924c8da44f55b2d3707ca58311. Other source files, solvers, inspection, correspondence builders, evidence and writers are shared with measured commit 9cdf1ef43d3dd46a0ae36ef8a68b74967cb56612. Both methods report the same run Git revision because they execute in this shared checkout; module bindings and exact source archives identify the substituted factor independently.

Numerical contract and quality

At 64 or more observations, cache read-only observation arrays within the factor and compute residuals in chunks of at most 4,096 observations. Replay the original Hamilton products and all four composed-rotation normalizations, including the two in point rotation. Preserve the translation composition, ordered dot products, square-root weights, residual order and numerical finite differences. Hessian/gradient accumulation, robust weighting and optimizer logic are unchanged. Small populations and nonfinite input/transform values retain the scalar path. Replacing the observation tuple refreshes the cache; changing the initial transform is observed on the next call. Retained arrays cost 17 float64 values per observation (136 bytes), plus array objects; temporaries are bounded by the chunk size.

Tests compare residual bits for static and per-point moving poses, chunk boundaries, weights, signed zeros, nonfinite and overflow cases; they compare complete numerical linearization, observability and solver history against the original scalar formula.

The optimizer still reports converged, but observability rank remains 3/6 and overall quality remains warn. Faster execution does not establish full six-DoF observability or absolute calibration accuracy. This single public capture is performance evidence; five repeats are not independent accuracy trials.

Inputs and reproducibility

Official source: Livox automatic calibration sample. Use the upstream public sample terms. Input files are not redistributed here.

• base_horizon_100432.pcd: 769,568 bytes, 23,921 finite points, SHA-256 19b3864cbee76b785d207fcb3fe48f6f506d142d8233f842d0a87fe4330fe407.
• target_horizon_100538.pcd: 761,408 bytes, 23,666 finite points, SHA-256 61a74eb8b8ff4eeba4ca0ff824a4046ae220e947e8156f27d93da452538f5b6f. The same 120 nonfinite header points are excluded by the reader in both methods.

From the measured source checkout, with the public pair downloaded:

python tools/download_public_dataset.py livox_horizon_horizon_pcd_sample
python tools/benchmark_livox_evaluation.py --workflow fixed-rig \
  --baseline-commit d017dfa03c6275924c8da44f55b2d3707ca58311 \
  --repetitions 5 --omp-threads 1 --output-dir outputs/livox-fixed-rig-rerun

Only child OMP_NUM_THREADS=1 is set; unrelated user activity and OS caches persist. No other Codex performance tests ran concurrently during the formal cohort. Host variability and small-sample medians prevent a portable speedup guarantee.

The schema-valid definition retains every trial, its complete canonical result, input/config/output/artifact hashes and module bindings. The aggregate is recomputed from that definition. Canonical results normalize only run IDs, timestamps, destination paths and Windows separators. Publication replaces private workspace prefixes only in commands and metadata; complete result strings remain byte-identical to the measurements. baseline.*.txt and measured.*.txt retain exact Git source bytes, with separate committed and working-byte hashes. The publication launcher checks each saved result schema, API/save equality and every saved artifact hash before publishing. Raw per-worker output files remain under outputs/public-livox-fixed-rig-batched-v1 in the local workspace.