CUDA spline plane fit comparison

Scope	NumPy planes	CUDA planes	Observed ratio
Complete worker	65.884 s	61.146 s	1.077x
Four fits, including covariance	10.239 s	3.506 s	2.921x

Two alternating fresh-process pairs replay the original public GLIM Livox bag's first six seconds, four map rounds, one BLAS thread and unchanged CUDA scan odometry. Every worker starts with a separate empty CuPy compiler cache. Initialization, compilation, all odometry passes, association, noise updates, all fits/covariance, holdout/controls, input hashing, saves and frozen-import verification remain inside whole time. Fit time includes evaluator construction, first plane kernel compilation and covariance. No warmup is discarded.

Baseline c3845d1eb42326e6da11a235b5e6239c33448160, measured candidate f4c8e5662a43f5e5ce24bb4181924c6ca5b3bb68. Bag SHA-256: f57a2e85339316d4235ab693ab2876aac5e113625c7f8b0f7873b62f92e09643. The dense odometry policy is 0.03 m voxels, three local-map scans, 1 nm coordinate quantization, 1 micrometre translation and 0.1 microradian rotation-component quantization. These grids remain off by default. Plane/IMU data, return budgets, convergence checks and evidence gates are unchanged. NumPy is still the default fitter plane backend.

The GPU computes FP64 spline plane residuals, analytic local derivatives and the seven-block control normal band. Each band entry has one owner and a fixed reduction order; there are no floating-point atomic reductions. The core keeps IMU residuals/derivatives, clock finite differences, LM acceptance/stopping, the solve, marginal covariance and all evidence evaluation on CPU.

The bound audit checks every result, all full per-round fits and real kernel identity/work against the frozen source and reported plane counts. Fit status, iterations and dimensions must match exactly. Trajectories/calibration/costs use abs 1e-9 / rel 1e-6, covariance abs 1e-11 / rel 1e-5, alongside the existing whole-result physical/evidence tolerances. This GPU path permits roundoff; it does not claim bit equality. The final verdict stays INCONCLUSIVE; the clip has no independent extrinsic truth. Host load is unreserved. Ratios describe this workload and these pairs, and are not multiplied across stages/cohorts.

All four full results, measurements/logs, deterministic source capsules, manifests, launcher/schema, publisher and source-bound correctness run are retained. Raw bag data and compiler caches are not redistributed.