Fused CUDA normal preparation

The CUDA scan-odometry adapter prepares complete dense targets without deduplicating every row on the GPU. Once all normals are ready, subsequent deskew rounds reuse the completion state. Sparse sources still compute only newly hit normals; completing a partially prepared target preserves cached rows.

A dedicated FP64 kernel computes each neighborhood's covariance, its lowest symmetric eigenspace, planarity and repair flags together. It scales the 3×3 matrix, solves its characteristic roots and chooses the largest row-cross-product for the normal. This is an independent implementation, without a new library dependency. Eigen's direct-solver documentation notes the accuracy tradeoff of direct roots; no Eigen code is copied into the core.

Native PCA repairs retain spatial-boundary membership, small eigengaps and classification boundaries. An eigenvector-residual guard catches inaccurate directions. The classification guard is more conservative near repeated upper roots. The existing nearest-neighbor policy, FP64 distance arithmetic, disabled FMA and coordinate/returned-motion settings are retained. No return or neighbor budget is reduced. Actual kernel calls, normal rows and native repairs are saved.

Complete previous/current CUDA comparison

Scope	Previous CUDA median	Fused-normal CUDA median	Speedup
Complete worker	116.955 s	90.017 s	1.299x
Normal preparation within the worker	32.733 s	8.693 s	3.766x
Four unchanged CPU fits, including covariance	20.671 s	17.056 s	1.212x

All three complete comparisons have zero rotation, lever-arm, clock, holdout and decision difference. The existing audit also checks every per-round covariance, trajectory, cost and convergence result.

Three alternating fresh previous/current CUDA pairs replay the first six seconds of the original public GLIM Livox bag with four map rounds and complete initial, gyro, feedback and final-segment odometry. The previous source is the spatial-pruning implementation. Both workers use one BLAS thread and an isolated initially empty CuPy compilation cache; runtime and driver caches may persist. Host background load is not reserved.

Both methods use the same explicit dense policy: 0.03 m voxels, three local-map scans, a 1 nm coordinate grid and returned-motion grids of 1 micrometre translation and 0.1 microradian rotation-vector components. CPU remains the default and all grids remain disabled by default. The B-spline fitter, map association, covariance and evidence evaluation remain on CPU.

Whole time includes imports, decoding, first GPU setup and compilation, every odometry pass, association, fits/covariance, noise updates, holdout, signed controls, hashes, saves and verified imports. Normal-preparation time records outermost preparation calls, including repair and cache-check work, without adding GPU synchronization. It is a part of whole time; the two ratios are not multiplied. First compilation belongs to both observed scopes.

The bound archive retains every complete result, full per-round covariance and trajectory, original log, exact source archive, import/kernel identity, launcher and schema. A CPU-only audit checks every comparison and the actual work. All final real-data verdicts remain INCONCLUSIVE: this clip supplies no independent extrinsic truth. The speed observation belongs to this explicit dense workload, not default coarse scans.

Reproduce

python -m tools.benchmark_bspline_lidar_imu --comparison cuda-cuda \
  --baseline-commit 3f070af51921f60451c28b192ea815cc3d466025 \
  --max-seconds 6 --max-rounds 4 --repeats 3 \
  --odometry-voxel-size-m .03 --odometry-local-map-scans 3 \
  --odometry-geometry-quantum-m 1e-9 \
  --odometry-motion-translation-quantum-m 1e-6 \
  --odometry-motion-rotation-quantum-rad 1e-7 \
  --output outputs/glim-cuda-normal-comparison
python -m tools.audit_bspline_lidar_imu_benchmark docs/assets/cuda_normal_preparation

Use a new output directory and committed runtime sources. The existing dense calibrex imu-lidar joint --odometry-backend cuda CLI settings apply without an additional switch. Correctness tests include rotated/scaled scenes, repeated-root classification boundaries, thin/tied neighborhoods, partial-cache completion and a target crossing the 65,536-row chunk boundary.