Resident LiDAR geometry evidence

This archive retains all nine workers of one predeclared, fresh, cyclic production cohort. CPU, ordinary CUDA and resident CUDA use the same six seconds of the spent public GLIM Livox bag, four map/fit rounds, one BLAS thread, density and explicit numerical grids. First initialization/compilation, decoding, all eight odometry passes, covariance, holdout/controls, saving and import verification enter whole-process wall time.

Protocol and all three triads are byte-exact copies of the executed records. The source capsules have complete manifests; correctness capsules bind the actual 47 GPU and 92 host tests, permitting only CRLF/LF normalization when comparing the tested working files to the measured Git archive. No measured source was edited.

The CPU-only auditor recomputes physical/evidence gates, every full fit, all ordinary/resident GPU result and fit float bits, every placement digest, both speed ratios, actual kernel identities, work, caches, imports and all payload checksums. Run from the repository root:

python -m tools.audit_resident_lidar_geometry docs/assets/resident_lidar_geometry --check

Read the performance page for observations and limits. Host resources were unreserved. The bag supplies no independent extrinsic ground truth; the final verdict remains INCONCLUSIVE. No default-density/zero-grid or cross-device speed claim is made.