Fused CUDA map registration: cold compiler cache

Scope	Previous (s)	Fused registration (s)	Ratio
Complete fresh process	19.201556	13.253226	1.449x
Complete map evaluation API	12.900916	7.182622	1.796x

All 6 processes are retained, including first initialization and compilation. The cold cohort has 3 alternating pairs. Cold mode starts each worker with a separate empty cache. Retained mode starts each method's cache empty and retains it across its subsequent fresh processes. No warmup, timing exclusions or discarded observations enter these medians.

Every original byte is preserved, including operator input paths in provenance. The prior revision is 22c19ebc75ec61715c83a2e77e9c3e8eea34dffc; the current revision is 8fcc8d63052908e8fb06c0153a2e975acfb0dc9f. Both source trees, actual-import hashes, exact launchers/schema, complete observations/results, timings and cache-file hashes are bound by publication-index.json. Raw scans and compiler binaries are external. The measurements use an unreserved host and are descriptive, with no portable speed guarantee. This is the existing copied-sensor KITTI execution control, not independent two-LiDAR calibration accuracy or bag decoding/odometry timing.

See scope and reproduction.

python -m tools.audit_cuda_map_preparation docs/assets/cuda_map_registration_fusion/cold