Rejected fused CPU scan covariance

An optional Numba candidate fused the target-normal gather, mean and six distinct FP64 covariance entries while retaining SciPy's neighbour ordering, native NumPy eigh and planarity checks. It avoided the gathered and centered (rows, neighbours, 3) arrays, compiled without fast math or parallel execution, and reported actual compiled/fallback work and adapter source identity.

The complete public-bag experiment did not demonstrate an improvement, so the candidate is not adopted. Current runtime under src/calibrex remains the native implementation from d612581, identical to 7b367723. The candidate source remains in commit 2d6e5caef559b4383bf4da44739bc9dc0146e619 and the bound source capsules; its CLI/API additions are not available on the final branch.

Complete public-data comparison

Scope	NumPy covariance	Fused Numba covariance	Observed ratio
Complete worker, including first compilation	165.422 s	177.744 s	0.931x
Lazy normal preparation, including neighbourhood queries	59.828 s	51.136 s	1.170x
Four unchanged fits, including covariance	14.396 s	18.932 s	0.760x

Two fresh-process pairs alternate NumPy/Numba then Numba/NumPy on an unreserved host. Both variants use identical frozen sources, eight CPU query workers and one BLAS thread. Every compiled worker starts with an empty, separate Numba cache; imports and first compilation belong to complete-worker timing. The fitter is unchanged, and its wall time also varied. These observations do not establish a stable slowdown ratio. The local preparation improvement was insufficient to justify an additional runtime/backend option.

The workload is the original public GLIM Livox bag's first six seconds with four map rounds, 0.03 m voxels and three local-map scans. Both variants explicitly use a 1 nm coordinate grid and returned-motion grids of 1 micrometre translation and 0.1 microradian rotation-vector components. These grids remain off by default. The candidate's tests also compare complete unquantized trajectories and registration quality exactly, including tied neighborhoods and gap scaling.

The bound archive retains all four complete results, every full per-round fit, original measurements/logs, both source capsules/manifests, launcher, schema, publisher and the correctness run with bound test sources. The CPU-only audit checks declared adapter choices, source identities, actual work and empty compilation caches before comparing all physical/evidence fields and fitted arrays exactly. Only independently validated execution metadata and wall times are excluded.

The candidate's retained correctness run passed 92 tests. Its receipt binds the source revision, source manifest, exact test sources, launcher and original log.

Both pairs preserve every complete physical/evidence field, including the stored initial estimate, and every complete fit record exactly. All final verdicts remain INCONCLUSIVE; this clip supplies no independent extrinsic truth. This comparison concerns dense CPU odometry and does not establish GPU gains or performance on default coarse scans.

To reproduce, use the candidate commit above in a separate checkout, install its numba extra, and use a new output directory:

python -m tools.benchmark_bspline_lidar_imu --comparison cpu-cpu \
  --prior-query-workers 8 --current-query-workers 8 \
  --prior-normal-backend numpy --current-normal-backend numba \
  --max-seconds 6 --max-rounds 4 --repeats 2 \
  --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-cpu-normal-covariance-pairs

The final branch can audit the archive without a GPU, JIT compiler or bag:

python -m tools.audit_bspline_lidar_imu_benchmark docs/assets/cpu_fused_scan_covariance