Rejected CPU worker-reuse pilot

A prototype reused a bounded thread pool within each CPU registration for correspondence searches and normal neighborhoods, processed short query batches on the caller and parallelized independent normal PCA rows. Each row kept its original SciPy query arithmetic, neighbor order, covariance and LAPACK solve. Workers were joined on normal completion and task failure.

The complete public-data comparison did not demonstrate an improvement, so the runtime retains the earlier parallel-query implementation. No additional thread-pool or parallel-PCA execution path is adopted.

Scope	Previous eight-worker runtime	Reusable-worker prototype	Observed ratio
Complete worker	249.395 s	255.982 s	0.974x
Four unchanged fits, including covariance	31.456 s	26.400 s	1.192x

Both workers used the same eight-worker budget. One fresh-process fixed-order pair replayed the original public GLIM Livox bag's first six seconds with four map rounds and one BLAS thread. Host load was unreserved; this observation does not establish a systematic slowdown. Unchanged CPU fit timings also varied. The complete-worker evidence did not justify adding the prototype's complexity.

Every complete calibration/evidence field and stored initial estimate agrees exactly after excluding declared source revisions, command provenance and timestamps. Every full per-round fit field, including covariance, trajectory, costs, status and iterations, agrees exactly after excluding fit wall times. Both results remain INCONCLUSIVE: the clip supplies no independent extrinsic truth.

Both runs retain 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. Those numerical grids remain disabled by default. Whole timing includes startup, imports, bag decoding, every initial/gyro/ feedback/final odometry pass, association, fitting/covariance, evidence, hashes, saves and source verification.

The bound archive retains every full result and fit, original measurement/log, both exact source capsules and manifests, the executed launcher and schema, plus the prototype's validation test sources and original test output. Its CPU-only audit checks actual import bindings, worker budgets, complete physical fields and recomputed ratios. Raw bag data is not redistributed.

Reproduce the rejected candidate

Use the candidate revision recorded in the archive's protocol, which includes the prototype. The current runtime has returned to the previous implementation. The new cpu-source comparison keeps explicit worker budgets while evaluating two source revisions; the existing cpu-cpu mode compares budgets on identical frozen sources.

python -m tools.benchmark_bspline_lidar_imu --comparison cpu-source \
  --baseline-commit 7b367723690d867c446742d327a59433c12a705d \
  --prior-query-workers 8 --current-query-workers 8 \
  --max-seconds 6 --max-rounds 4 --repeats 1 \
  --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-worker-reuse-rejected
python -m tools.audit_bspline_lidar_imu_benchmark docs/assets/cpu_worker_reuse_rejected