Odometry candidate screening

All ten candidate strategies remain experimental. No runtime change is adopted. The best complete target-construction/registration screen is only 1.158x; no whole-bag improvement is inferred from these small fixed-input observations.

Candidate / scope	Previous	Candidate	Observed ratio
Stable GPU voxel reduction, raw scans	0.022896 s	0.021742 s	1.053x
Stable GPU voxel reduction, processed scans	0.016450 s	0.020034 s	0.821x
Fixed k=11 query	0.392656 s	0.387144 s	1.014x
Unrolled neighbor insertion	0.392656 s	0.375619 s	1.045x
Leaf-packed reference coordinates	0.392656 s	0.394731 s	0.995x
GPU radix tree with lazy native repairs	0.457124 s	0.394617 s	1.158x
Tree-ordered normal queries	0.377905 s	0.380340 s	0.994x
Morton-ordered source queries	0.377905 s	0.385769 s	0.980x
CPU-tree-ordered source queries	0.377905 s	0.400486 s	0.944x
Guarded GPU batches of four iterations	0.360379 s	0.371897 s	0.969x
Guarded GPU batches of eight iterations	0.360379 s	0.353789 s	1.019x

Five separate cohorts retain all 744 observations and their own baselines. They are not pooled. Four cohorts rotate the candidate order over five repeats of eight captured public GLIM registration inputs. One first-case compilation warmup per variant is outside their timing. Scope includes construction, normal preparation, native repairs and registration. The voxel cohort retains seven rounds over eight original raw scans and eight processed inputs; its six-round retained-cache medians exclude round zero explicitly, with round zero retained. Its first candidate round takes about 2.935 seconds, including on-demand operation compilation; module setup is outside this pilot's timing. There is no fresh-process/cold performance claim and no complete-bag candidate evaluation. One BLAS thread is used; host CPU/GPU activity is unreserved.

All registration decisions, iterations and correspondence counts match. Layout, radix and ordering pilots retain exactly equal numeric registration fields; batched updates use abs 1e-9 / rel 1e-6 for transforms/RMSE/eigenvalues. All batched fixed-case runs report zero full-path fallbacks. This does not validate their unexercised fallback/boundary cases. The voxel launcher asserts coordinate bit equality and preserves point counts; coordinate payloads are not archived, so its CPU-only audit checks counts/timings/source bindings, not the coordinates.

The CPU-only audit recomputes every timing aggregate and checks all observations, retained registration fields/work, source capsules, launcher/protocol hashes, scope and input identities. Raw scans and compiler caches are not redistributed. The launchers are diagnostic scripts with fixed repository-relative inputs and output directories; they are frozen as text and are not production APIs.

The separately retained cProfile run diagnoses the current pipeline. Its instrumentation and eight fixture captures add overhead; its roughly 88 second worker profile is not compared with uninstrumented bag benchmarks. Runtime source is 3b006a4be3babcf2b0e5c537993105b7d1c68b46 and bag SHA-256 is f57a2e85339316d4235ab693ab2876aac5e113625c7f8b0f7873b62f92e09643. This clip supplies no independent calibration truth.

python -m tools.audit_odometry_candidate_screen docs/assets/odometry_candidate_screen