Batched probabilistic Camera–LiDAR pose evaluation

Scope	Previous (s)	Current (s)	Observed ratio
Complete worker	29.919726	10.352140	2.890x
Artifact-backed API, including schema validation	21.964982	2.005127	10.954x

Repeat	Variant	Complete seconds
0	prior	30.445932
0	current	10.453279
1	current	10.251000
1	prior	29.393520

Two alternating fresh-process pairs, one BLAS thread, no discarded warmup. Whole time includes Python imports, artifact parsing, invariant capture, every pose candidate, disjoint holdout, acceptance, full trace, saving, schema validation and frozen-source verification. The host was unreserved; the two complete pair ratios are 2.913x and 2.867x. This four-worker sample establishes no portable timing guarantee. Ratios apply to probabilistic refinement only, not raw-sensor feature extraction or the full calibration stack.

Nine original KITTI raw clouds from 2011_09_30_drive_0018_sync provide 1,024 deterministically selected points per frame, 9,216 total. A declared generated camera pose projects them through pinhole, MEI and double-sphere models, with fixed 0.35 px noise and anisotropic covariance. These are generated execution controls, not independently measured image correspondences or calibration accuracy evidence. Seven frames (7,168 correspondences) train; two (2,048) hold out. Both implementations converge and accept the candidate after exactly 325 evaluations. No frame, point, candidate, iteration or holdout is removed by the optimization. Raw KITTI coordinates and generated correspondence files remain local; the archived builder and raw input hashes reproduce the fixture.

All complete result fields match exactly, including every trace objective, candidate/selected transforms, acceptance and holdout. Only creation timestamps and source commit provenance differ. The declared quaternion normalization and Hamilton arithmetic order, camera projection powers, reductions and confidence gates remain unchanged. Tiny frames retain scalar evaluation. Scalar fallback semantics are retained above the bounded capture budget. Captures belong to one call and never enter a shared cache.

Prior runtime 2c82d9ef625673d9c236320af75b60b7dab6765f; candidate runtime 0ea59903716ab2e269abf8f7d9e8f93444137fb7. Source capsules/manifests, complete outputs, launcher, source-freezer helper, input hashes, schemas, publisher and 73-test receipt are bound by SHA-256. The original scalar fallback body is also checked against the prior evaluator. The earlier dense-only candidate cohort is retained in earlier-dense-v2 and is not pooled with this final-runtime comparison. The failed schema-kind validation attempt remains in earlier-dense-v2/failed-validation-v1, outside successful timing medians. Audit with:

python -m tools.audit_probabilistic_refinement_benchmark docs/assets/probabilistic_pose_batching