CPU frame parallelism on public KITTI raw

Four schema-valid JSON files preserve evidence from serial NumPy and a two-thread NumPy candidate on 2011_09_30_drive_0018_sync/image_02, original Velodyne scans, and frozen FeatDepth provider SHA256 7e3ef953a360a7c01ed0cdc3cf947db97fb0aff5912ced805d80bebe210d3360. The same 25 frames contain 3,130,279 points: 20 fit and five holdout, fixed before optimization and histogram-range selection. The declared prior adds local x rotation 0.5 degrees and camera-frame x translation 0.5 m to the dataset reference. Search bounds, steps, and the 800-evaluation budget use defaults. The reference is never the search target.

Each method was invoked three times from the same problem bytes and prior. Serial runs first on repetitions one and three; the candidate runs first on repetition two. There is no separate warm-up and every invocation is retained. The candidate reuses one two-thread pool across search and initial/final holdout evaluation. Frame calculations run concurrently and results are reduced in input order. Thread startup is included in solve time.

Repetition	Execution order	Serial solve (s)	Two-thread solve (s)
1	Serial, candidate	64.057565	60.098134
2	Candidate, serial	72.149458	54.548826
3	Serial, candidate	70.177615	42.316660
Median		70.177615	54.548826
Range		64.057565–72.149458	42.316660–60.098134

The ratio of solve medians is 1.2865x. Pipeline phase medians are 73.583689 / 57.475266 s (1.2803x), including input verification/loading, backend preparation, solve, and holdout/pool shutdown. Shared demo setup, result serialization, and HTML rendering are excluded. Measurements ran on an i7-9750H, Windows, Python 3.12.10, and NumPy 2.3.4, without isolating other running work or changing numerical library thread settings. Variation is visible; three repetitions on one workload do not establish a portable speed guarantee or statistical confidence. The artifact's bootstrap fields are schema summaries and must not be interpreted as confidence here.

All six runs converged after exactly 229 evaluations. Complete solver records, metrics, final poses, skipped frames, and initial/final holdout evaluations match exactly, including the earlier serial/CUDA x-positive trace. The canonical complete numerical payload SHA256 is fdf4d36e8cbe9b8a88a5cb765f4c387f67fa2021df8e51271ce461bb4f9e3160. Final reference differences are 0.390923 degrees / 0.160078 m and holdout MI is 1.293495 nat. Convergence and CPU parity do not prove calibration accuracy: quality remains warn, covariance and observability are unknown, raw scans are not motion compensated, and FeatDepth may have KITTI training overlap.

comparison.definition.json and comparison.benchmark.json retain all six attempts, per-invocation timings, source/inputs/configuration, and original external standard-result SHA256 bindings. serial.result.json and parallel.result.json retain the first pair's complete traces and holdout records; subsequent complete numerical payloads were checked against this pair, with the shared payload digest recorded in every benchmark trial. Dataset/workspace paths are replaced with ${DATASET_ROOT} and ${WORKSPACE_ROOT}. Original result hashes bind original external bytes; published representative files have separate hashes in source_artifacts and retain their original-result hash under run.provenance.publication. Raw data, depth weights, generated HTML, and images remain outside Git.

The measured implementation is commit 5809c30. After measurement the report's CSS was updated to wrap long checkpoint hashes on mobile; the original HTML and all original numerical JSON remain in outputs/public-kitti-six-dof-cpu-frames-v1. HTML was regenerated only for presentation and checked at desktop/mobile widths, including expanded provenance, nine decoded images, and nine SVG overlays.

Reproduce with:

calibrex demo kitti-six-dof \
  /path/to/2011_09_30_drive_0018_sync /path/to/featdepth-provider.json \
  --candidate-backend numpy --candidate-frame-workers 2 --repetitions 3 \
  --output-dir outputs/kitti-six-dof-cpu-frames --json