Complete guarded point-plan reuse comparison

Whole-worker median: 23.324646 -> 9.522338 seconds (2.449466x). See workload and bounds. Each run retains 106 preparation requests; actual builds fall from 106 to two.

protocol.json binds both source revisions, full-point config, input hashes/counts, exact launcher, schemas and source capsules. attempts.json retains every worker. Each worker folder contains the complete result, full event/parameter trace and array hashes, measurement and both logs. summary.json is recomputed by the public auditor; index.json binds every published file. Source capsules and measured files preserve their original bytes. The measured launcher equals the public auditor.

Raw PCDs are external public inputs. The historical upstream manifest is copied verbatim; actual header counts are pinned in the protocol. This development capture remains degenerate with rank 9; this archive supports timing/parity and does not establish independent calibration accuracy.

Baseline: c517dca9b98c6bc092a184707406db4f0c56a908. Measured runtime: dfd9df76515fc36b3b252e48a6c58d605b636b93. Protocol SHA-256: e3491dbe5a3bf15ebffccff56db23859545d9ff27461fbb3cf3ef12fb473b7ff.

python -m tools.benchmark_joint_point_plan_reuse --audit-only docs/assets/joint_lidar_point_plan_reuse