Reusing fixed-trajectory holdout evaluations

Joint LiDAR-IMU calibration scores 32 signed known-bad controls, plus baseline gyro and accelerometer chi-square values. Previously, each of those 34 scores prepared the trajectory and evaluated all three residual families, including the unused LiDAR planes. Only one family's sum of squares was retained.

The evidence stage now uses a private fixed-trajectory residual session. It evaluates only the requested family and shares prepared spline states with the preceding holdout RMS evaluation. Calibration perturbations still recompute the requested residual, while clock perturbations prepare the original shifted timestamps and segment selections. Bit-identical gyro and accelerometer timestamp arrays share their state recursion; unequal arrays remain separate.

The session reuses the fitter's 64 MiB retained-array bound and eviction rules. Train and holdout have independent sessions. Inputs stay fixed during evidence evaluation, and the session ends with that evaluation. Public residual calls still create fresh sessions and observe measurement edits between calls. Public APIs, result schemas, samples, summation order, control magnitudes and decision thresholds are unchanged. This change runs on CPU.

Complete public-data comparison

Scope	Previous	Reused holdout evaluations	Observed ratio
Complete worker	82.051 s	83.963 s	0.977x
Signed controls after holdout RMS	2.073258 s	0.013284 s	156.076x
Unchanged four CPU fits, including covariance	12.768 s	13.567 s	0.941x

The complete worker is 2.3% slower in these observations, despite removing about 2.06 seconds from the signed-control scope. Unchanged fit and GPU normal preparation timings also vary. These two pairs do not isolate the cause of the whole-worker difference, and establish no full-calibration speed gain. The implementation is retained for its reduced CPU evidence work; this cohort is not counted as successful end-to-end acceleration.

Two alternating fresh-process pairs replay the original public GLIM Livox bag's first six seconds with four map rounds. Both variants use the adopted CUDA scan-odometry path, one BLAS thread and separate initially empty CuPy compilation caches. The host's background load is unreserved. Whole time includes all preparation, imports, decoding, odometry, fits/covariance, evidence, hashes, saves and source verification.

The control timer starts after holdout RMS evaluation, including its existing baseline calls and all signed scores. Prepared states retained from that RMS are already available to the current implementation; this timing is a stage measurement, not a cold standalone score or a complete-worker speedup.

The dense odometry policy remains 0.03 m voxels and three local-map scans, with an opt-in 1 nm coordinate grid and returned-motion grids of 1 micrometre translation and 0.1 microradian rotation-vector components. These grids remain disabled by default. Both implementations execute the same GPU kernels.

The bound archive retains all four complete results, full fit records, measurements/logs, source capsules, launcher, schema and publisher. --require-exact-results checks every physical and evidence field and execution report after excluding only declared run metadata. Every fit record must also match exactly, excluding wall time. The archive audit independently checks equality and the 34 scored calls.

The real-data verdict remains INCONCLUSIVE and this clip supplies no independent extrinsic truth. Stage and complete-worker ratios have different scopes; neither is multiplied by observations from other cohorts.

python -m tools.benchmark_bspline_lidar_imu --comparison cuda-cuda \
  --baseline-commit e59600c1354dc14f0b5afd06f92d90bf61d56f41 \
  --require-exact-results --max-seconds 6 --max-rounds 4 --repeats 2 \
  --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-holdout-evaluation-reuse-pairs
python -m tools.audit_bspline_lidar_imu_benchmark docs/assets/holdout_evaluation_reuse

Use committed runtime sources and a fresh output directory. Keep the launcher, comparison schema and repository HEAD fixed until all workers finish.

Local validation found a separate, existing screw-linear synthetic recovery test returning max_iterations instead of converged. Frozen baseline sources reproduce the same assertion failure; its source-bound validation receipt/log are retained in the archive. The B-spline evidence tests and both synthetic bag calibration tests pass. No assertion or solver setting was relaxed.