Solid-state LiDAR v0.4 train-only candidate selection

This is a candidate-selection note, not the final v0.4 benchmark. The selection used public AgRob Modular-e data only and read train diagnostics; temporal holdout metrics were not used to choose the setting.

Frozen candidate

The current adaptive_mad profile uses continuous_time_outlier_mad_scale=3.5. The v0.4 candidate fixes the same profile at 2.5, with the existing minimum_threshold_m=0.02 m, support floor, adaptive voxel policy, solver budget, capture windows, and sampling policy unchanged.

The diagnostics schema records the evidence used for this decision:

slac.continuous_time_lidar_train_diagnostics/v0.1
selection_metric=train_rmse_m
holdout_used_for_selection=false

AgRob train-only probes

Each row uses seed 0 and a different temporal train prefix. The baseline values are the v0.3 adaptive_mad artifacts; candidate values were replayed with the v0.4 setting. Lower train RMSE/P95 is preferred, while rank and support are retained as constraints.

Train prefix	Baseline train RMSE (m)	Candidate train RMSE (m)	Candidate P95 abs (m)	Candidate kept / rejected	Rank
early (0.60)	0.3458	0.3010	0.5861	494 / 47	6
middle (0.70)	0.3306	0.2836	0.5950	464 / 56	6
late (0.80)	0.2747	0.2064	0.4476	436 / 94	6

The candidate is therefore frozen before the v0.4 holdout matrix. Its performance on TIERS and the GLIM identity control remains an evaluation question; those datasets are not used to tune this setting.

The reproducible declaration is examples/public_datasets/solid_state_cross_dataset_benchmark_v04.yaml.

The holdout matrix is now complete; see the v0.4 benchmark report for the independent evaluation result.