Direct metric-depth calibration protocol

Compare a bounded robust metric estimator against unchanged MI on the frozen analytic scene, all five conditions and ± initial seeds plus separate reference-start controls (10 perturbed / 5 drift trials). This follows failed MI selection validation; it is adaptive development, not unseen evaluation. Do not change scene, seeds, point populations or full-pose accuracy gates.

Metric objective: transformed LiDAR optical z minus bilinearly sampled camera z, with radial images converted to optical z before interpolation. Use all original fitting points, never the held-out lattice. No reference extrinsics enter fitting. Four valid positive depth neighbors are required. Each unsupported point retains a constant 1 m residual rather than disappearing; this prevents free deletion but does not constitute a smooth/physical visibility model. Retain this limitation. No occlusion filtering or residual-based population pruning.

Use SciPy TRF least_squares, two-point finite differences, soft-L1 scale .1 m, rotation bounds ±2 degrees / translation ±1 m in the same initial-pose coordinates, x_scale=(1,1,1,.1,.1,.1), all three tolerances 1e-8, max_nfev=100. Count EVERY residual call, including numerical-Jacobian calls, with a hard 800-call ceiling; abort and retain logs on exceeding it. The effective 100-function-evaluation cap is also reported and may stop earlier. MI retains its 800-objective-call ceiling. These objectives and per-call costs differ; this is not an isolated optimizer comparison. Report actual calls and wall time. Support counting/holdout diagnostics are separate from optimization calls and included in elapsed metric runtime only as implemented.

Require optimizer success and >=90% of initial supported points (>=64) in EACH frame, plus the unchanged <0.5-degree/<0.20-m reference-scored pose gates. Retain all failed/boundary cases; convergence is not evidence of pose correctness. No reference-start rescue or threshold tuning. Evaluate the frozen held-out metric validator only after fitting and never use it to select parameters or accept a pose as ground truth. Its known false-accept behavior remains part of the evidence.

Freeze source/settings before scoring. Recompute MI and verify exact baseline reproduction. Unit tests verify metric units, subpixel interpolation and fixed missing-support population. A positive outcome remains limited to one analytic scene, and depth bias/occlusion/true-start drift must be reported before deciding whether further geometry tests or a real-data comparison is warranted.

PYTHONPATH=src:. OPENBLAS_NUM_THREADS=1 /workspace/work/p0-venv/bin/python \
  tools/evaluate_joint_metric.py docs/assets/joint_metric_20261029/plan.json \
  docs/assets/joint_metric_20261029/evaluation.json