Fixed-budget joint multistart diagnostic

Freeze before running the unchanged known-geometry cases. Compare the production single-start solver (800 calls maximum) against five fixed starts with 160 calls maximum each. Both use exactly the same MI, histogram ranges, step sizes, stopping tolerance and bounds in coordinates relative to the original initial pose: rotation ±2 degrees per axis, translation ±1 m per axis. Multistart uses an experimental global-coordinate pattern-search adapter; it does not shift bounds around each restart. Actual evaluations/time are reported; equal ceilings do not mean equal used computation. No unused budget redistribution.

Starts are zero plus tetrahedral rotation offsets (degrees): (.6,.6,.6), (.6,-.6,-.6), (-.6,.6,-.6), (-.6,-.6,.6); all translation offsets zero. They depend only on the provided initial pose. Select highest MI across all five outputs, using the earlier start on ties, regardless of convergence/accuracy. A selected unfinished run is a failed recovery even if its pose meets the accuracy gates. Do not choose the best ground-truth pose or omit failed starts.

Use all five conditions × two perturbed seeds plus reference start: 10 recovery trials and five separately reported drift controls, 15 baseline fits and 75 restart fits. Same <0.5-degree/<0.20-m gates, including convergence. Retain all traces, counts, times, selected index, and outputs. Recomputed baseline must match the previous study. No real-data rerun or default promotion unless this control improves recovery without unacceptable drift; no tuning based on the results. Unit controls verify budget, global bounds, best-value retention and flat-objective termination. This compares an allocation strategy under a fixed maximum budget, not a proof about all multistart or global optimization methods.

PYTHONPATH=src:. OPENBLAS_NUM_THREADS=1 /workspace/work/p0-venv/bin/python \
  tools/joint_multistart_control.py docs/assets/joint_multistart_20261029/plan.json \
  docs/assets/joint_multistart_20261029/evaluation.json