Add an independent IK baseline and restart ablation

[Dylan-Gallagher]

Summary

• add a predeclared, deterministic comparison of arm-lab IK with independently maintained k 0.32.0 on the existing 1,000-target full-pose cohort
• retain all 1,000 target rows and all 4,000 solver-result rows, with every success rechecked through arm-lab FK under one pose-and-limit rule
• separate cold-start performance from eight fixed restart starts and replay every artifact in CI
• correct the README's old description of the 97.8% result: it is home plus up to eight seeded restarts, not a cold-start result

Results

Variant	Common FK-checked success
arm-lab, home only	713/1,000 (71.3%)
arm-lab, home + up to 8 restarts	978/1,000 (97.8%)
k, home only	328/1,000 (32.8%)
k, home + the same 8 restart vectors	718/1,000 (71.8%)

For the restart-enabled variants, both solve 714 targets, arm-lab alone solves 264, k alone solves four, and neither solves 18. Their union solves 982/1,000. The exact paired sign/McNemar value is 8.998e-73, reported only as a descriptive statistic for this frozen cohort.

The pre-existing 97.8% arm-lab control was known before the protocol was frozen and is explicitly labeled non-blind. The cold-start, independent-solver, paired-cell, and union outcomes were not known before the evaluator ran.

Reproducibility

• protocol committed before evaluator and new outcomes: d9078f1b95a159cc09acf2995f895cd91c23aa6e
• evaluator committed before the first new result-bearing run: 6b3320a5e73406eaa9c497eeba3bc96960322e62
• a post-outcome corrective commit changes only tiny-probability formatting, not any target, result row, success rule, or analysis
• two consecutive full runs reproduced every result byte
• CI now replays the 4,000 rows in addition to the existing multi-scene, randomized, and independent-planner artifacts

Validation

• cargo fmt --all -- --check
• strict workspace Clippy across all targets and features
• 68/68 release tests
• full deterministic checks for the multi-scene, randomized, OxMPL, and new IK artifacts
• fresh pre-publication IK regeneration: every generated byte matched

Claim boundaries

Targets are generated by FK from in-limit configurations, so they are reachable under this exact model but are not Cartesian-uniform or task-representative. Evaluation is collision-unaware and simulation-only. Maximum starts and iteration caps are matched, but per-iteration work, update rules, damping, and early stopping differ; this is not a runtime, compute, hardware, sim-to-real, workspace-wide, or universal solver ranking.

Generated-by: OpenAI Codex (GPT-5; accessed 2026-08-18)