Add an independent OxMPL planner comparison

[Dylan-Gallagher]

Summary

• add a predeclared comparison against independently maintained OxMPL 0.6.0
• retain all 1,085 paired seed trials over the 217 frozen blocked-direct queries
• independently revalidate every external path with arm-lab's MuJoCo checker
• commit the full CSV, generated report, exact-host two-run record, and fast CI artifact verifier

Result

• in-repository RRT-Connect: 271/1,085 successes
• OxMPL RRT-Connect: 290/1,085 successes
• paired cells: 271 both, 0 in-repository only, 19 OxMPL only, 795 neither
• independently valid OxMPL paths: 290/290; invalid paths: 0
• blocked queries with at least one successful seed: 56/217 in-repository, 58/217 OxMPL

The protocol was committed before the external dependency, evaluator, or result-bearing run. A write followed by an immediate full check matched every status and exact path digest after normalizing wall time.

Validation

• cargo fmt --all -- --check
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test --release --workspace --all-targets
• multi_query_bench --check
• randomized_eval --check
• oxmpl_baseline --check
• oxmpl_baseline --verify-artifacts
• independent 1,085-row CSV key/seed/status invariant audit

Boundaries

This is a frozen-cohort cross-implementation comparison, not a universal or compute-matched planner ranking. The in-repository rows use a 2,000-iteration budget; OxMPL uses a predeclared 250-ms wall-time budget. The integer seeds are paired labels but feed different PRNGs. The seed-level McNemar/binomial calculation is retained as descriptive only because five seeds share each query. Collision checks remain sampled simulation checks, with no continuous-safety, hardware, or sim-to-real claim.

Development disclosure

OpenAI Codex assisted with implementation, testing, analysis, and documentation. Dylan Gallagher is responsible for reviewing the changes, evidence, and public claims.