Add predeclared randomized planning evaluation

[Dylan-Gallagher]

Summary

• add a protocol frozen in its own commit before the evaluator and result-bearing runs
• retain 300 independently sampled joint-space queries across the three shipped scenes
• record all 2,253 planner outcomes and 276 nominal controller replays in raw CSV artifacts
• add generated aggregate reporting, exact artifact verification, and a bounded CI reproduction

Results

• direct interpolation: 83/300 sampled-clear queries (27.7%, 95% Wilson 22.9--33.0%)
• canonical RRT-Connect: 138/300 successes (46.0%, 95% Wilson 40.4--51.7%), including 55/217 recoveries on blocked-direct queries
• blocked-query goal-bias ablation: 271/1,085 default successes versus 270/1,085 with zero goal bias
• position PD: 0/138 numeric and full tracking passes
• PD + velocity feedforward: 104/138 numeric passes and 90/138 full zero-sampled-penetration passes
• retained negatives: 162 canonical planning failures and 37/276 tracking cases with sampled penetration, maximum depth 3.2225 mm

Validation

• cargo fmt --all -- --check
• cargo clippy --locked --workspace --all-targets --all-features -- -D warnings
• cargo test --locked --release --workspace --all-targets
• cargo run --locked --release -p arm-lab-demo --bin multi_query_bench -- --check
• cargo run --locked --release -p arm-lab-demo --bin randomized_eval -- --check
• separate row-level audit of cohort sizes, unique queries, seed formulas, planner layouts, paired controller replays, gates, and reported aggregates

Scope

This changes evaluation and documentation only; the planner, controllers, collision predicate, trajectory limits, actuator model, and thresholds are unchanged. The evidence is deterministic simulation with sampled collision checks. It is not hardware, continuous-safety, grasping, workspace-uniform, or sim-to-real evidence.