Add multi-scene planning and penetration-aware tracking evidence

[Dylan-Gallagher]

Summary

• add a deterministic multi-scene planning and tracking benchmark covering
  three shipped MuJoCo scenes and nine fixed scene/query fixtures
• preserve raw planning/tracking CSVs and a generated result ledger in the
  repository
• audit executed robot contacts at every sampled 2 ms state and require zero
  sampled penetration for a tracking pass
• run the bounded artifact replay in CI, exact-comparing every deterministic
  field and outcome while ignoring only wall-clock planner timing
• clarify that the scripted pick/place cube is visual/mocap-only and is not
  part of the planner's collision geometry

Frozen result

• planning: 30/30 direct-free and 15/15 obstructed trials succeed
• tracking: position PD passes 0/18 cases
• PD plus desired-velocity feedforward meets numeric gates in 14/18 cases and
  the numeric-plus-zero-sampled-penetration gate in 13/18
• four of 36 executed cases retain penetration rather than being discarded;
  all are the tabletop_pillar/reverse_cross_workspace fixture, with 25–72
  affected path samples and 0.050–0.079 mm maximum penetration

The result is intentionally evidence-bounded: the tasks are hand-designed,
planner collision checks are discrete, executed collision checks detect
sampled penetration rather than continuous swept-volume safety, and there is
no hardware or sim-to-real claim. The carried cube is not modeled as planner
collision geometry in this branch.

Reproducibility

cargo fmt --all -- --check
cargo clippy --workspace --all-targets -- -D warnings
cargo test --release -p arm-lab
cargo test --release -p arm-lab-demo --bin multi_query_bench
cargo run --release -p arm-lab-demo --bin multi_query_bench -- --check
cargo build --release -p arm-lab-demo

Local release verification on the exact head passes 34 arm-lab tests, seven
benchmark tests, strict lint/formatting, the full deterministic artifact check,
and the demo build.

Audit history

The commit sequence deliberately retains the evidence corrections instead of
squashing them away: the benchmark matrix was locked before outcomes, executed
contact auditing was then added, and the collision claim was narrowed from a
positive threshold to MuJoCo-emitted sampled penetration (dist < 0). This
history makes the limitations and retained failure cases reviewable.