Add pair-scoped attached-payload collision checks

[Dylan-Gallagher]

Summary

• add an end-effector-attached 50 mm box proxy for Demo 3 payload planning
• evaluate explicit cube-to-floor/table/pillar MuJoCo geometry distances with a
  strict 5 mm payload-only planning buffer
• retain the robot's existing strict zero-penetration predicate and exclude
  the intended payload/attachment-region contact by construction
• integrate the combined predicate only into the carry segment and preserve a
  pre-result protocol/result commit boundary

Fixed-query evidence

The original straight carry is rejected by both predicates: across 23 samples,
it has nine robot-collision samples and 11 payload-margin violations, with a
minimum cube/pillar distance of -0.065709578 m.

The fixed-seed planner returns six waypoints and 47 densified samples with:

• zero sampled robot-collision states
• zero sampled payload-margin violations
• 0.011848632 m minimum cube/pillar distance
• 0.006848632 m reserve above the 5 mm payload gate
• 2.251250420 rad joint-space path cost

The improvement is not presented as free: the carry grows from four to six
waypoints and representative worst tracking error rises from 0.0037 to
0.0067 rad.

Claim boundary

This is a sampled planning-path result for one deterministic rigid box/query at
0.05 rad joint-space edge spacing. It is not a continuous swept-volume or
hardware safety certificate. The experiment does not model grasp stability,
slip, uncertainty, fingers, compliance, calibration error, or payload-versus-
robot self-collision outside the explicitly allowed attachment contact. The
visual/mocap weld remains a scripted grasp stand-in, and executed measured
states are not audited for payload clearance in this experiment.

Reproducibility

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

The exact branch passes 40 arm-lab tests, seven benchmark tests, strict
lint/formatting, the full existing multi-scene artifact check, compiled-pose
transform validation below 1e-10, and independent source/evidence review.

Audit history

The first commit fixes geometry, pair scope, query, seed, 5 mm threshold, and
acceptance gates before implementation/results. The second commit contains the
implementation and measured evidence. Please preserve both commits rather than
squashing away that boundary.