Add regression tests for predictor sweeps, async snapshot, and harvest siblings

[cursor]

Summary

Adds focused regression coverage for high-risk paths that recent execution/planning work left thinly tested. Complements (does not duplicate) open draft PRs #17/#18.

Risky behavior now covered

• Predictor-gated obstacle sweeps (precompute_obstacle_polygons): movers without SingleTrajectory stay static boxes; predicted agents sweep forward; far-behind agents are not beside-swept; just-behind agents use beside_sweep_time
• Async combined planner worker: one get_sensor_frame per iteration shared by replan+perceive; no fetch when no stage module is active
• stack_event harvest siblings: GlobalPlan.stack_event and localization pm.stack_event notify once then clear
• DETECTION GT list-copy: stack clears cannot wipe world NPC lists; DETECTION-off clears stack agents only
• update_waypoint_by_xy_forward: never moves current_wp backward; honors min_wp floor

Test files added/updated

• test/c50_common/test_c55_collision_checking.py
• test/c40_execution/test_c42_sensor_snapshot.py
• test/c40_execution/test_c43_task_strategy.py
• test/c50_common/test_c54_trajectory_waypoint_update.py

Why these tests materially reduce regression risk

These paths sit on lattice collision costs, CARLA-async sensor timing, task automation signals, GT detection isolation, and HDMap waypoint tracking. Silent regressions here produce unsafe lattice costs, stage desync on shared sensors, missed task events, despawned world NPCs, or path-index backjumps after sharp turns.

Validation

pytest test/c50_common/test_c55_collision_checking.py \
  test/c50_common/test_c54_trajectory_waypoint_update.py \
  test/c40_execution/test_c42_sensor_snapshot.py \
  test/c40_execution/test_c43_task_strategy.py -v

All 35 tests passed. No production code changes.