Add regression tests for velocity collision glue, Kalman tracks, and OpenDRIVE geometry

[cursor]

Summary

Adds focused regression coverage for high-risk paths that remained untested on dev, while avoiding duplication of open drafts #17/#20/#22/#23/#24/#25/#26/#27/#28.

Risky behavior now covered

• VelocityLocalPlanner.profile_trajectory obstacle_polygons glue: with agents, check_collision receives precomputed polygons (and precompute gets total_time / thresholds / inflation, not lattice beside_* windows); without agents the kwarg stays None and precompute is skipped
• KalmanTracker association lifecycle: persistent track IDs across frames with positive velocity estimate; detections beyond gate_distance spawn new IDs; tracks prune after max_missed; None perception model raises
• OpenDRIVE geometry / lane offset: arc quarter-circle endpoints; zero-curvature spiral falls back to line; _get_lane_offset_at_s selects the latest applicable polynomial segment

Test files added/updated

• test/c20_planning/test_c27_local_behavioral_and_velocity_planners.py
• test/c10_perception/test_c15_kalman_tracker.py (new)
• test/c10_perception/test_c18_hdmap_parser.py

Why these tests materially reduce regression risk

Velocity profiling is the default local-plan speed path; dropping obstacle_polygons silently changes collision semantics vs predictor-gated precompute. Kalman association owns persistent agent IDs and estimated velocity used by prediction/planning. Arc/offset parsing feeds HD-map centerlines—wrong geometry corrupts global plans on curved roads.

Validation

pytest \
  test/c20_planning/test_c27_local_behavioral_and_velocity_planners.py \
  test/c10_perception/test_c15_kalman_tracker.py \
  test/c10_perception/test_c18_hdmap_parser.py -v

All 33 tests passed. No production code changes.