Add regression tests for lap S-crossover, lattice boundaries, and collision precompute glue

[cursor]

Summary

Adds focused regression coverage for high-risk planning/tracker paths left thinly tested after recent closed-loop track_end_s / lattice work. Complements (does not duplicate) open drafts #17/#20/#22/#23/#24.

Risky behavior now covered

• Race lap S-crossover (LocalPlanningStrategy.step): lap increments only when previous s is past 0.8 * track_end_s and new s is near start; does not treat the stale path_s[-2] band as near-end; mid-track and race_mode=False do not increment
• Lattice._check_boundary_violation: left(+)/right(−) boundary convention with clearance inset; missing parent Frenet path returns False
• generate_lattice_from_nodes hot-path glue: with agents, every check_collision call receives precomputed obstacle_polygons; without agents the kwarg stays None
• convert_sd_orientation_to_xy_orientation: world yaw = Frenet theta + heading of the cursor current_wp→next_wp segment

Test files added/updated

• test/c20_planning/test_c23_local_planning_pipeline.py
• test/c20_planning/test_c28_local_lattice_planners.py
• test/c50_common/test_c54_trajectory_waypoint_update.py

Why these tests materially reduce regression risk

Lap counting and lattice feasibility gate race sessions and overtaking. Silently reverting to path_s[-2], weakening boundary clearance, or dropping precomputed polygons would mis-count laps, accept off-road edges, or fall back to the slow collision path with different CV-sweep semantics. Orientation composition feeds spawn/teleport helpers that already depend on in-bounds next_wp.

Validation

pytest test/c20_planning/test_c23_local_planning_pipeline.py \
  test/c20_planning/test_c28_local_lattice_planners.py \
  test/c50_common/test_c54_trajectory_waypoint_update.py -v

All 59 tests passed. No production code changes.