Add regression tests for LidarLocalization ICP and BasicSim kinematics

[cursor]

Summary

Adds focused regression coverage for high-risk paths that remained untested on dev, while avoiding duplication of open drafts #17/#20/#22–#32.

Risky behavior now covered

• LidarLocalization ICP: first scan seeds the reference map without mutating ego; reset() clears map/pose; pure translation and yaw recovery on a seeded asymmetric cloud; None/empty/<3-point/3D z-band-empty scans are no-ops
• BasicSim bicycle ego kinematics: XY integrates with pre-update velocity, then accel updates v, then yaw uses post-update v with default axle length 2.5
• BasicSim NPC plant: with npc_control=True, agents advance and steer is slew-limited to |Δsteer| ≤ 3·dt; with npc_control=False, agents stay frozen while ego moves
• State.get_bb_corners: axis-aligned corners and 90° yaw length/width swap used by LiDAR raycasting and collision geometry

Test files added/updated

• test/c10_perception/test_c16_lidar_localization.py (new)
• test/c40_execution/test_c46_basic_sim_kinematics.py (new)
• test/c10_perception/test_c11_state_geometry.py (new)

Why these tests materially reduce regression risk

LidarLocalization is the only non-GT localization path and silently writes ego pose from scan alignment—wrong ICP/guards corrupt the whole stack state estimate. BasicSim bicycle/NPC integration is the plant for Control Step/Steer and multi-agent sims; integration-order or steer-slew regressions change every closed-loop demo. Bounding-box corners feed LiDAR hits and corridor collision; wrong rotation quietly shifts perceived obstacles.

Validation

pytest \
  test/c10_perception/test_c16_lidar_localization.py \
  test/c40_execution/test_c46_basic_sim_kinematics.py \
  test/c10_perception/test_c11_state_geometry.py -v

All 14 tests passed. No production code changes.