Add regression tests for HD-map chopping, plan load validation, and FTG safety

[cursor]

Risky behavior now covered

1. OpenDRIVE lane-sign path chops — chop_path / chop_path_from_two_sides reverse positive-id lanes and slice start/goal indexes. A regression here emits a reversed or truncated HD-map global path. _inset_d and smoothen_path_savgol keep corridor bounds and velocity samples aligned.
2. GlobalPlan file validation and zero-start ramp — is_loadable rejects missing keys, empty arrays, nested bounds, and non-JSON. from_file replaces velocity[0] <= 0 with c20_min_ramp_start_velocity so a stored 0-speed first waypoint does not pin the ego at a standstill.
3. Follow-the-Gap safety filters — 3D z-band discard, footprint bubble, no-forward / all-in-bubble zero command, and preferring a narrow interior corridor over wider ±90° edge openings.
4. GoalArrivalMonitor missing-goal guards — no local planner, missing/short goal_point, and reset() re-arming the rising edge so a false GOAL_ARRIVED cannot stop the run.

Test files added/updated

• test/c20_planning/test_c24_global_hdmap_planners.py (added)
• test/c20_planning/test_c21_planning_model.py (added)
• test/c30_control/test_c35_pure_pursuit.py (updated)
• test/c40_execution/test_c47_execution_tasks.py (added)

Why these tests materially reduce regression risk

These are shared, high-blast-radius surfaces that landed on dev without dedicated tests: HD-map route geometry, stored global-plan ingest, LiDAR steering safety, and goal-stop event gating. They were not claimed by open coverage/bugfix drafts #17/#20/#22–#38.

Validation

PYTHONPATH=/workspace python3 -m pytest \
  test/c20_planning/test_c21_planning_model.py \
  test/c20_planning/test_c24_global_hdmap_planners.py \
  test/c30_control/test_c35_pure_pursuit.py \
  test/c40_execution/test_c47_execution_tasks.py -q

41 passed. Deterministic unit tests; no slow / requires_data markers.