Fix closed-loop Frenet wrap snap to s≈0 on last lap meters

[cursor]

Bug and impact

On closed race lines with duplicated endpoints (first==last, e.g. bundled Yas Marina), ego on the last segment got Frenet s≈0 (or slightly negative) instead of s≈track_end. Lattice sampling / lap detection then acted as if the car were at the start/finish while it was still meters before the line — wrong local plans and premature lap counts every lap.

Trigger: Load a closed race line → drive (or place ego) onto the final polyline segment before S/F. Reproduced on Yas Marina: true s≈5250.7 reported as s≈-0.3.

Root cause

PR #21 fixed path_s init and open-path corner segment scoring, but XY→SD still only scored segments adjacent to the KD nearest waypoint. At the finish, first==last makes the tree return index 0, so only segment (0,1) was considered — never the wrap segment (n-2, n-1).

Fix

• Cache __closed_duplicated_endpoints at init.
• _frenet_segment_candidates also scores (n-2, n-1) when closest_wp==0 on those paths.
• Shared by scalar and numpy Frenet converters.

Validation

PYTHONPATH=/workspace python3 -m pytest test/c50_common/test_c54_closed_loop_path_s.py -v

7 passed (includes new square + Yas Marina wrap-segment regressions).

Broader test/c50_common + lattice/velocity/PP: 127 passed (2 failures are env tkinter missing, unrelated).

Memory cleanup

• Removed merged PR Fix closed-loop path_s corruption and Frenet corner false CTE #21 from automation memory.
• Not re-reporting open PR Fix StopExecAtGoal async restart and combined stage order #22 (async stop restart / combined stage order).