hybrid_astar_dyn_pp: control experiment shows dyn-aware search alone is brittle

[rsasaki0109]

Summary

A control experiment that complements PR #6 (hybrid_astar_dwa). Same Hybrid A* search as the baseline hybrid_astar_pp, but the search's collision check inflates each dynamic obstacle's predicted position along its time stamp. The tracker remains the same pure-pursuit. So the only difference vs. blind hybrid_astar_pp is the dynamic-aware search, and the question is: does global predictive avoidance close the gap without per-step local replanning?

Answer: no. The dyn-aware variant is worse than blind on hard cells.

planner	cells	solved	mean coll	hard solved	hard coll
hybrid_astar_pp	30	21	6.23	3/12	15.58
hybrid_astar_dyn_pp	30	20	6.37	2/12	15.92
hybrid_astar_dwa	30	30	0.00	12/12	0.00

The linearised constant-speed prediction is offset from the simulator's acceleration-from-rest schedule by ~1 s; the search avoids the ghost position and walks into the real obstacle. Inflating the safety buffer beyond 1 m either kills the path through the pincer convergence or collapses to the static path. The negative result tightens the case for hybrid_astar_dwa by showing that a per-step local controller is not optional once the obstacle dynamics matter — a smarter global search alone doesn't compensate.

What changed

• include/hybrid_astar_pp.h: hybrid_astar_plan now takes an optional vector<DynamicObstacleSpec> + t_offset. The Node tracks a per-node time stamp; each sub-step rejects expansions whose predicted obstacle position is inside (o.r + robot_radius). Lattice key is still (x, y, theta) (the standard 4D-without-time approximation).
• benchmark_diff_mppi.cu: ensure_hap_path_planned takes an include_dynamic flag. When set, the search uses target_speed for its v_search (so search-t roughly matches sim-t) and inflates robot_radius by hap_dyn_inflation = 1.0 m to absorb the constant-speed-search vs. accelerate-from-rest mismatch.
• Registered as hybrid_astar_dyn_pp (planner_kind=5).
• docs/local_planner_comparison.md regenerated from a new 11-planner sweep.
• docs/hybrid_astar_baseline.md extended with the negative-result section before the hybrid_astar_dwa section.

Test plan

• Build: cd build && cmake --build . --target benchmark_diff_mppi -j\$(nproc) (clean).
• Hard-cell smoke test on dynamic_pincer / crossing / slalom: confirmed dyn_pp collisions match pp ballpark, dwa is collision-free.
• Full sweep (3 scenarios × 5 speed × 2 radius × 4 seeds × 11 planners) regenerated docs/local_planner_comparison.md.

Known CI failure

validate_design_workflow fails on experiments.horizon_selection is missing PROBLEM_KIND — same pre-existing breakage on master. PR #7 fixes that; merge order is independent.