S2.3 — Goal-region re-sampling for RRT-Connect (OMPL parity)

#task #oxmpl #oxmpl/epic #oxmpl/story/size-M

Moved to backlog 2026-10-08 (grilling decision): a feature-parity gap, not determinism/performance. Parked in oxmpl - backlog § Goal-region re-sampling; revisit with the samplers work. Cancelled as a ticket; the backlog entry carries it.

Intent

Found in blind review 2026-09-22 during S1.2 (deferred per scope discipline — feature gap, not error-handling).

oxmpl's RRT-Connect seeds its goal tree with a single sample_goal() root and never adds more. OMPL re-samples goals periodically (criterion: getSampledGoalsCount() < tGoal_->size()/2) during the grow loop, which materially strengthens RRT-Connect when the goal is a sampleable region (not a single state) — multiple goal roots give the bidirectional search more connection opportunities and typically faster, shorter-path convergence on goal regions.

Note: S1.2's goal-tree lazy-seed change (seed moved from setup into solve, seeded rng) creates the exact site this feature would extend.

Acceptance criteria

Goal regions get multiple roots in the goal tree during the grow loop (criterion inspired by OMPL's sampled-goals-count heuristic; exact rule decided in-story)
Single-state goals behave identically to today (no regression)
Benchmark on a goal-region problem showing improved path quality / solve time vs. single-root baseline (justify the feature with data)
just test-all green; new tests follow the planner×space matrix convention

Tasks

Decide re-sampling rule (when to add a goal root; cap?) — OMPL parity vs. simpler
Implement in RRT-Connect grow loop
Benchmark + regression tests (Rust/py/js matrix)