- Extract _compute_mutant helper with circular angle diff (fixes 0/2π boundary bug)
- Fix currenttobest1bin (remove non-standard noise term), add rand1bin strategy
- Decoupled mutation: independent F ranges for position vs theta
- Configurable crossover mode: per-device (default) or per-dimension
- Discrete angle snapping in normal 3N DE (joint mode, replaces hybrid as default)
- Stop auto-injecting prefer_orientation_mode into DE
- Expose DE hyperparameters (mutation, theta_mutation, recombination, strategy, angle_mode) via API
Replace scipy differential_evolution with custom DE loop for per-device
crossover, circular θ wrapping, and configurable mutation strategy
(currenttobest1bin default, best1bin as turbo mode).
Key improvements:
- Graduate ALL hard constraints during DE (proportional penalty instead of
flat inf), giving DE smooth gradient for reachability, min_spacing, etc.
Binary inf preserved for final pass/fail reporting.
- 2-axis sweep-and-prune AABB broad phase for collision pair pruning
- Multi-seed injection from multiple seeder presets + Gaussian variants
- snap_theta_safe: collision-check after angle snapping, revert on violation
- Weight normalization (100 distance / 60 angle / 5× hard multiplier)
- Constraint priority field (critical/high/normal/low → weight multiplier)
with LLM intent interpreter setting priority per constraint type
- Final success field now checks user hard constraints in binary mode
- arm_slider added to mock checker reach table (1.07m)
Tests: 202 passed, 24 new tests added (optimizer 7, constraints 6, broad_phase 11)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
