Goal
Improve C-WAH action sequencing by tracking held-object state and avoiding incoherent grab/placement choices.
Context
#201 added local precondition hints and setup actions, but the bounded real comparison still produced failed action records dominated by grab and putin. The next missing layer is hand occupancy and held-object-aware sequencing.
Scope
- Infer per-agent held objects from agent-local symbolic observations.
- Add hand/held-object metadata to legal actions.
- Mark grab actions as blocked or lower priority when already holding an object.
- Prefer placement actions for held goal objects before grabbing unrelated objects.
- Prefer setup actions for placement targets when target is not close or needs opening.
- Preserve no evaluator/full-graph/debug leakage.
Acceptance Criteria
- Unit tests cover holding-object blocks unrelated grab, held goal object prioritizes placement, target-close setup, and existing leakage behavior.
- Existing C-WAH tests and targeted CRAFT leakage/regression tests pass.
- Mock baseline workflow passes.
- Bounded real CoELA comparison is run if runtime assets are available.
Goal
Improve C-WAH action sequencing by tracking held-object state and avoiding incoherent grab/placement choices.
Context
#201 added local precondition hints and setup actions, but the bounded real comparison still produced failed action records dominated by grab and putin. The next missing layer is hand occupancy and held-object-aware sequencing.
Scope
Acceptance Criteria