Auto-sync: 2026-06-30 03:27:59 (part 4)

tests/test_maniskill_policy_rollout.py

@@ -21,6 +21,7 @@ from dovla_cil.eval.maniskill_policy_rollout import (
     _attach_retrieved_residual_candidates,
     _diagnostic_candidate_indices,
     _effective_lattice_candidate_count,
+    _lattice_candidate_include_mask,
     _lattice_candidate_mask,
     _lattice_candidate_type_bonus,
     _load_state_archive,
@@ -311,6 +312,33 @@ def test_retrieval_residual_margin_can_abstain_to_policy() -> None:
     assert index.tolist() == [0]
 
 
+def test_retrieval_residual_challenger_can_override_primary_anchor() -> None:
+    import torch
+
+    mean = torch.zeros(1, 1, 1)
+    offset = torch.full_like(mean, 0.4)
+    model = _StubModel(torch, mean, best_offset=offset)
+    residuals = torch.tensor([[[[0.0]], [[0.1]], [[0.4]]]], dtype=torch.float32)
+
+    actions, index = _select_action_chunk(
+        model,
+        observations=torch.zeros(1, 3),
+        instructions=["a"],
+        torch=torch,
+        selection_mode="retrieval_residual",
+        num_candidates=1,
+        candidate_sigma=0.0,
+        selection_seed=0,
+        selection_margin=1.0,
+        action_candidates=residuals,
+        challenger_mask=torch.tensor([[False, False, True]]),
+        challenger_margin=0.05,
+    )
+
+    assert torch.allclose(actions, mean + offset)
+    assert index.tolist() == [2]
+
+
 def test_candidate_oracle_prefix_keeps_deployed_candidate_first() -> None:
     import torch
 
@@ -554,6 +582,39 @@ def test_lattice_candidate_mask_can_exclude_exact_composite_without_parts() -> N
     assert mask.tolist() == [[True, True, False, True]]
 
 
+def test_lattice_candidate_include_mask_uses_exact_candidate_types() -> None:
+    import torch
+
+    case = _RolloutCase(
+        group_id="g",
+        task_id="PickCube-v1",
+        source_dataset=Path("."),
+        state={},
+        observation={"features": [0.0]},
+        instruction="pick",
+        oracle_score=1.0,
+        oracle_success=True,
+        expert_score=1.0,
+        expert_success=True,
+        best_action_values=[[0.0]],
+        candidate_action_values=[[[0.0]], [[0.2]], [[0.4]]],
+        candidate_types=[
+            "policy_residual",
+            "residual_near_miss",
+            "residual_near_miss+residual_wrong_gripper",
+        ],
+    )
+
+    mask = _lattice_candidate_include_mask(
+        [case],
+        torch=torch,
+        device="cpu",
+        include_types=("residual_near_miss",),
+    )
+
+    assert mask.tolist() == [[False, True, False]]
+
+
 def test_lattice_candidate_type_bonus_can_sum_composite_parts() -> None:
     import torch