auto-sync 2026-07-02T17:27:17Z workspace (part 4)

workspace/tests/test_causal_action_metrics.py

@@ -0,0 +1,58 @@
+from __future__ import annotations
+
+import math
+
+from dovla_cil.eval.metrics import (
+    branch_causal_action_regret,
+    candidate_prefix_causal_metrics,
+    causal_action_decomposition,
+    normalized_causal_action_regret,
+    positive_tangent_recall_at_k,
+    selector_regret_at_k,
+    support_gap,
+)
+
+
+def test_causal_action_regret_metrics() -> None:
+    assert branch_causal_action_regret(2.0, 1.25) == 0.75
+    assert selector_regret_at_k([1.25, 2.0, 0.5], selected_index=0) == 0.75
+    assert support_gap(3.0, 2.0) == 1.0
+    assert positive_tangent_recall_at_k([0.4, 0.7], base_utility=0.6) == 1.0
+    assert positive_tangent_recall_at_k([0.4, 0.6], base_utility=0.6) == 0.0
+    assert math.isclose(
+        normalized_causal_action_regret(2.0, 1.25, 1.0),
+        0.75,
+        rel_tol=1.0e-6,
+    )
+
+
+def test_candidate_prefix_metrics_use_policy_residual_as_base() -> None:
+    metrics = candidate_prefix_causal_metrics(
+        branch_scores=[1.0, 1.4, 0.2],
+        branch_types=[
+            "retrieval_residual_policy_residual",
+            "retrieval_residual_residual_near_miss",
+            "retrieval_residual_residual_wrong_gripper",
+        ],
+        valid_mask=[True, True, True],
+    )
+
+    assert math.isclose(metrics["car_to_proposal_oracle"], 0.4)
+    assert math.isclose(metrics["selector_regret_at_k"], 0.4)
+    assert metrics["ptr_at_k"] == 1.0
+    assert metrics["base_utility"] == 1.0
+
+
+def test_causal_action_decomposition_matches_support_selector_story() -> None:
+    decomposition = causal_action_decomposition(
+        base=0.2974,
+        selected=0.3890,
+        proposal_oracle=0.4435,
+        same_state_oracle=0.5699,
+        full_oracle=0.6933,
+    )
+
+    assert math.isclose(decomposition.clean_gain, 0.0916)
+    assert math.isclose(decomposition.selector_gap, 0.0545)
+    assert math.isclose(decomposition.support_gap, 0.1264)
+    assert math.isclose(decomposition.target_for_gap_closure(0.65), 0.474525)

workspace/tests/test_cil_schema.py

@@ -5,9 +5,12 @@ import pytest
 from dovla_cil.data.schema import (
     CIL_VERSION,
     ActionChunk,
+    CILBenchBranch,
+    CILBenchGroup,
     CILGroup,
     CILRecord,
     FailureInfo,
+    OutcomeVector,
     RewardInfo,
     StructuredEffect,
     compute_regret_and_ranks,
@@ -96,3 +99,62 @@ def test_deterministic_record_ids_and_state_hashes() -> None:
     assert make_record_id("g", "a", 1) != make_record_id("g", "a", 2)
     assert compute_state_hash(b"state") == compute_state_hash(b"state")
     assert compute_state_hash(b"state") != compute_state_hash(b"other")
+
+
+def test_outcome_vector_from_reward_and_utility() -> None:
+    reward = RewardInfo(
+        progress=0.6,
+        success=True,
+        terminal_success=True,
+        dense_components={
+            "contact_quality": 0.5,
+            "task_stage_quality": 0.25,
+            "smoothness": 0.8,
+            "recovery": 1.0,
+        },
+    )
+
+    outcome = OutcomeVector.from_reward(reward)
+
+    assert outcome.success == 1.0
+    assert outcome.progress == 0.6
+    assert outcome.contact_quality == 0.5
+    assert outcome.safety_violation == 0.0
+    assert outcome.lexicographic_utility() > reward.score
+
+
+def test_cilbench_group_roundtrip_and_same_state_contrast() -> None:
+    base_action = ActionChunk(action_id="base", horizon=1, values=[[0.0, 0.0]])
+    repair_action = ActionChunk(action_id="repair", horizon=1, values=[[0.1, 0.0]])
+    group = CILBenchGroup(
+        group_id="chart-0",
+        task_id="PickCube-v1",
+        split_id="train",
+        simulator_state_hash="hash",
+        instruction="pick the cube",
+        observation_ref=None,
+        observation_inline={"rgb": "obs/000.png"},
+        scene_metadata={"target_object": "cube"},
+        anchor_policy="h16_bc",
+        branches=[
+            CILBenchBranch(
+                branch_id="base",
+                action=base_action,
+                branch_family="anchor",
+                outcome=OutcomeVector(success=0.0, progress=0.2),
+            ),
+            CILBenchBranch(
+                branch_id="repair",
+                action=repair_action,
+                branch_family="recovery_tangent",
+                outcome=OutcomeVector(success=1.0, progress=0.8, recovery=1.0),
+            ),
+        ],
+    )
+
+    restored = CILBenchGroup.from_dict(group.to_dict())
+
+    assert restored == group
+    assert restored.same_state_causal_contrast("repair", "base") > 0.0
+    with pytest.raises(KeyError):
+        restored.same_state_causal_contrast("repair", "missing")

workspace/tests/test_maniskill_policy_rollout.py

@@ -135,6 +135,10 @@ def test_policy_rollout_summary_includes_candidate_oracle_when_present() -> None
             "candidate_oracle_restore_error": 3e-7,
             "candidate_oracle_candidate_type": "retrieval_residual_residual_no_op",
             "candidate_oracle_valid_mask": [True, True],
+            "candidate_oracle_types": [
+                "retrieval_residual_policy_residual",
+                "retrieval_residual_residual_no_op",
+            ],
             "candidate_oracle_branch_successes": [False, True],
             "candidate_oracle_branch_progress": [0.1, 0.8],
             "candidate_oracle_branch_scores": [0.1, 1.8],
@@ -162,6 +166,10 @@ def test_policy_rollout_summary_includes_candidate_oracle_when_present() -> None
             "candidate_oracle_restore_error": 2e-7,
             "candidate_oracle_candidate_type": "retrieval_residual_policy_residual",
             "candidate_oracle_valid_mask": [True, True],
+            "candidate_oracle_types": [
+                "retrieval_residual_policy_residual",
+                "retrieval_residual_residual_near_miss",
+            ],
             "candidate_oracle_branch_successes": [True, True],
             "candidate_oracle_branch_progress": [0.9, 0.8],
             "candidate_oracle_branch_scores": [1.9, 1.8],
@@ -188,6 +196,10 @@ def test_policy_rollout_summary_includes_candidate_oracle_when_present() -> None
         summary["candidate_oracle_branch_score_gains_over_selected"],
         [0.0, 0.8],
     )
+    assert np.isclose(summary["candidate_oracle_car_to_proposal_oracle"], 0.85)
+    assert np.isclose(summary["candidate_oracle_selector_regret_at_k"], 0.85)
+    assert summary["candidate_oracle_ptr_at_k"] == 0.5
+    assert summary["candidate_oracle_base_trace_coverage"] == 1.0
 
 
 def test_policy_rollout_loads_state_archive(tmp_path: Path) -> None: