Auto-sync: 2026-06-29 08:12:42 (part 2)

scripts/build_paper_analysis.py

@@ -315,6 +315,22 @@ METHODS = [
             "k4_grid035040045_safe_margin0p20_mean_by_type_noopbonus0p03_consensus0p10_summary.json"
         ),
     ),
+    MethodSpec(
+        key="residual_k4_compose_grid035040045",
+        label="K4 composed type-consensus tangents, scales 0.35/0.40/0.45",
+        summary_path=(
+            "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
+            "k4_compose_grid035040045_safe_margin0p20_summary.json"
+        ),
+    ),
+    MethodSpec(
+        key="residual_k4_compose_grid035040045_noopbonus003",
+        label="K4 composed type-consensus tangents, scales 0.35/0.40/0.45, no-op bonus 0.03",
+        summary_path=(
+            "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
+            "k4_compose_grid035040045_safe_margin0p20_noopbonus0p03_summary.json"
+        ),
+    ),
     MethodSpec(
         key="repair_nearmiss_k4_grid025035050_margin020",
         label="K4 near-miss-to-expert repair tangent, scales 0.25/0.35/0.50, margin 0.20",

scripts/build_paper_table_status.py

@@ -655,6 +655,26 @@ SPECS = [
         story_role="train-neighbor tangent-consensus confidence on the current best typed prior",
         pending_job="14903390/14903391",
     ),
+    ResultSpec(
+        key="retrieval_residual_k4_compose_grid035040045",
+        label="K4 composed type-consensus residual retrieval, scales 0.35/0.40/0.45, margin 0.20",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_compose_grid035040045_safe_margin0p20_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="local tangent composition without typed priors",
+        pending_job="14905956/14905963",
+    ),
+    ResultSpec(
+        key="retrieval_residual_k4_compose_grid035040045_noopbonus003",
+        label="K4 composed type-consensus residual retrieval, scales 0.35/0.40/0.45, margin 0.20, no-op bonus 0.03",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_compose_grid035040045_safe_margin0p20_noopbonus0p03_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="local tangent composition on the current best typed prior",
+        pending_job="14905957/14905964",
+    ),
     ResultSpec(
         key="retrieval_repair_nearmiss_k4_grid025035050_margin020",
         label="K4 near-miss-to-expert repair tangent, scales 0.25/0.35/0.50, margin 0.20",

scripts/eval_maniskill_policy_rollout.py

@@ -218,11 +218,13 @@ def main(argv: list[str] | None = None) -> int:
             "mean_by_type",
             "median_by_type",
             "kernel_mean_by_type",
+            "compose_mean_by_type",
             "field_softmax",
         ),
         default="none",
         help="Optional consensus reduction over retrieved residuals with the same candidate type. "
         "'kernel_mean_by_type' weights source residuals by train-state retrieval distance; "
+        "'compose_mean_by_type' also adds pairwise sums of type-consensus tangents; "
         "'field_softmax' forms a field-weighted tangent barycenter before final scoring.",
     )
     parser.add_argument(

tests/test_maniskill_policy_rollout.py

@@ -895,6 +895,44 @@ def test_retrieval_residual_reducer_builds_type_consensus() -> None:
     )
 
 
+def test_retrieval_residual_reducer_builds_composed_type_tangents() -> None:
+    residuals, candidate_types = _reduce_residual_candidates_by_type(
+        [
+            [[0.0, 0.0]],
+            [[0.2, 0.0]],
+            [[0.0, 0.0]],
+            [[0.4, 0.2]],
+            [[-0.5, 0.0]],
+        ],
+        [
+            "policy_residual",
+            "residual_no_op",
+            "policy_residual",
+            "residual_no_op",
+            "residual_wrong_gripper",
+        ],
+        mode="compose_mean_by_type",
+    )
+
+    assert candidate_types == [
+        "policy_residual",
+        "residual_no_op",
+        "residual_wrong_gripper",
+        "residual_no_op+residual_wrong_gripper",
+    ]
+    assert np.allclose(
+        np.asarray(residuals, dtype=np.float32),
+        np.asarray(
+            [
+                [[0.0, 0.0]],
+                [[0.3, 0.1]],
+                [[-0.5, 0.0]],
+                [[-0.2, 0.1]],
+            ]
+        ),
+    )
+
+
 def test_retrieval_residual_reducer_penalizes_low_consensus_tangents() -> None:
     residuals, candidate_types, bonuses = _reduce_residual_candidates_by_type(
         [