Match task roles to compatible asset combinations

agent_context/MAP.yaml

@@ -445,11 +445,12 @@ topics:
 - id: asset-generalization
   title: Asset Generalization
   aliases: [asset generalization, asset provider, 资产泛化, 资产语义匹配]
-  keywords: [AssetDescriptor, PreparedAsset, AssetProvider, AssetRequirements, AssetRoleRequirement]
+  keywords: [AssetDescriptor, PreparedAsset, AssetProvider, AssetRequirements, AssetRoleRequirement, ResolvedAssetSet, resolve_asset_set]
   paths: [topics/asset-generalization/asset-generalization.md]
   source_of_truth:
   - embodichain/data/asset_catalog/
   - embodichain/lab/task_program/semantics/asset_requirements.py
-  watch_paths: [embodichain/data/asset_catalog/, tests/data/asset_catalog/, embodichain/lab/task_program/semantics/asset_requirements.py, tests/lab/task_program/semantics/test_asset_requirements.py]
+  - embodichain/lab/task_program/semantics/asset_matching.py
+  watch_paths: [embodichain/data/asset_catalog/, tests/data/asset_catalog/, embodichain/lab/task_program/semantics/asset_requirements.py, tests/lab/task_program/semantics/test_asset_requirements.py, embodichain/lab/task_program/semantics/asset_matching.py, tests/lab/task_program/semantics/test_asset_matching.py]
   related_topics: [data-assets, task-programs, gen-sim, scene-services]
   status: active

agent_context/topics/asset-generalization/asset-generalization.md

@@ -33,3 +33,14 @@ the complete package into an immutable snapshot. `verify_prepared_asset` is the
 common integrity boundary for any provider's prepared output; its package-root,
 file-inventory and manifest contract is documented in that function. Changing
 source content creates a new identity; changing cached content is an error.
+
+## Task matching
+
+`semantics/asset_matching.py` owns bounded role matching. It resolves explicit
+category aliases, checks hard properties/capabilities and dimensions, and prunes
+upright support/containment assignments as roles are selected. It never builds
+the full Cartesian product. Candidate and search-node budgets are independent;
+exhaustion produces structured rejection evidence rather than a relaxed match.
+`ResolvedAssetSet` freezes selected content and requirement identity with a seed.
+Absolute cache paths are excluded from its identity. Same asset selection does
+not establish reachable grasps, settled state or measured task success.

docs/source/api_reference/asset_generalization.rst

@@ -146,3 +146,28 @@ Directory and explicit providers retain candidate metadata without claiming task
    :members:
 
 .. autofunction:: embodichain.data.asset_catalog.local.verify_prepared_asset
+
+embodichain.lab.task_program.semantics.asset_matching additions
+---------------------------------------------------------------
+
+Bounded deterministic selection applies hard role constraints and upright geometric relations before freezing an asset set. Unknown metadata cannot satisfy required constraints; a resolved set does not certify physical success.
+
+.. currentmodule:: embodichain.lab.task_program.semantics.asset_matching
+
+.. autosummary::
+
+   AssetRejection
+   AssetResolutionError
+   ResolvedAssetSet
+   resolve_asset_set
+
+.. autoclass:: embodichain.lab.task_program.semantics.asset_matching.AssetRejection
+   :members:
+
+.. autoclass:: embodichain.lab.task_program.semantics.asset_matching.AssetResolutionError
+   :members:
+
+.. autoclass:: embodichain.lab.task_program.semantics.asset_matching.ResolvedAssetSet
+   :members:
+
+.. autofunction:: embodichain.lab.task_program.semantics.asset_matching.resolve_asset_set