* fix(langgraph-agent): describe selected primary source
* fix(langgraph): isolate specialized engine loading failures
* refactor(egress): centralize engine classification
* fix(langgraph-agent): describe primary source neutrally
* fix(langgraph-agent): address review findings on primary-source metadata
- restore the pre-refactor AttributeError degradation for
retriever_registry.get_metadata in _load_specialized_engine_tools:
an older/minimal registry without the metadata API hands the decision
to evaluate_retriever (metadata=None, PDP fails closed as
retriever_unclassified) instead of being swallowed by the broad
per-engine skip; regression test drives the real loop with a
metadata-less registry
- test the real format_primary_search_description prose unmocked, and
prove end-to-end that an identifier-laden collection primary leaks no
engine key, UUID, display name, or config text into the model-visible
web_search description
- cover the retriever-primary PUBLIC and UNSPECIFIED classification
branches in _build_tools
- pin the two downstream consumers of the collection is_local=True fix:
settings UI keeps public collections in Local RAG, and the MCP engine
listing reports them is_local=true
- drop the unenforced '# allow: SIZE_OK' marker and the inert
is_public/is_local attributes on the routing test's engine stand-in
LearningCircuit
a8d7f7127a