[namedgraph]

Why

The XML serialization's executor has a core Reconcile that matches names such as a place of birth, a home, an employer or a school against Wikidata. This PR puts the same operation into the spec and the Python implementation, so the two serializations keep the same catalog (§5).

What

• formal-semantics.md §4.7 Reconciliation specifies Reconcile as a query effect, URI × Literal × Maybe Literal × Maybe Literal → Result:
  • It speaks the W3C Entity Reconciliation Service API to any endpoint; Wikidata's is one of them.
  • The manifest's identifierSpace resolves the candidate ids to URIs.
  • The query goes as the protocol's queries form field, with an optional type id and limit.
  • The Result binds entity, name, description (when the service gives one), score (xsd:decimal) and match (xsd:boolean), best first.
  • Errors: a manifest without an identifier space, or a response that isn't JSON, raises ValueError; a non-2xx answer propagates as transport, as for GET.
• Python:
  • operations/reconcile.py, a READ operation.
  • ReconciliationClient in client.py, with the same constructor as the other clients.
• Docs: the README catalog, the planner's prompts/system.md with a runnable example, and an Unreleased CHANGELOG entry.

Tests

• tests/unit/test_reconcile.py (7) covers:
  • rows with their ids resolved;
  • the type and limit sent with the query;
  • JSON arguments;
  • an empty result;
  • a missing identifier space;
  • type checks;
  • that the operation reads.
• tests/http_stub.py answers the protocol offline, as it answers SPARQL: a manifest for a JSON GET, and no candidates for a POST of queries. test_system_prompt routes a reconciliation POST to that default handler.
• The suite passes with uv run --with pytest python -m pytest tests -m "not network": 1120 passed.
• Live against Wikidata: "Aarhus University" restricted to universities (Q3918) returns Q924265, score 100, match true.

Also: GET asks for N-Triples, then Turtle

GET now ranks the RDF media types it accepts: application/n-triples, text/turtle;q=0.9, application/ld+json;q=0.8, application/rdf+xml;q=0.7.

• The bug: unranked, a server choosing among equals picked freely, and Wikidata picked RDF/XML.
• Why that fails: Wikidata writes its rdf:nodeIDs as hashes, and most of them start with a digit. RDF/XML requires XML NCNames, which can't, so rdflib refused every Wikidata entity: rdf:nodeID value is not a valid NCName. The same blank nodes are fine in N-Triples and Turtle.
• Upstream: reported as T252731 in 2020 and declined in favour of T244341, which is still open. w3c/rdf-xml#127 proposes relaxing the RDF/XML grammar.
• What it unlocks: Reconcile followed by GET of the chosen ?entity now works in Python, as it already did in the Java executor (Jena is lenient).
• Checks: a new test in tests/unit/test_get.py covers the header, and a live GET of http://www.wikidata.org/entity/Q25319 returns 7060 statements, coordinates included. The suite passes: 1121.

🤖 Generated with Claude Code