fix(search): frame the root subject, not the first graph node

[ddeboer]

Problem

frameByType groups each root subject with the triples of the one-hop nodes it references, then yields framed['@graph'][0]. When a root subject references another subject of the same root type, jsonld.frame({'@type': rootType}) returns several root nodes and [0] can be the referenced one – so it is emitted twice and the referencing subject is dropped entirely.

This is not hypothetical. In the NDE Dataset Register a dataset can list a terminologySource (or other reference) whose IRI is itself a separately registered dcat:Dataset. On the full production corpus this silently dropped about 1% of datasets – each dropped one replaced by a duplicate of the dataset it referenced – with no error, because both the projection and the Typesense import still succeed.

Fix

Thread the root subject IRI through groupByRoot and yield the framed node whose @id matches that root, instead of blindly taking [0].

Validation

• New regression test: a root subject that references another root-typed subject – both are projected, and the referenced one is not duplicated.
• Reproduced against the full production Dataset Register read: before, projectGraph yielded 2625 documents but only 2602 unique (23 dropped); after the fix, 2625 / 2625 / 0 duplicates, and the previously missing datasets (e.g. the Gouda Tijdmachine knowledge graph) are present.

Local vitest/tsc could not run in my worktree (its shared node_modules predates the @tpluscode/rdf-ns-builders dependency on main); CI runs the suite with the correct dependencies.