Interface GlobalDocumentReference<TType>

The engine's reference value — deliberately NOT a lake reference, and spelled differently on purpose (id/type, never _ref/_id):

• id is a URI, not a bare doc id, because workflow references cross resources (another dataset, a media library) and a bare _id is only meaningful inside one dataset. Everything resource-aware — client routing, spawn discovery, guard target resolution — reads the resource out of this URI. It is stripped to a bare lake id at exactly one boundary: paramsForLake (core/params.ts), when a query actually runs against a lake.
• type carries the target's schema type so consumers don't need a fetch — as stamped by the ref's constructor, never verified against the target. Engine-built refs stamp it from real data (a projection row's _type, the releaseRef constructor); a caller-supplied ref's type is a trusted claim. Three advisory surfaces read it: the declared-subject-type contract (an entry's types: [...] accepts/rejects incoming refs — refTypeIssues at start initialFields and spawn with), start applicability ("which definitions accept a doc of this _type" — acceptsDocumentType), and guard match.types inference from resolved targets (resolveMatchTypes). All must work when the target lives in a resource the reader can't (or shouldn't) fetch from.

In rendered GROQ this shape is what a reference-held value exposes ($fields.<entry>.id / .type) — whereas _id/_type belong to documents you hold (lake rows, $row, a dereferenced singular doc.ref). The spelling is not a namespace guarantee: a content document can carry its own id/type fields — what you're reading follows what the entry's declared kind puts in your hand, never the key's spelling.

TType preserves the type literal the constructors (refDataset, gdrRef, …) were called with, so a constructed ref satisfies literal-typed targets (e.g. a release.ref value's "system.release").