Close the derive hole: make path the whole statement of what the projection reads

[ddeboer]

Close the derive hole: make path the whole statement of what the projection reads.

Problem

derive receives the framed node and reads arbitrary predicates through irisOf / literalsOf / firstLiteralOf – exported, per project.ts, “so derivations can read arbitrary paths”. So SearchField.path is not a complete statement of what the projection reads.

In the Dataset Register’s Dataset type, 10 predicates are reachable via path and 13 only from inside a closure: class, format, conformsTo, additionalType, validUntil, iiifEntities, manifestsSampled, manifestsValidated, quadsValidated, schemaApNdeConformant, subjectUrisSampled, subjectUrisResolved, subjectNamespaceDurable. More than half.

Two consequences:

• A schema-driven extraction generator (Generate extraction CONSTRUCTs from the search schema #548) would emit a graph missing every one of them – silently breaking the class facet, the format facet, status and all five compatibility booleans.
• Generate extraction CONSTRUCTs from the search schema #548’s acceptance criterion “contract test fails on schema / CONSTRUCT drift” is unachievable, because the test cannot know what a closure reads either.

The drift is already real, not hypothetical: register-source.ts CONSTRUCTs dr:dateRead, and no field in the schema reads it. Nothing catches that.

Scope

Split the two concerns: path says what to read from the graph; derive says what to compute from what was read.

• derive: (document) => unknown – drop the FramedNode argument. A derive reads fields declared before it, exactly as status_rank and terms already do.
• Every predicate a derive needs becomes a real field with a path.
• Honour the model’s existing promise for every kind. schema.ts says “a field exposes exactly the roles it declares”, so a field declaring no role (output / searchable / filterable / facetable / sortable) is internal: projected into the document so later derives can read it, pruned before the writer, absent from the collection definition. text already behaves this way (applyText and typesenseFields both no-op); keyword / reference / integer / number / date / boolean write and declare unconditionally. Make them consistent.
• Unexport irisOf / literalsOf / firstLiteralOf.

Acceptance

• The set of field.path over a type is exactly the set of predicates the projection reads.
• A zero-role field of every kind reaches neither the document handed to a writer nor the collection definition.

Consequences

Breaking for @lde/search: consumers migrate derives to internal fields plus document reads.

Net code goes down. The Dataset Register additionally deletes parseNumber, parseBoolean and the qualityMeasurements WeakMap memo – an internal field is projected once and read by N derives, so the memo becomes structural rather than hand-rolled. An internal field costs nothing at the engine: not stored, not indexed, no RAM.

Relates to

Precondition for #548. Design context: #534.