feat(pipeline-void): merge namespace-alias variants into one VoID partition

[ddeboer]

Adds a namespaceAliases option to the VoID stages that merges http:// and https://schema.org/ variants into a single partition per canonical class/property. Fixes the duplicate void:classPartition reported in netwerk-digitaal-erfgoed/dataset-knowledge-graph#334.

Problem

Partitions are keyed on an opaque MD5(class[, property[, …]]) IRI, so a dataset that types entities under both http://schema.org/CreativeWork and https://schema.org/CreativeWork produces two void:classPartition nodes that, once the class IRI is normalized, describe the same class. Consumers that group by class URI break on the duplicate (it blanked the Dataset Register browser's detail page). Consumers can't cleanly fix it from the published summary either: void:distinctObjects can never be merged from the counts alone (object sets overlap), and void:entities can only be summed under a subject-disjointness assumption. The analysis is the right place to normalize — it has the raw object values to dedupe distinctObjects, and it keeps the duplicate partition nodes from ever being published — but for the summed measures it makes the same disjointness assumption a consumer would.

Approach

Merge after aggregation, keeping the analysis queries plain:

• mergeNamespaceVariants — a per-stage QuadTransform that buffers a stage's VoID output, re-mints each partition IRI from its canonical key components (replicating the queries' MD5(CONCAT(STR(…))), pinned by a test against the real #334 hashes), collapses the duplicate partition nodes, and sums void:entities / void:triples. Attached to the class and property partition stages.
• The class selector canonicalizes and de-duplicates its bindings, and the reader expands each canonical class back to its namespace variants, so a class's variants are co-located in one batch — the merge is self-contained per stage, no pipeline-core change needed.
• The three void-ext queries emit their cp → pp parent chain (void:class, void:property) so the transform can re-key them from an otherwise opaque hash.
• class-properties-objects.rq keeps query-time normalization — its void:distinctObjects is a distinct-count over object values that overlap across namespace variants (two works can share a name), so summing would over-count. It's deduped in the query and never reaches the transform. The integration test pins this: a shared object value gives distinctObjects = 3 where a naive sum would give 4.
• NamespaceAlias moves to @lde/pipeline with shared canonicalizeIri / aliasVariants helpers; re-exported from @lde/pipeline-shacl-sampler.

Assumptions

Summing pre-aggregated counts is exact only under subject/class disjointness (class-partition and triple sums) and predicate-namespace disjointness (property-partition entity sum). Both hold for the schema.org datasets that motivate this — the variants appear as disjoint subsets. A resource typed under both variants over-counts its class entities; this is documented on the transform, in ADR 7, and pinned by a test. With no aliases configured the transform is a no-op and behaviour is unchanged.

Verification

Unit tests for the transform (incl. the MD5-minting coupling and the malformed-input guards) plus an endpoint-backed integration test over the full VoID output of a mixed-namespace fixture: one merged class partition (4 entities), one property partition (4 entities, distinctObjects 3), merged datatype/language/object-class partitions, and the documented over-count under a disjointness violation.

Follow-up (dataset-knowledge-graph, separate repo): pass namespaceAliases: [{ canonical: 'https://schema.org/', alias: 'http://schema.org/' }] to voidStages() and drop the write-time schemaOrgNormalizationPlugin rewrite.