@lde/search-pipeline: composition + object indexing

[ddeboer]

Composition package for the search-as-pipeline design (epic: #534). Depends on the core change, the Typesense sinks, and per-reference label sources.

Scope

• New @lde/search-pipeline package composing Reader -> Stage(s) / Projection -> Writer over @lde/pipeline + @lde/search + @lde/search-typesense. @lde/search stays domain-agnostic and pipeline-free.
• Object-grain wiring: per-entity extraction (select URIs -> CONSTRUCT each, VALUES-bound), bounded VALUES batches; the Projection frames each batch (framing fused, one collection); the Writer loads documents.
• Memory bounded by the VALUES batch size (streaming by reuse; no whole-dataset buffering, no new framer).
• Reuses Skip reprocessing unchanged datasets: source signal + consumer-declared pipeline version #450 skip-unchanged (provenance store scoped by a search-pipeline IRI) and the @lde/pipeline: emit per-distribution validity + record reachability fingerprint #469 distribution health signal.

Acceptance

• End-to-end object index build against a Typesense testcontainer.
• Memory-flatness test: flat memory across a large synthetic input.

Design and rationale: #534.