feat(search-pipeline)!: project inside the batch, per root type (ADR 13)

[ddeboer]

The scalability landing of the search projection: projection moves into per-type pipeline stages and the buffering writer retires, so memory is bounded by batchSize roots instead of the whole dataset. Design: ADR 13.

What changed

@lde/search (breaking)

• projectRoots is now the only projection entry point – one root type over the roots the caller supplies.
• Removed the whole-schema path: projectGraph, frameByType, rootsByType, and buildSubjectIndex's rootTypes parameter (it no longer scans for rdf:type).
• TypedSearchDocument moved out to @lde/search-pipeline; @lde/search yields a bare SearchDocument and stays pipeline-free.

@lde/search-pipeline (breaking)

• searchStages(options) builds one projecting Stage<TypedSearchDocument> per root type: each selects its roots, extracts each root's quads, and projects the root-complete batch (projectRoots) into documents tagged with their SearchType. selectByClass(searchType) is a convenience selector for the object grain (not a default).
• searchIndexWriter is now a Writer<TypedSearchDocument>: it keeps ADR 9's per-collection fan-out and run lifecycle but stops projecting and stops buffering, routing each tagged document straight to its type's engine run.
• A search pipeline is now new Pipeline<TypedSearchDocument>({ datasetSelector, stages: searchStages(...), writers: searchIndexWriter(...) }) – one terminal, N stages.

Memory is flat, asserted by counting

test/search-stages.test.ts runs a synthetic input at 10 and 10 000 roots (batchSize: 10) against a recording fake writer and asserts: project is invoked ceil(roots / batchSize) times; the max quads.length handed to project is identical at both sizes; and the peak concurrently-live document count is identical at both sizes. The old buffering writer's peak was every document, so this fails against it and passes here.

Validation

npx nx affected -t lint typecheck test build --base=origin/main is green (search, search-pipeline, search-typesense, search-api-graphql). The Typesense testcontainer integration test (multi-collection.integration.test.ts) runs against a live container and passes – same documents land in the same collections. Coverage is 100% on the new/changed source; no new lint warnings.

Design choices not spelled out in the issue

• searchStages options omit writerFor. The issue sketch listed it, but stages do not write – the terminal does. searchStages takes only { schema, types }; writerFor stays on searchIndexWriter.
• searchStages re-resolves each type from the schema by class. assertTypeInSchema is an identity check, so the stage projects with the schema's own declaration object even if a class-equal lookalike is passed (covered by a test).
• Added queueCapacity per stage type. ADR 13 motivates it (a document is far heavier than a quad); exposed as an optional knob.
• selectByClass(searchType, rootVariable = 'root') takes an optional root variable so the selector's projected variable matches the stage's rootVariable.

Part of #606
Implements #623