@lde/search-pipeline: searchStages, retire projectGraph and the buffering writer

[ddeboer]

Compose per-type projecting stages in @lde/search-pipeline, retire projectGraph and the buffering searchIndexWriter, and prove the memory bound. The scalability landing — ADR 13. Depends on Stage<Out> and projectRoots.

Scope

• searchStages(options): Stage<TypedSearchDocument>[] — one stage per root type. Sketch:

  { schema, types: [{ searchType, rootVariable, itemSelector, readers, batchSize? }], writerFor }

  Each stage's project closure: projectRoots(quads, bindings.map(b => b[rootVariable].value), schema, searchType), tagged { searchType, document }. rootVariable couples the selector's projected variable to what the projection reads — settle this interface here, against a working Stage<Out>.
  • selectByClass(searchType) helper: ?x a <class> — one word for the object grain, where class is the source class. Not a default (three of DR's four catalog types have no source class).
• Move TypedSearchDocument from @lde/search to here: it exists only because a terminal routes, which is glue, not projection.
• searchIndexWriter → Writer<TypedSearchDocument>: keeps ADR 9's per-collection fan-out (route by searchType.class to per-type runs) and the run lifecycle; stops projecting and stops buffering.
• Delete projectGraph, rootsByType, frameByType from @lde/search.

Acceptance

• Memory flat by counting (ADR 12): run a synthetic input at 10 and 10,000 roots, batchSize: 10, against a recording fake Writer; assert project invoked ceil(roots/batchSize) times, and peak live document count identical at both sizes. Fails against today's searchIndexWriter; passes here.
• No behavioural regression: the DR-shaped multi-collection integration test (Typesense testcontainer) — same documents land in the same collections.
• projectGraph gone from the workspace.

Design: ADR 13. Parent: #606.