[ddeboer]

Fix #1065

Extraction put 10 roots in each CONSTRUCT, whatever the source. On a QLever import, 100 roots per query ran the CreativeWork stage 5.5 times faster. A live endpoint may be shared with other users and fail on big queries (#838), so it keeps 10.

@lde/pipeline

• StageOptions.batchSize can take one value per source kind: { endpoint, imported }.
• The pipeline passes each stage RunOptions.source. It is imported when the resolver (or the fallback) imported a dump, and endpoint otherwise.
• A stage whose sourceFor swaps in another distribution counts as endpoint. A sourceFor that hands back the resolved distribution unchanged keeps its kind; registrySource does that for a file: dataset. A chained child counts as imported, because it reads its parent's output from a store the pipeline serves itself.

@lde/search-pipeline

• searchStages defaults batchSize to { endpoint: 10, imported: 100 }.
• searchIndexerPipeline has two new options: tuning (batchSize, maxConcurrency) applies to every type, and tuningByType overrides it per type name. The per-type override is merged field by field. A name the schema doesn't declare throws.
• With these, per-stage tuning is no longer a reason to compose the pipeline by hand.

@lde/search-indexer

• Three new variables: EXTRACTION_BATCH_SIZE, EXTRACTION_CONCURRENCY and EXTRACTION_TUNING (JSON, keyed by type name).
• QLever now starts with --num-simultaneous-queries set to the busiest stage's concurrency, capped at 4. The memory default stays at 4G, which the issue measured to be enough even at -j 8.

@lde/sparql-qlever

• QleverServerOptions gains num-simultaneous-queries.

Breaking: IndexerConfig has a new required extraction field. Search stages that set no batchSize now batch 100 roots from an imported dump.

To decide:

• Set values apply everywhere. As agreed, a batch size you set applies to both kinds of source. That includes registry-sourced types, which read the shared registry endpoint. So EXTRACTION_BATCH_SIZE=200 also sends 200-root queries to the registry.
• Not re-tested. With 100-root batches and -j 4, queries queue at QLever while its 30 s default-query-timeout still runs. The issue's harness ran these sizes without timeouts, but I haven't re-run it.
• Overlaps with feat(pipeline)!: fetch selector roots in pages of 1000 #1079. That PR's selector pages are independent of this one. Together they give the 7.5× from the issue.