refactor: shared bulk processor pool sized by cluster CPU count

[oblodgett]

Move RoutedBulkIndexer pool from per-species to a shared pool managed by SourceDocumentCreationManager. Pool size is dynamically calculated from the ES cluster's total available CPUs * 2, queried via the new _nodes/os endpoint on ElasticSearchInterface.

JSONProducer work buckets now use byte-size limit (1MB default) instead of document count, configurable via VARIANT_SOURCE_DOCUMENT_CREATOR_JSON_QUEUE_BUCKET_SIZE.

Shared jsonQueue capacity scales with active species count (250 * species).

Move ElasticSearchInterface from agr_indexer to agr_java_core for reuse across modules, add getNodesOs() method for cluster info queries.

[claude]

Code Review

This is a clean refactor that moves the bulk indexer pool from per-species to shared management, with dynamic sizing from ES cluster CPU count. The changes look correct.

One minor observation (non-blocking):

In SourceDocumentCreationManager.java, the jsonQueue drain loop (while (!jsonQueue.isEmpty()) { Thread.sleep(1000); }) followed by interrupt()+join() relies on RoutedBulkIndexer completing its current in-flight batch after being interrupted. This is the same pattern that existed before (just hoisted up a level), so it's not a regression — but worth being aware that correctness depends on RoutedBulkIndexer.run() finishing its current bulk request before honoring the interrupt.

Everything else looks good:

• JSONProducer byte-size bucketing with proper flush on interrupt is correct
• getClusterCpuCount() has appropriate try-catch with sensible fallback
• Thread-confined workBucket/workBucketBytes locals avoid concurrency issues
• ElasticSearchInterface move to agr_java_core for cross-module reuse makes sense
• Import in LiteratureIndexer updated to match the new package location