Is your feature request related to a problem? Please describe.

cuvs_bench has a pluggable backend system but currently only supports C++ executables (cpp_gbench). Users need to benchmark (1) cuvs-lucene (Lucene connector for cuVS CAGRA) and (2) Elasticsearch GPU (cuVS-backed HNSW over HTTP)--two different runtimes that both use cuVS under the hood. There's no unified way to benchmark these alongside the existing C++ algorithms.

Describe the solution you'd like

Add two pluggable backends to cuvs_bench:

1. CuvsLuceneBackend -- Subprocess to run the cuvs-lucene JAR (java -jar ). The JAR performs build and search; cuvs_bench parses output into BuildResult/SearchResult. Config via jar_path (or CUVS_LUCENE_JAR), dataset, dataset_path. Supports investigations into concurrent searches on cuvs-lucene.

2. ElasticBackend -- HTTP client to talk to an Elasticsearch GPU instance. Create index, bulk index vectors, run kNN search, compute recall. Config via host, port, HNSW params (m, ef_construction, num_candidates). Reuse logic from cuvs-bencher.

Both implement BenchmarkBackend, have ConfigLoaders, return BuildResult/SearchResult, and plug into the existing orchestrator. Registration keys: cuvs_lucene and elastic.

Describe alternatives you've considered

• HTTP for cuvs-lucene: Rejected--cuvs-lucene is a JAR, not a server; subprocess matches the CppGoogleBenchmarkBackend pattern.
• JAR for Elasticsearch: Rejected--Elasticsearch runs as a service; benchmarking requires an HTTP client, not a JAR.
• Separate tools (e.g. cuvs-bencher): Guidance is to contribute backends to cuvs_bench instead of maintaining a separate SDK.

Additional context

• PRD: CuVs-smoke/docs/benchmarking/CUVS-LUCENE-BACKEND-PRD.md
• cuvs_bench structure: backends/base.py (BenchmarkBackend ABC), backends/cpp_gbench.py (subprocess pattern), orchestrator/config_loaders.py (ConfigLoader ABC). Base class already supports network backends (requires_network, initialize(), cleanup(), _check_network_available()).
• cuvs-bencher: ElasticsearchBackend.run_benchmark() in packages/cuvs_bencher_elastic/ provides reusable ES HTTP benchmarking logic.
• CAGRA params for cuvs-lucene: graph_degree, intermediate_graph_degree, itopk, search_width (see config/algos/cuvs_cagra.yaml, bug_issue_93_reproducer.cu).
• Elasticsearch HNSW params: m, ef_construction, num_candidates (see ES-GPU-API-REFERENCE.md).