[benchmark]: Add multi-retriever benchmarking, Nova 2 Lite support, and PGA bio/stat split

[oussamahansal]

Description

Adds multi-retriever benchmark infrastructure, fixes Nova 2 Lite extraction compatibility,
and introduces PGA bio/stat split evaluation.

Changes

• Fix null handling in batch extractors for Nova 2 Lite (returns None for some docs)
• Add topic-beam-chunk_only and topic_beam_search retrievers to benchmark harness
• Fix topic-beam-chunk_only to use SemanticGuidedChunkRetriever (was silently failing)
• Add 3K token budget (per_topic_cap) and increase max_search_results from 5 to 10
• Add PGA bio/stat split evaluation support (BENCHMARK_DATASET env var)
• Add multi-retriever loop (run_all_retrievers.sh) for single-command full benchmark
• Fix boto3/botocore version pins to use ranges for notebook compatibility

Problem

Related issue (if any): #

Testing

• Unit tests added/updated
• Integration tests added (as appropriate)
• Existing tests pass (pytest)
• Tested manually (describe below)

Checklist

• Code follows existing style and conventions
• License headers present on new files
• Documentation updated (if applicable)
• No breaking changes (or clearly documented)

---

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[github-actions]

Lexical Graph Coverage Report: The coverage is at 61.9% (target: 80%). Download the HTML report here.

[mykola-pereyma]

A few questions:

1. Three concerns bundled in one PR — The Nova 2 Lite null fix is a library-level change, the benchmark infra is test tooling, and the boto3 range is dependency management. Would it be possible to split into separate PRs? Makes review/revert cleaner if something regresses.

2. See inline comment on on-start.sh about the boto3 version range width.

3. See inline comment on retriever_factory.py about the max_search_results scope change.

4. batch_topic_extractor_sync.py is missing a newline at end of file — minor but may trigger linting.

[github-actions]

Lexical Graph Coverage Report: The coverage is at 61.9% (target: 80%). Download the HTML report here.

[github-actions]

Lexical Graph Coverage Report: The coverage is at 61.9% (target: 80%). Download the HTML report here.