[FEATURE] Add S3 doc-store option to WikiHow extraction benchmark

[noel-improv]

Description

The extraction benchmark hardcoded FileBasedDocs (local disk), so it could not measure the cost of S3-backed document storage. This adds a BENCHMARK_DOC_STORE switch in the shared run_benchmark_extract path, so any dataset (ConcurrentQA, WikiHow, PGA) can use it. It defaults to local disk, so with the variable unset every dataset runs as before. The scale runs below use WikiHow.

Changes

• benchmark_extract.py: read BENCHMARK_DOC_STORE (file default, or s3). When s3, use S3BasedDocs with BENCHMARK_S3_JSONL selecting the per-chunk or JSONL path, and collection_id=None (timestamp per run) so repeated runs land in separate collections. Log the resolved store, flag, and collection id.
• benchmark_extract.py: when s3 is selected, validate AWS_REGION_NAME/S3_RESULTS_BUCKET/S3_RESULTS_PREFIX are set and raise a ValueError naming any that are missing, instead of a bare KeyError (these vars are not otherwise required in non-batch prototype runs).
• build-tests.sh: propagate BENCHMARK_DOC_STORE and BENCHMARK_S3_JSONL into .env.testing so they reach the notebook.

Problem

benchmark_extract built FileBasedDocs unconditionally, so there was no way to benchmark extraction against S3. The switch lives in the shared helper, not one dataset's test, because the document store is a generic NodeHandler swap that any run should be able to opt into. The build-tests.sh change is needed too: without it the variables stay in the caller's shell, the notebook never sees them, and the run silently falls back to FileBasedDocs.

Related issue (if any): #

Testing

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

Ran the WikiHow 5,000-doc extraction on a SageMaker notebook across all three settings: file (33m 33s), s3 with BENCHMARK_S3_JSONL=false (42m 21s), and s3 with BENCHMARK_S3_JSONL=true. Each run logged the expected Doc store: line and produced the expected object layout (16,536 per-chunk objects vs one JSONL per source document). The default (file, unset) reproduces the prior behaviour unchanged.

Checklist

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

No new files (existing headers unchanged). Default behaviour is identical to before when the new variables are unset.

---

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