Revision-controlled corpus-cfg.TOML for RAG deployment + chunking/granularity flags

Read #349, then inspect backend/generate_metadata_jsonl.py, backend/create_datastore_gcs.py, backend/create_app_gcs.py, and constants.py to map the existing arguments and location setting. Verify how the proposed TOML schema should represent each RAG tool and how CLI overrides interact with it. Done means the documented corpus-cfg.TOML is consumed by default, the listed flags work as overrides, and --datastore-id is repeatable.

Follow-up from #349. yangm2 raised that as we accumulate multiple RAG tools with unique settings, reproducible deployment should be driven by a checked-in config file rather than ad-hoc CLI args.

Proposed config (backend/corpus-cfg.TOML or similar): revision-controlled settings for each RAG tool — bucket list, datastore ID, app ID, chunking strategy, granularity, annotation handling, location.

RAG experimentation flags to wire up (from yangm2's review table in #349):

Impact	Experiment	Script	Arg
Highest	Section-boundary chunking	generate_metadata_jsonl.py	--granularity [file|section] — switches from one entry per .txt file to one entry per 90.XXX section
High	Cross-reference expansion	generate_metadata_jsonl.py	--expand-refs — resolves inline ORS X.XXX citations, appends target section text to each entry's struct_data
Medium	Chunk size / overlap	create_datastore_gcs.py	--chunk-size INT and --chunk-overlap INT passed into DocumentProcessingConfig (pairs with --granularity file)
Medium	Annotations datastore separation	generate_metadata_jsonl.py + create_datastore_gcs.py	--exclude-annotations on metadata generation; run twice to produce two metadata.jsonl files; --datastore-id on create_app_gcs.py made repeatable (action="append") so both datastores attach to one app

Most of these naturally live in the TOML once it exists, so they should land together (or the TOML should land first and these flags become overrides).

Also covers yangm2's comment on DEFAULT_VERTEX_AI_SEARCH_LOCATION in constants.py — long term that constant should be sourced from the TOML rather than a Python constant.

Acceptance

• corpus-cfg.TOML exists in the repo with documented schema.
• generate_metadata_jsonl.py, create_datastore_gcs.py, create_app_gcs.py consume the TOML by default; CLI flags above act as overrides.
• --datastore-id on create_app_gcs.py is repeatable.

Refs: #349