feat: replace _extract_keywords with LLM-generated tags in metadata node

[wmobley]

Background

The metadata analysis node currently uses a hardcoded keyword candidate set (_extract_keywords) to populate CKAN tags. This set is domain-specific and misses most real-world vocabulary (e.g. drone, photogrammetry, lidar, ODM).

The function was partially removed in the ODM metadata fix PR — the direct call in _guess_dataset_metadata is gone. The function itself remains in use by _text_metadata and _json_metadata to surface candidate keywords as evidence for the LLM.

What to explore

Compare these approaches for tag generation in _prompt_guided_metadata:

1. Current (partial): LLM receives _extract_keywords matches as evidence in the file report; LLM generates final tags from that evidence + user message + file content preview.
2. Full removal: Remove _extract_keywords from _text_metadata / _json_metadata entirely; let the LLM generate tags purely from the file previews and user context.
3. Domain-vocabulary seeding: Pass a domain vocabulary (e.g. from the schema or organization) to the LLM prompt rather than a hardcoded code-level list.

Goal

Find the approach that produces the most accurate, specific, and useful tags for diverse dataset types (geospatial, photogrammetry, hydrology, notebooks, etc.) without missing common terms.

Acceptance criteria

• Benchmark tag quality on 3–5 representative datasets (ODM outputs, MODFLOW notebooks, tabular CSV)
• No hardcoded candidate list remains in the production code path
• Tags are generated or filtered by the LLM, not string matching