GraphQL approach had broken pagination (totalCount key missing, crashes
silently). S3 approach at least completes cleanly even if the metadata
table currently lacks a description column.
- Replace S3 bigquery_tables metadata lookup with paginated GraphQL API call
to fetch table and column descriptions from Base dos Dados
- Add gera_schemas.py for schema compilation and S3 inventory
- Add schemas.json and file_tree.md as generated reference artifacts
- Add websocket proxy in Caddyfile for ttyd on port 7681
- Ignore generated context/ artifacts in .gitignore
- Add openai to requirements.txt
