1. config.py — moved _MIN_SIMILARITY into Settings as mcp_default_similarity
   (configurable via .env, default 0.7)

2. models.py — added sort_by to SearchMetadataInput:
   Literal["relevance", "price_asc", "price_desc", "date_desc", "area_desc"]
   Default "relevance". Full Field description with all options.

3. main.py — major refactor:
   - Tool signatures collapsed from 13 individual filter params into
     nested Pydantic models: filters: MetadataFilters, pagination: PaginationParams
     → AI agent now sees rich Field descriptions for every filter field
     through $defs/$ref in generated JSON schema
   - search_by_metadata_tool gains sort_by param
   - Removed manual dict construction (filters={"deal_type": deal_type, ...})
     → cleaner, less error-prone
   - Added SERVER_INSTRUCTIONS to FastMCP(instructions=...) —
     system-level guidance sent to every MCP client (Claude Desktop, etc.)
     covering: server purpose, language, tool selection strategy,
     filter best practices, sorting, fallback, pagination
   - search_guide prompt now reuses SERVER_INSTRUCTIONS (DRY)
   - describe_schema_tool docstring rewritten to be more compelling

4. tools.py — added _build_order_by_clause(sort_by) for metadata search:
   relevance → ts_rank_cd DESC
   price_asc/desc → price ASC/DESC NULLS LAST
   date_desc → publish_date DESC NULLS LAST
   area_desc → total_area DESC NULLS LAST
   Uses settings.mcp_default_similarity instead of hardcoded constant.
   describe_schema updated with sorting section.

5. test_models.py — 12 tests, covers:
   - sort_by defaults and validation
   - nested schema generation ($defs)
   - existing pagination, filter, serialization tests

12 tests pass, ruff clean.

Co-Authored-By: Claude <noreply@anthropic.com>