Import should promote matching extracted properties instead of creating duplicates

Problem

After the PropertyReference refactoring, dedup between extracted and imported properties only works in one direction.

Extraction → existing property (works): When enrichment.py extracts a property, it calls Property.find_matching() to check if a matching property already exists (whether from Wikidata or a previous extraction). If found, it calls add_reference() to attach the new source to the existing property instead of creating a duplicate.

Import → existing extracted property (broken): When a Wikidata re-import brings in a statement that matches an existing extracted property, the importer creates a new Property row. This happens because the batch upsert conflicts on statement_id, and extracted properties have statement_id = NULL — so there's no conflict, and a new row is inserted. Previously soft_delete_matching_extracted() handled this, but it was removed in the PropertyReference refactoring.

The result: if a politician has an extracted property (e.g. birth date = 1960, no statement_id) and Wikidata later adds a statement for that same fact, you end up with two Property rows for the same fact.

Proposed fix

Use find_matching() during import as well. In _insert_politicians_batch, before batch-upserting each property, check for a matching extracted property. If one exists, promote it: set statement_id and references_json on the existing row. This preserves its PropertyReferences (the extraction sources) while marking it as Wikidata-backed. Promoted properties are then excluded from the batch_upsert.

This makes find_matching() the single generic dedup mechanism used in both directions:

• Extraction finds existing Wikidata property → add_reference()
• Import finds existing extracted property → promote it (set statement_id, references_json)

Start in enrichment.py and trace the import path through _insert_politicians_batch, then inspect Property.find_matching() and the batch upsert behavior. Done means a matching extracted property is promoted without creating a duplicate, its existing PropertyReferences remain intact, and promoted properties are excluded from the batch upsert.