fix: write geography columns explicitly instead of inferring points from value shape

[joaquimds]

Problem

An Airtable data source whose records have columns named lat and lng failed to import with invalid input syntax for type json. The PointPlugin converted any object with lat and lng keys into a PostGIS point before every query, so the whole JSON row was turned into SRID=4326;POINT(...) and rejected by the jsonb column. This affects any data source type with those two column names.

Fix

Geometry is no longer inferred from a value's shape.

• PointPlugin now only parses geography values on read (unchanged behaviour).
• Geography columns (data_record.geocode_point, placed_marker.point, geocode_cache.point) are typed as GeographyColumn<...> so their insert/update type is a SQL expression. A raw { lat, lng } object is a type error at the write site.
• New helper toGeography(point) builds the ST_SetSRID(ST_MakePoint(...), 4326)::geography expression.
• upsertDataRecords and upsertPlacedMarker wrap the point internally, so callers are unchanged. The two direct geocodeCache writes wrap it themselves.
• Turfs and areas already used explicit expressions and are untouched.
• AGENTS.md updated to describe the rule.

Tests

• New tests/feature/geography.test.ts: regression test importing a CSV with lat/lng columns, and a round-trip test for toGeography on both geography columns.
• Import, geocode cache, geojson API and data record tests pass.
• The original failing Airtable source imports all 142 records locally with the fix.

🤖 Generated with Claude Code