tars-location-mcp: Instructions file for Codex

Repository instructions

• Never commit a real position, address, place name, token or connection string. Tests use synthetic coordinates; the two in the fixtures are city centres, not anyone's home.
• A migration reviewed by reading is a guess. Every schema change runs against a real Postgres before it lands: docker run -d --name location-test -e POSTGRES_PASSWORD=test -p 5433:5432 postgres:16, then LOCATION_TEST_DATABASE_URL=... pytest.
• Migrations are additive and idempotent. create ... if not exists and create or replace, because upgrading an archive someone already filled is the normal case.
• The core schema never references people. That join lives in 0002_people_bridge.sql, which refuses to install without one.
• A duration is measured, never modelled. If a change makes the code invent a boundary between two fixes, it is the wrong change.
• "No record" is not "was not there". Anything that returns empty points at location_coverage, and the gaps it reports are computed from the data rather than written down.
• location_sql stays read-only enforced by Postgres, not by prompting. Do not add a path that runs caller SQL outside the read-only transaction.
• A chance of rain is counted over ensemble members, never lifted off a single deterministic run, and the weather panel holds one model per independent centre. Adding a regional model that falls back to someone else's global run makes the panel report agreement it has not got.
• psycopg is the only runtime dependency. The geocoder, the HTTP endpoint and the JSON-RPC server are standard library and stay that way.
• Run ruff check . and pytest after Python changes.