The tg_ticket_geom trigger resolved feed coords -> cluster centroid -> none,
never consulting tickets.geo_locations, so every 20-min delta ingest re-upserted
changed rows and downgraded previously-resolved 'location' geoms back to the
cluster centroid. Live effect: only 51 of 114k INC (and 0 of 42k CRQ) rows kept
the precise geocode the LocationIQ budget paid for.
- migration 18: trigger now resolves feed -> geo_locations (precise) -> cluster
-> none, mirroring resolve_ticket_geoms() precedence; ends with one resolve
pass to repair the backlog. Dry-run against the live DB (rolled back) repaired
7,481 rows: INC location 51 -> 5,339, CRQ 0 -> 2,193.
- pipeline.ingest(): re-resolve after every applied run that ingested files, so
geoms self-heal even before migration 18 lands.
- run_ingest.sh: chain an incremental --geocode-clusters pass (0 API calls when
no new clusters) so new clusters map without a manual command (FT-BUG-02).
- Dockerfile/.dockerignore: pinned installs from uv.lock, non-root user (FT-SEC-02).
- 20260618_bug.txt removed (stale review of a since-rewritten file).
Numbered 18 to coexist with 17_drop_unused_geo_indexes.sql (parallel 260702
change). Audit + plan + work log in docs/260702_*. Local only; not applied to prod.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
bb38d354e5