Summary

Adds regression coverage for the PostGIS spatial-filter GeoJSON binding bug tracked in constructive-io/constructive-planning#724. Three layers:

1. ORM integration test — graphql/orm-test/__tests__/postgis-spatial.test.ts (~60 tests, live PG) organised in six sections:

   • A. GeoJSON input-shape binding smoke test — intersects × every GeoJSON shape (Point, LineString, Polygon, MultiPoint, MultiLineString, MultiPolygon, GeometryCollection) × both geometry and geography Point columns.
   • B. Every ORM-exposed operator — all 27 geometry operators + 6 geography operators exercised against seeded data with known-correct expected rows.
   • C. Column-type showcase — one representative query per non-Point subtype (Polygon, MultiPolygon, LineString, MultiPoint, MultiLineString, GeometryCollection, PointZ including intersects3D).
   • D. Combinations — AND/OR/NOT, spatial + scalar, spatial + orderBy.
   • E. Edge cases — isNull on a nullable geometry column, empty polygon input, CRS-qualified GeoJSON.
   • F. Direct shape-match mirrors of the agentic-db fix: bypass grafast middleware in graphile-test to preserve adapted withPgClient #724 regression-guard.

2. Seed fixture — graphql/orm-test/__fixtures__/seed/postgis-spatial-seed.sql — 10 tables (7 geometry subtypes + 2 geography + 1 PointZ) with known city coordinates so expected row IDs are obvious. Loaded via seed.sqlfile([...]) following the mega-seed.sql convention.

3. Structural unit guard — new it(...) in graphile/graphile-postgis/__tests__/connection-filter-integration.test.ts asserting every registered spatial operator spec overrides value binding (via resolveSqlValue, resolveInput, or resolveInputCodec) so GeoJSON goes through ST_GeomFromGeoJSON rather than a raw text bind.

4. CI matrix wiring — .github/workflows/run-tests.yaml: adds graphile/graphile-postgis and graphql/orm-test to the test matrix. These packages were previously not in CI, which is why the regression was able to land on main without tripping any workflow.

No production code is modified. The spatial-filter fix will ship in a follow-up PR.

Updates since last revision

• ORM inflection verified. Ran the suite against a live PostGIS container locally and inspected generated code — codegen pluralises table identifiers, so tables are accessed as orm.citiesGeom, orm.regionsGeom, orm.territoriesGeom, etc. (not singular). Fixed all 10 addresses and pushed as a follow-up commit. Filter field names (loc, shape, regions, path, points, paths, shapes, loc3d, secondaryLoc) are confirmed correct — local + CI both get past the GraphQL layer.
• Observed CI pattern matches expectation. Two jobs fail as designed:
  • graphile/graphile-postgis: 216 passed, 1 failed — the structural guard fails on bboxAbove (first operator alphabetically), confirming no geometry/geography op defines resolveSqlValue / resolveInput / resolveInputCodec.
  • graphql/orm-test: ~53 failed, ~15 passed in postgis-spatial.test.ts. The 15 passes are exactly what I'd expect to pass today without a fix: bbox operators on geometry (&&, ~, <<, >>, etc.), isNull on a nullable geometry column, empty-polygon input, and CRS-hint GeoJSON. Everything that routes through the broken binding path fails — including every operator in Section F and every geography op in Section B.2.
• No new CI failures outside the two target jobs. Adding graphile/graphile-postgis and graphql/orm-test to the matrix did not surface any unrelated pre-existing failures — all other 46 jobs still pass.

Expected CI behavior

This PR is expected to land with failing tests. That is the entire point — the tests codify the contract the follow-up fix must satisfy.

Review & Testing Checklist for Human

• Verify expected-row assertions are correct for the tests that currently fail. The 53 failing tests all report expect(r.ok).toBe(true) → false. That tells me the GraphQL layer errored; it does not on its own prove each failure is the GeoJSON-parse error vs. a wrong expected-ID on my part. Worth spot-checking 3–5 assertions (especially directional-bbox expectations in Section B.1 — bboxOverlapsOrLeftOf, bboxOverlapsOrRightOf, bboxOverlapsOrAbove, bboxOverlapsOrBelow) so that once the fix lands CI actually goes green without needing another test correction round. bboxOverlapsOrLeftOf is one currently-failing bbox-geometry test worth reviewing carefully since it's the only geometry bbox op that doesn't pass today.
• withinDistance argument shape. I used { point: <GeoJSON>, distance: <meters> } per within-distance-operator.ts. Only confirmed by code reading, not by landing a passing test (the fix hasn't landed yet). Worth double-checking.
• Structural guard coupling. The guard is strict (resolveSqlValue OR resolveInput OR resolveInputCodec must be set). If the upstream fix takes a different shape — e.g. changing the codec's toPg method to emit wrapped SQL — the guard needs updating. Flag if that's the direction.
• Seed determinism. Worth a pass on postgis-spatial-seed.sql — especially the GeometryCollection row and the 3D polygon prism used by intersects3D.

Notes

• New file count: 2 (seed + test). Edits: 2 (CI matrix + structural guard test). No source-code changes.
• Estimated suite runtime on CI: ~15s for the postgis-spatial suite after the beforeAll codegen step.
• Seeded row IDs are stable integers (1..6 for cities, 1..4 for regions, etc.) so assertions read as simple arrays.

Link to Devin session: https://app.devin.ai/sessions/c5eeee65a3c546c4ac6753bb05fa03e0
Requested by: @pyramation