Summary

Fixes #724. All 26 PostGIS spatial operators registered by createPostgisOperatorFactory() were relying on the default connection-filter resolveSqlValue pipeline, which binds the GeoJSON object as raw text cast to ::geometry / ::geography. PostgreSQL's geometry_in / geography_in parsers do not accept GeoJSON text — they require ST_GeomFromGeoJSON(...) wrapping. Every spatial filter with a GeoJSON input therefore failed at runtime with parse error - invalid geometry.

The change, per operator registration:

• Override resolveSqlValue: () => sql.null to disable the default bind.
• In resolve(...), wrap the input via <schema>.st_geomfromgeojson($1::text), appending ::<schema>.geography when the registration is on a geography-based type.

This is the same pattern already used correctly by within-distance-operator.ts — now applied consistently across the broader operator set.

Unit tests in connection-filter-operators.test.ts are rewritten to assert the new contract: every operator's compiled SQL must contain the schema-qualified st_geomfromgeojson($1::text) wrap, and geography-type registrations must additionally append the ::geography cast.

The full end-to-end regression suite (live PG, 66 tests across all spatial operators × 7 GeoJSON input shapes × both codecs) lands via PR #989 and flips green once this is merged.

Updates since last revision

• Added graphile/graphile-postgis to the CI matrix in .github/workflows/run-tests.yaml so the ST_GeomFromGeoJSON wrapping contract (the 218-test unit suite for this package, now including the new assertions) runs on every PR — previously the package was not in the matrix, which is how fix: bypass grafast middleware in graphile-test to preserve adapted withPgClient #724 slipped through in the first place. All 47 matrix jobs pass on this PR head.

Review & Testing Checklist for Human

• Confirm the geography cast logic is correct — spec.baseType === 'geography' appends ::<schema>.geography after the ST_GeomFromGeoJSON(...) wrap. Verify this produces the geography overload of each ST_* function rather than the geometry one, especially for operators registered on both codecs (intersects, covers, coveredBy, exactlyEquals, bboxIntersects2D).
• Confirm sql.identifier(schemaName, 'st_geomfromgeojson') and sql.identifier(schemaName, 'geography') resolve correctly when PostGIS is installed to a non-public schema — there's a unit test for a postgis schema, but no live integration test in this PR. (PR test(postgis): regression coverage for GeoJSON spatial filter binding (#724) #989 exercises live PG but assumes the default public schema layout.)
• Spot-check a couple of previously-working callers (e.g. anything that was passing a raw WKT string or geometry-typed bind via its own pipeline) to confirm they haven't regressed. The default pipeline is now disabled for all 26 operators.

Notes

• The within-distance operator is not touched here — it already had the correct wrapping.
• During the test audit we discovered a separate pre-existing issue: WithinDistanceInput is registered by graphile-postgis but does not surface on the generated GeometryInterfaceFilter schema type. This is unrelated to the GeoJSON binding bug and is tracked as a follow-up; the two test cases in the regression suite that exercise withinDistance are marked xit with a FIXME trail.
• This PR is scoped to the fix only. The regression test suite lives in PR test(postgis): regression coverage for GeoJSON spatial filter binding (#724) #989, which is designed to land red on main until this fix is merged and then rebased → green.

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