Epic: native geometry computation core for vortex-spatial

[HarukiMoriarty]

Replace the geo crate as vortex-spatial's computation engine with an in-house core that runs directly over the native columnar geometry layout. geo stays only as a dev-dependency: test oracle and bench baseline.

Status

Active

Motivation

Each row-oriented kernel (ST_Intersects, ST_Contains, ST_Distance, ST_Area, ST_ConvexHull) currently decodes every batch Vortex → Arrow → geoarrow → per-row geo_types::Geometry before computing. Owning the core:

• removes that materialization — measured ~3-4x on ST_Area after the swap;
• puts geometry semantics (degenerate inputs, NaN policy, panics) under Vortex's control;
• aligns with the kernels that are already native (ST_Length, ST_MakeLine, ST_Collect, ST_Envelope).

Approach

• vortex-spatial/src/algorithms/: pure 2-D functions over borrowed slice views (GeometryRef) — no vortex-array types, no per-row allocation.
• GeometryBatch canonicalizes a column once and hands out zero-copy row views; a constant operand is a one-row batch.
• Kernels keep the existing dispatch shells; only the decode bridge and the per-row math change.
• The robust crate supplies exact orientation predicates where a sign test decides topology.

Correctness

• Every algorithm is differential-tested against geo: curated fixed cases (empties, unclosed rings, holes past the exterior, mixed winding) plus proptest properties per geometry family — failures shrink to a minimal counterexample and persist to proptest-regressions/.
• Comparisons are exact, not epsilon: the core mirrors geo's float operation order, so drift fails loudly.
• The oracle is geo as the kernels actually used it, through the geo_types constructors. Phase 1 caught two traps this hides: rings are implicitly closed, and a polygon's bbox is its exterior ring only.
• Pre-existing kernel tests stay green unchanged. Deliberate divergences (e.g. hull ring rotation) get a documented canonical form with their own tests.
• On disagreement, PostGIS is the tie-breaker — geo is not automatically right.

Phases

• Differential test harness: generated native columns vs the geo oracle, covering ST_Length and ST_Distance first — test(vortex-spatial): add differential test harness against the geo oracle #9391
• Core skeleton: GeometryRef views, Aabb, area, bounding-rect; swap ST_Area, the GeometryAabb aggregate, and the prune rules
• Native ST_Distance: point/segment/segment-segment primitives plus point-in-polygon
• Native ST_Intersects: segment intersection and the full type-pair matrix, with a columnar bbox prescreen
• Native ST_Contains: interior/boundary-aware matrix (no generic DE-9IM relate)
• Native ST_ConvexHull: monotone chain, canonical ring form, native polygon output sink
• Remove the geo_types decode path; move geo to dev-dependencies

Out of scope

• Overlay/boolean ops (ST_Intersection, union, difference) stay on geo::BooleanOps; no generic DE-9IM relate.
• Arrow/WKB interop (geoarrow, geoarrow-cast, wkb) is untouched — this Epic covers computation only.