The import phase that rebuilds measured relations and writes every edge a provider asserts.

The two paths compose rather than one replacing the other: a provider whose areas nest spatially or by code returns :rebuild from relations/1, and a provider whose source carries a hierarchy no overlap test can derive -- a parent column on every row, or a source with no geometry at all -- returns edges from asserted_relations/2. A provider can do both, and both land in the same release.

rebuild_relations/1 runs first, so a rebuilt measured relation is on the release before any asserted edge is written. If a provider asserts an edge for a pair the rebuild also measured, the assertion wins: it upserts onto the same row, nulling the measured overlap columns and overwriting the geometry-classified relation_type with no warning that a measurement was replaced. A well-behaved provider keeps its asserted edges disjoint from the pairs its own geometry would measure.

Asserted edges are written after normalization, so an edge may name any area the run produced rather than only the ones its own row emitted -- GeoGenius.Catalog.put_relation_many/4 requires both areas to already be members of the release, which every area GeoGenius.Pipeline.Normalize wrote already is by the time this phase runs.

Staging is read in pages, and each page is one heartbeat, the same as GeoGenius.Pipeline.Normalize.normalize/1: the import this exists to serve stages tens of thousands of rows, and before asserted edges existed this phase was a single SQL call under :timeout -- streaming rows with no heartbeat between them would let a long asserting pass outrun the run's lease with nothing renewing it.

A provider's own defect -- an edge naming an area absent from the release, an unknown relation_type -- raises out of Catalog.put_relation_many/4 and is left to propagate. It is not rescued here: GeoGenius.Pipeline catches whatever a phase raises and records it as {:exception, exception, stacktrace}, the same structured detail every other phase's provider call is recorded with. Catching it again here first would only discard the exception's type and stacktrace in favor of a flatter string.