dbt & Modern Data Stack

The dbt adapter emits both the `schema.yml` declarations (sources, models, columns, tests, descriptions) and the `models/*.sql` files for each Magellan mapping. Snowflake, BigQuery, Redshift, Databricks, and Postgres dialects are supported; the SQL dialect is selected by the target connection profile, not hard-coded in the adapter.

Every emitted `.sql` file carries a header comment with the source mapping ULID, the `ApproverReleaseSigned` entry ULID, the SHA-256 chain head at export time, and the SHA-256 of the canonical mapping payload. A `dbt parse` step in CI re-reads this header and refuses to compile if the chain head is older than the latest published mapping version — which is the cheapest possible way to catch a stale spec before it touches the warehouse.

Column-level lineage is emitted as dbt `meta` blocks under each column, so dbt's own lineage graph and Magellan's Automated Lineage view stay aligned. The annotations include the source schema ULID, source attribute path, and the algorithm class that generated the mapping (exact_match, synonym_match, abbreviation_match, context_hierarchy, domain_equivalence) — so a dbt user can see not just where a column came from, but on what basis it was matched.

Tests defined in Visual Rules for Targets — uniqueness, not-null, accepted-values, ranged-numeric — are emitted as native dbt tests in `schema.yml`. Custom rules emit as `tests/*.sql` files. The tests are the same tests Magellan validates against during draft; dbt is not running a parallel ruleset.

Each export writes `ArtifactExported` carrying the dbt project identifier, the artifact format (`dbt-yaml`, `dbt-sql`, `dbt-tests`), and the SHA-256 of the emitted bytes. A `dbt build` failure can be traced back to a specific `ArtifactExported` ULID, which can in turn be traced to the `ApproverReleaseSigned` entry that authorized it — the trail from warehouse failure to signed Magellan decision is two hops.

Re-runs are byte-deterministic. The same mapping ULID against the same dbt target connection profile produces a byte-identical project. Pull requests that touch generated dbt files should always be empty diffs unless a Magellan release has happened in between; this is enforceable as a CI rule.