Automatic MetaModel Conversion

A spreadsheet row is not a mapping. It is a description of a mapping in a particular columnar shape, in a particular dialect of how that team writes transformations. The conversion pass turns that description into a MetaModel mapping object — a typed, hierarchical artifact that participates in Schema & Map Version Control, Automated Lineage, and Impact Analysis without per-row glue code.

The conversion produces a versioned MetaModel **mapping** (`<Mapping Name>`) together with the **source schema(s)** and **target schema(s)** it references — projected from the distinct source paths and target paths in the spreadsheet. A single mapping can reference more than one schema on either side: when the spreadsheet pulls from multiple source systems or lands in multiple target structures, each schema is qualified (e.g. `<Mapping Name> Source Schema — HRIS`) so it stays uniquely addressable. Every artifact produced by one import is written at the **same version number at the moment of import** (typically `1.0.0` on first import). After import, schemas and the mapping evolve **independently** — the mapping pins the specific schema versions it was authored against, and re-importing the same spreadsheet re-establishes a shared version only across the artifacts produced by that re-import.

Source and target paths are resolved against registered schemas using their dot-delimited path notation, including array indices (`orders[0].lines[1].item`). Unresolved references write `MappingImportDeferred`; resolved references write `MetaModelProjected` carrying the resolved schema ULIDs on both sides plus the canonical mapping payload hash.

Transformation expressions are parsed against the MetaModel's expression grammar. Common spreadsheet idioms — `IF(...)`, `CONCATENATE(...)`, `LEFT(field, n)`, `LOOKUP(...)` — are recognized and converted to the equivalent rule node, so the imported mapping participates in Visual Rules for Targets immediately. Unrecognized expressions are preserved verbatim as a `derivation_source` annotation and surfaced for review.

Datatype assertions are computed from the resolved schemas, not from the spreadsheet text. A row that maps `varchar(50)` → `varchar(20)` writes a `TypeNarrowingDetected` entry alongside `MetaModelProjected`, which is what allows Predictive Mapping to flag silent precision loss before a mapping is published.

The projection is a pure function of the spreadsheet bytes, the column-mapping profile ULID, and the registered schema ULIDs at the moment of import. Re-running the projection at the same input ULIDs produces a byte-identical MetaModel — which is what makes spreadsheet-imported mappings defensible under 21 CFR Part 11 §11.10(e) and SOX §404 ITGCs without an asterisk.

From this point forward the spreadsheet is provenance, and the MetaModel is the operational artifact. Approver release signatures (`ApproverReleaseSigned`), CI/CD promotions (`ReleasePromoted`), and downstream ETL exports (`ArtifactExported`) all reference the MetaModel mapping ULID. The spreadsheet ULID is one transitive lookup away in the ledger when an auditor asks where the mapping originated.