embeddings: enforce model and source provenance beyond vector dimension

[ragnorc]

Problem

Dimension checks prevent some incompatible vectors, but equal dimensions do not establish that stored vectors and query embeddings share a model, revision or preprocessing contract. Manually supplied vectors can lack enough provenance to detect mixing two same-dimensional embedding spaces.

Shared embedding configuration and query-time model checks already exist where identity is available. The residual gap is persistent, enforceable provenance for stored/manual vectors and their source content.

Proposed change

Define the model/provider revision, dimension, relevant preprocessing identity and source revision/hash associated with an embedding field or value. Specify import/load validation, legacy/unknown provenance, model migration and search behavior when a corpus has mixed or stale vectors.

Do not infer compatibility solely from vector length, and do not silently label arbitrary imported vectors as produced by the currently configured model. Provide a deliberate explicit contract for externally generated embeddings.

Acceptance criteria

• Two distinct models with the same dimension cannot be silently treated as one verified space.
• Source edits and model changes make stale coverage observable and trigger the documented backfill path.
• Import, branch merge, export/reload and restart preserve provenance.
• Query-time and stored-vector identity checks agree; legacy unknowns have a documented migration/refusal policy.

Code: embedding configuration/validation, query-time embedding resolution. This is not a report that all dimensional mismatches are currently accepted.