Record embedder model/dimension and stub re-embed migration

[sid-ak]

Category: Model Coupling & Portability
Priority: Medium
Phase: Phase 3 (Embeddings + dense retrieval)

Problem
Swapping the embedder later requires a full re-embed plus schema migration, since the dimension is baked into vector(N).

Solution
Two changes, together making an embedder swap additive rather than in-place.

• Dimension as an explicit parameter
• Per-model embedding columns. Each active embedder gets its own vector(N) column and its own index, tagged by a model identifier.
• See 0008-per-model-embedding-columns.md

[sid-ak]

0008-per-model-embedding-columns.md

[sid-ak]

Removed from Phase 1 (#2) and kept as Phase 3 work, where it originally sat.

Rationale: ADR 0008 exists to make an embedder swap additive instead of a destructive rebuild of the shared vector column — but that cost only bites once there are rows and a migration history to preserve. Right now there are neither: initialize_schema() is CREATE ... IF NOT EXISTS, there is no data, and the standing rule for schema changes is docker compose down -v && up. So reshaping the schema later is free.

It also has no consumer yet. No embedder exists (Phase 3) and no retrieval reads the vector (Phase 4), so building the registry, per-model columns, current-model marker, and re-embed migration now would be designing the swap machinery ahead of everything that uses it — speculative and likely reworked once the embedder interface is real. Implementing 0008 in Phase 3, against an actual embedder, is both correct and zero-migration-cost.

Phase 1 keeps the single embedding vector(1024) column as an intentional pre-0008 placeholder. The drift is already tracked in packages/rag_core/AGENTS.md (Known drift from ADRs, ADR 0008 bullet), which stays as the reminder.