feat(db): IVFFlat index on episodes.embedding [STORY-014]

[khoks]

Summary

• Adds the cosine-distance IVFFlat index that EPIC-014 (RAG / agent memory) will need to build nearest-neighbour queries on top of. Declared in the Drizzle schema (so drizzle-kit owns it) and migrated as 0001_clammy_sir_ram.sql.
• Cosine ops (vector_cosine_ops) because OpenAI text-embedding-3-* and Voyage voyage-3 return L2-normalised vectors. lists=100 is a sensible cold-start default — the index is mostly useless on an empty table; REINDEX after the first batch of episodes lands.
• Adds 1 introspection unit test asserting the index exists with method: "ivfflat".
• Closes STORY-014. Schema-level pgvector work is now complete; retrieval logic comes in v1 under EPIC-014.

Test plan

• pnpm typecheck — clean.
• pnpm lint — clean.
• pnpm test — 20 schema tests pass (was 19; added the IVFFlat index assertion).
• pnpm format:check — clean.
• pnpm --filter @learnpro/db db:generate produces a 1-line CREATE INDEX IF NOT EXISTS ... USING ivfflat ... migration.
• End-to-end migration apply against real Postgres: deferred to STORY-005 (auth shell / smoke harness).

🤖 Generated with Claude Code