RAG vector store has no migration — course_chunks/match_course_chunks exist only as dashboard DDL

[AndresL230]

Split from the #323 audit (verified @ d1d724f). course_chunks, the match_course_chunks RPC, and CREATE EXTENSION vector appear in ZERO .sql files (425 checked — migrations/, seed.sql, archive/). A database replayed purely from python -m db.migrate has RAG silently dead end-to-end: retrieve_chunks swallows the RPC failure into [] (rag_service.py:137-139), _get_catalog_chunk degrades to "", and indexing failures vanish into a fire-and-forget log line — the tutor answers ungrounded with no error, metric, or user-visible signal, and nothing in the suites or oracles asserts the table exists.

Fix: a numbered migration (IF NOT EXISTS, mirroring 0032's reconcile pattern) codifying the extension + table + RPC to match the live dashboard DDL (shape per docs/superpowers/specs/2026-07-03-rag-document-chunking-design.md — verify against staging before writing). This is the highest-leverage remaining RAG item and it is a migration, not a feature.

[AndresL230]

Fixed in #495 (9ca30d4). Closing — backend-only, no UI surface to verify.

Migration 0039 codifies CREATE EXTENSION vector, course_chunks (with its indexes) and the match_course_chunks RPC. Shape verified against live production and staging by reading a real row and calling the RPC — not from the design doc, which describes intent while the database is what the code actually talks to.

Two things that probe turned up, worth recording:

• The live RPC is 768-dim and the code's _OUTPUT_DIM is 768 — they match, so there was no latent dimension bug hiding behind the missing migration.
• Direct psycopg to staging is unreachable from here (the IPv6-only endpoint refuses and the pooler hosts reject), so the probe went through PostgREST.

Every statement is IF NOT EXISTS / CREATE OR REPLACE, mirroring 0032's reconcile pattern — a no-op where the objects already exist.

The ragstore oracle closes the gap this issue names ("nothing in the suites or oracles asserts the table exists"). It checks the extension, the table and the RPC — deliberately not their contents, since an empty course_chunks is normal on a fresh stack and a missing one is the bug.

Proven the only way that counts

A normal e2e cycle would not have caught this: e2e-up runs against a database that already has the objects. So this was verified with a from-empty replay — supabase db reset plus every migration from scratch, then the new oracle against the virgin DB:

RESET_EXIT=0 # every migration replays clean
RAGSTORE_EXIT=0 # 0 findings — vector store present after a pure from-migrations rebuild
JOURNEYS 35 passed
ORACLES 0 finding(s)

Also fixed here (part of #482)

The oracle's NULL-embedding check led back to the cause: index_document_chunks upserted chunks whose embedding never landed. match_course_chunks ranks by vector distance and skips NULLs, so those rows were unretrievable by construction while still counting toward the "indexed N chunks" the caller logs — which is how a total embedding outage read as complete success. They're now dropped, logged, and only real indexing is reported.

Two tests changed because they pinned that bug rather than a contract. The remaining #482 items (silent-empty observability, DBOS durability inversion) are still open there.