fix(rag): register pgvector adapter on get_connection()

[cipher813]

Summary

• Silent-broken-since-day-one fix found during tonight's 2026-04-19 RAG recovery run.
• `rag.pipelines.filing_change_detection` reads `c.embedding` (pgvector `vector` column) and passes to `np.array(dtype=float32)`. Without `pgvector.psycopg2.register_vector(conn)`, psycopg2 returns the vector as a stringified list and the numpy call raises `ValueError: could not convert string to float`. Registered per-connection (psycopg2 scopes type adapters to the connection, not globally).
• Steps 1-4 (ingest_sec_filings, ingest_8k_filings, ingest_earnings_finnhub, ingest_theses) only write vectors — writes serialize Python lists → pgvector's string format fine. Step 5 is the first reader of vector columns, which is why the bug was latent until tonight's real-run reached it.

Impact

• Every Saturday SF RAG step has been exiting non-zero at step 5 since filing_change_detection was added. Downstream Research was blocked because RAGIngestion SF step failed, even though steps 1-4 had written the embeddings Research actually queries.
• Output `rag/filing_changes/latest.json` is not consumed by any code in alpha-engine-research or alpha-engine-data (grepped both). Pure analytic signal for lazy-filing detection. Bug impact = SF state fire, not trading-decision impact.

Changes

• `rag/db.py` — `register_vector(conn)` after `psycopg2.connect()` in `get_connection()` context manager. One-line import, 5-line comment explaining the per-connection scoping.
• `requirements.txt` — add `pgvector>=0.2` (explicit pin; was transitively available but not declared).

Test plan

• `pytest tests/` — 109 passed
• Post-merge: re-run RAG ingestion end-to-end, verify step 5 (filing_change_detection) completes and writes `s3://alpha-engine-research/rag/filing_changes/latest.json`
• Verify heartbeat `AlphaEngine/Heartbeat{Process=rag-ingestion}` emits (was suppressed by the step-5 crash under `set -euo pipefail`)

🤖 Generated with Claude Code