Canon-only RAG: single-pass retrieval, canon DB, port config

[aniongithub]

Summary

Pins the vector DB to the canon-filtered release and simplifies the RAG path to a single-pass retrieve → rerank → generate flow, plus fixes container port handling so the chat client and API always agree.

Changes

• Dockerfile — pin MEMORYALPHA_DB_RELEASE to v0.5.2 (Tier 2 strict canon-only DB; non-canon real-world pages dropped upstream in memoryalpha-vectordb). Text collection is now ~24.8k canon docs (was 37.5k).
• api/memoryalpha/rag.py — single-pass no-tools ask() / ask_stream() (retrieve → cross-encoder rerank → stuff → generate) with real token counts; the legacy tool-calling loop is preserved behind ask_with_tools().
• api/memoryalpha/ask.py — GET/POST /memoryalpha/rag/ask and GET/POST /memoryalpha/rag/stream.
• .env / docker-compose.yml / chat.sh — split container-internal APP_PORT (default 8000) from host-published API_PORT (default 18000), so the published mapping and the uvicorn listen port can't drift. chat.sh targets APP_PORT with a RAG_API_URL override for host use.
• wait-for-ollama.sh — fix empty-model handling when pulling DEFAULT_MODEL.
• README.md — document endpoints, host-vs-container ports, and canon filtering.

Validation

Rebuilt the devcontainer on v0.5.2 and ran live queries:

• DB: `memoryalpha_text` = 24,817 docs; non-canon confirmed absent (The Pegasus (episode), Gauntlet, Patrick Stewart); canon present (Jean-Luc Picard, Transporter, Prime Directive).
• "Who is Captain Picard?" / "What is the Prime Directive?" / "Who is Sarek?" → clean canonical in-universe answers, no novel/comic/reference-book pollution.

Notes

• Image (CLIP) collection is empty in v0.5.2 (Fandom blocks image fetches from datacenter IPs in CI) — tracked as a follow-up in `memoryalpha-vectordb`. The text RAG path doesn't use it.