feat(backend): AGI tutor RAG pipeline over course content (#406)

[Degentle12]

Summary

Implements the AGI tutor RAG pipeline over course content (#406). The tutor now answers from indexed course material instead of a static surface, with grounded answers, citations, and a safe "I don't know" fallback.

What's included

RAG pipeline (backend/src/services/tutor/)

• embeddings.ts — deterministic, keyless embeddings (feature-hashed character n-grams, L2-normalized) so no external embedding API is required.
• vectorStore.ts — QdrantVectorStore (REST) plus an in-memory fallback used in tests / no-Docker environments.
• contentProvider.ts — course material provider (chunks seed course content; interface ready for a DB-backed provider).
• generators.ts — extractive answer generator (default) and an OpenAI-compatible generator used automatically when OPENAI_API_KEY is set.
• groundingMetrics.ts — faithfulness (n-gram containment), citation coverage, retrieval confidence, and an overall confidence score.
• ragPipeline.ts — orchestrates retrieval → generation → citation + grounding metrics → safe fallback when confidence is below threshold.

Indexing worker (backend/src/workers/indexingJob.ts)

• Started on server boot (fails open if the vector store is unreachable); embeds and upserts course chunks idempotently, skipping re-indexing when content is unchanged.

API

• POST /api/agi-tutor/rag/ask — grounded answer with citations and grounding metrics (or the "I don't know" fallback).
• POST /api/agi-tutor/rag/index — trigger content indexing.
• GET /api/agi-tutor/rag/status — indexing/vector store status.

Infra & config

• docker-compose.yml — new Qdrant vector store container.
• backend/.env.example — RAG/Qdrant/LLM configuration variables.

Tests

• backend/src/__tests__/ragPipeline.test.ts — 16 tests covering embeddings, retrieval, indexing (including skip-on-unchanged), grounded answers with citations, confidence fallback, and metrics.

Acceptance criteria

• Course material indexed into a vector store
• Queries retrieve relevant context before generation
• Answers include citations to source material
• Grounding/faithfulness metrics tracked
• Fallback to a safe "I don't know" when confidence is low

Verification

• npm run build -w backend passes (tsc, strict mode).
• ESLint clean on all new/modified files.
• New test suite passes: npx jest src/__tests__/ragPipeline.test.ts.
• OpenAPI spec generation + swagger-parser validation passes with the new routes.
• docker-compose.yml parses with the new qdrant service.

Closes #406