[FEAT] Build LLM microservice for structured card generation

[MikeNovikoff]

Feature description

Implement a standalone Python/FastAPI microservice (llm-service) responsible for all AI-powered card generation. The service must communicate with OpenRouter (OpenAI-compatible API) and return strictly validated JSON using Pydantic structured outputs.

What is needed to be done

Scope

llm-service/server.py: FastAPI app with /health and /generate
llm-service/generator.py: OpenAI client wrapper, Pydantic models (MultipleChoiceCard, SingleAnswerCard), system prompts
llm-service/Dockerfile: Python 3.13-slim image
Environment: OPENROUTER_API_KEY, LLM_MODEL (default: openai/gpt-oss-20b:free)

Acceptance Criteria

• POST /generate accepts text, card_type (multiple_choice|single_answer), and count (1–20)
• Returns JSON array of cards with type, question, options, correct_indices, correct_answer, hint
• MultipleChoiceCard generates exactly 4 options with 0-based correct indices
• SingleAnswerCard generates a free-text correct answer
• GET /health returns {"status":"ok"}
• Service is containerized and starts with uvicorn