[TASK] (llm): setup lightweight OpenRouter client with structured outputs

[MikeNovikoff]

Description

LangChain introduces unnecessary overhead for direct text-to-JSON generation. We will use the native openai package patched with instructor to enforce strict Pydantic schemas via OpenRouter

To do

• Install dependencies: openai, instructor, pydantic, python-dotenv
• Setup .env and .env.example
• Create core/llm_client.py using instructor.patch(OpenAI(base_url="https://openrouter.ai/api/v1"))
• Add a basic ping_llm healthcheck script.

Notes

Acceptance Criteria

• Client returns a validated Pydantic model on a dummy prompt, not just a raw string