v2.4.0 — GenAI / LLM epic

The GenAI / LLM layer everyone asks about — six demos, each with a free-provider live path (Groq / OpenRouter / Hugging Face / local Ollama, via their OpenAI-compatible endpoints) and an honest offline simulation so the whole epic runs on stage with no network and no paid key.

• #98 Prompt playground — one task, four prompts (vague → engineered), scored on a 5-point rubric
• #99 RAG — retrieve from a company knowledge base, answer grounded, refuse when out of scope
• #100 Summarization — extractive offline / abstractive live, with the trade-off spelled out
• #101 Structured extraction — messy messages → a clean table with rules, model for the messy rest
• #102 Zero-shot classifier — route tickets into named categories with no training data
• #103 Guardrails & hallucination — why models invent facts, and a grounding fact-checker

Shared workshop/llm.py auto-detects the provider and labels every answer's source in the UI, so a simulation is never passed off as a real model. Full offline suite: 277 passed, 96% coverage.