Export models and wire into dashboard #102

Summary

Register winning GGUF models in Ollama and wire the dashboard to use them via model-client.ts.

Depends On

• Train and evaluate Qwen 3.5-4B + Gemma 4 E4B on Colab #101 (trained + evaluated models)

Tasks

• Create Ollama Modelfiles for each task with matched system prompts:
  • bishop-state-narrator:{size}
  • bishop-state-summarizer:{size}
  • bishop-state-explainer:{size}
• Register models via ollama create
• Verify each model responds correctly via ollama run
• Refactor explain-pairing/route.ts to use generateExplanation() from model-client.ts
• Refactor query-summary/route.ts to use generateSummary() from model-client.ts
• Remove duplicate createOpenAI instances from refactored routes
• Update generate_readiness_scores.py default model string for --enrich-with-llm
• Add env vars to .env.example: MODEL_BACKEND, OLLAMA_BASE_URL, MODEL_SIZE, SCHOOL_CODE
• Test with MODEL_BACKEND=ollama: all 3 tasks serve correctly
• Test with MODEL_BACKEND=openai: fallback still works

Ollama Model Naming

bishop-state-narrator:{size}
bishop-state-summarizer:{size}
bishop-state-explainer:{size}

Acceptance Criteria

• MODEL_BACKEND=ollama serves all 3 tasks without OpenAI
• MODEL_BACKEND=openai still works as before
• No duplicate OpenAI client instantiations remain in refactored routes