feat: Add Ollama LLM provider for local model support

[SolariSystems]

/claim #96

Summary

Fixes #96

Implements a complete Ollama LLM provider for self-hosted local model support, enabling:

• Local Model Inference via Ollama API (/api/chat, /api/generate)
• Model Management - list, pull, delete, show models
• Tool Calling with Beams, Prisms, and Lenses
• Health Checking and automatic model availability verification
• Configuration via environment variables and config files

Features

Core LLM Provider (lib/lux/llm/ollama.ex)

• Implements @behaviour Lux.LLM interface
• call/3 - Main inference endpoint with tool support
• list_models/1 - List locally available models
• pull_model/2 - Download models from Ollama library
• delete_model/2 - Remove local models
• show_model/2 - Get model information
• ensure_model/2 - Auto-pull if model not available
• health_check/1 - Verify Ollama server accessibility

Configuration (config/ollama.exs)

• Model presets: default, smartest, fastest, coding, embeddings
• Environment variable support (OLLAMA_HOST, OLLAMA_MODEL)
• HTTP client configuration (timeouts, pool size)

Integration Tests (test/integration/ollama_test.exs)

• Health check tests
• Model listing tests
• Inference tests (text and JSON responses)
• Error handling tests
• Conditional skip when Ollama not available

Configuration

config :lux, :ollama_models,
  default: "llama3.2",
  smartest: "llama3.2:70b",
  fastest: "llama3.2:1b",
  coding: "codellama"

config :lux, Lux.LLM.Ollama,
  endpoint: "http://localhost:11434",
  receive_timeout: 120_000

Usage

# Basic inference
Lux.LLM.Ollama.call("What is the capital of France?", [], %{})

# With specific model
Lux.LLM.Ollama.call("Explain quantum computing", [], %{model: "llama3.2:70b"})

# List models
Lux.LLM.Ollama.list_models()

# Ensure model is available (auto-pull if needed)
Lux.LLM.Ollama.ensure_model("mistral")

Test Plan

• Health check returns :ok when Ollama running
• list_models/1 returns available models
• call/3 returns valid Lux.Signal response
• JSON response mode returns structured output
• Error handling for non-existent models
• Tests skip gracefully when Ollama not available

---

Generated by Solari Bounty System
https://github.com/SolariSystems

[SolariSystems]

Closing in favor of #344 which includes this work plus the provider abstraction layer.