feat: migrate LLM backend from OpenAI gpt-5-nano to Claude Sonnet

[panchgonzalez]

Background

Production is hitting OpenAI 429 insufficient_quota errors with gpt-5-nano. This causes silent failures in the lead classification pipeline. Beyond the quota issue, gpt-5-nano is underpowered for the 3-stage pipeline (triage → research → scoring) — classification quality suffers at the cheap tier.

Switching to Claude Sonnet (claude-sonnet-4-6) removes the OpenAI dependency entirely and gives us a significantly more capable model for lead scoring.

Proposed Change

This is a provider swap — pydantic-ai supports Anthropic natively, so no framework changes are needed.

Files to modify

1. pyproject.toml

   • Change pydantic-ai-slim[openai,duckduckgo] → pydantic-ai-slim[anthropic,duckduckgo]

2. src/leads_agent/agent.py (lines 9-10, 105-106)

   • Remove OpenAIProvider, OpenAIChatModel, OpenAIChatModelSettings imports
   • Add AnthropicModel from pydantic_ai.models.anthropic
   • Update agent_factory() to instantiate AnthropicModel(model_name=llm_model_name) — no base_url needed

3. src/leads_agent/config.py (lines 52-54)

   • Change default LLM_MODEL_NAME from gpt-5-nano → claude-sonnet-4-6
   • Remove LLM_BASE_URL field (Anthropic client doesn't take a custom base URL)
   • Rename OPENAI_API_KEY → ANTHROPIC_API_KEY

4. .env.example

   • Update env var name: OPENAI_API_KEY → ANTHROPIC_API_KEY
   • Update default model: gpt-5-nano → claude-sonnet-4-6
   • Remove LLM_BASE_URL line (or leave as optional for Ollama use cases)

5. src/leads_agent/core/init_wizard.py (lines 45-47)

   • Update interactive setup prompt default to claude-sonnet-4-6

Out of Scope

• No changes to agent prompts or system instructions
• No changes to output schemas or Pydantic models
• No changes to Slack integration or Docker configuration (beyond .env)

Testing

1. Set ANTHROPIC_API_KEY in .env
2. Run docker compose up locally
3. Send a test lead through the Slack trigger
4. Confirm all 3 stages (triage, research, scoring) complete without errors in logs
5. Verify ✅/❌ reaction appears on the Slack message as expected