feat(ai): wire real LLM provider for AI infrastructure summaries

[marcuspat]

v1.1.0 target

The current AIService uses a mock LLM/embeddings adapter. Wire a real provider (OpenAI, Anthropic, or OpenRouter) for production AI infrastructure summaries.

Scope:

• Implement real LLM adapter (suggest: OpenRouter for multi-model flexibility)
• Wire real embeddings for RAG pipeline in discovery domain
• Add provider config via env vars (AI_PROVIDER, AI_API_KEY, AI_MODEL)
• Update skipped AI service tests to run against real or sandbox provider

Acceptance criteria:

• GET /api/v1/dashboard/summary returns LLM-generated narrative, not fixture
• AI adapter tests removed from skip list
• Provider abstraction allows swapping models via config (no code change)

Context: v1.0.0 mock returns static fixture text. Port exists — adapter swap only.