Provider adapter: OpenAI-compatible and Ollama local LLM routing

[XucroYuri]

Context

L2MAS routes agent work through capability names rather than hard-coded model IDs. The MVP already has deterministic mock providers; the next useful step is a real local LLM adapter that can target OpenAI-compatible servers and Ollama.

Target capabilities

• script.plan
• quality.review

Scope

Add a provider adapter that can call:

• OpenAI-compatible local endpoints, including vLLM and LM Studio
• Ollama HTTP API, either directly or through an OpenAI-compatible bridge

Keep provider selection controlled by config/provider_registry.example.json and ProviderRouter.

Acceptance criteria

• A local provider can be selected by capability and privacy_mode: local-only.
• The adapter accepts endpoint/model configuration from the provider registry.
• The deterministic mock provider remains the default fallback when no local service is running.
• Unit tests cover provider selection and adapter fallback behavior.
• Documentation explains how to point the registry at Ollama, vLLM, LM Studio, or llama.cpp-style servers.

Validation

ruff check .
python3 -m py_compile $(rg --files -g '*.py')
python3 -m unittest discover -s tests -v
python3 examples/test_end_to_end.py