pi_llm: LiteLLM stream, tools assemble, credentials, errors

[maozhen520]

Parent

• Spec: #18
• Handoff: #17

What to build

A thin LLM adapter that, against recorded/handwritten chunks (no live API), can stream completions for UI, assemble modern tool calls before execution, probe capabilities, resolve credentials (env then optional auth.json), and surface OpenAI-mapped errors.

Acceptance criteria

• Public stream/completion surface uses LiteLLM OpenAI Chat Completions contract (acompletion, modern tools/tool_choice)
• Streaming yields incremental assistant text suitable for UI consumers
• Tool calls are fully assembled before any execute handoff
• Capability probes report unsupported features without crashing the caller
• Credentials resolve env-first, then optional ~/.pi/agent/auth.json
• Provider/auth failures map to an OpenAI-shaped error surface
• Pytest coverage uses fixtures only — no live LLM required

Blocked by

• Scaffold uv workspace, four packages, and CI