Test: LLM adapter integration tests (Anthropic, OpenAI, Gemini, Copilot)

[JackChen-me]

Context

PR #53 brought overall src/ coverage from 37% to 71%. The remaining low-coverage modules are the LLM adapter chat() and stream() implementations:

Module	Current Coverage
src/llm/anthropic.ts	6%
src/llm/openai.ts	9%
src/llm/gemini.ts	6%
src/llm/copilot.ts	0%

What needs to be done

Mock the underlying SDK HTTP clients to test the adapter chat() and stream() methods without making real API calls:

1. Anthropic adapter — mock Anthropic.messages.create() and messages.stream(), verify toAnthropicMessages() / fromAnthropicContentBlock() round-trip
2. OpenAI adapter — mock OpenAI.chat.completions.create() for both stream: false and stream: true paths
3. Gemini adapter — mock GoogleGenerativeAI client
4. Copilot adapter — mock OAuth device flow + chat completions

Approach

Use vi.mock() or vi.spyOn() to intercept SDK client methods. Each test should verify:

• Correct wire-format conversion (framework types → SDK types → framework types)
• Token usage extraction
• Tool call handling (including streaming tool call assembly)
• Error propagation (API errors, malformed responses)
• Abort signal forwarding

Priority

P2 — the conversion helpers (openai-common.ts) are already at 100% coverage. The untested code is mostly SDK call wrappers. Most likely to break on SDK version bumps, not on framework changes.