test(llm): lock event reducer laws

[nexxeln]

Summary

• strengthen reducer tests for event order, terminal finish, partial completion, and usage precedence
• assert OpenAI Responses interleaved reasoning/text assembles into semantic message content
• assert Anthropic reasoning signature metadata lands on the assembled reasoning part

Tests

• cd packages/llm && bun typecheck
• cd packages/llm && bun test
• pre-push hook passed: bun turbo typecheck (29/29 tasks)