fix: send Anthropic Messages effort

[musnows]

Summary

• send output_config.effort for Anthropic Messages adaptive thinking requests
• keep off mapped to thinking: { type: "disabled" }
• add a model client regression test for /v1/messages reasoning effort

Root cause

The UI and runtime preserved the selected reasoning effort, but the Anthropic Messages adapter collapsed every non-off effort into thinking: { type: "adaptive" }. That enabled adaptive thinking, but dropped the concrete effort level, so compatible upstream services could show an empty reasoning strength.

Anthropic's current docs state that adaptive thinking is enabled with thinking: { type: "adaptive" }, while thinking depth is controlled by the output_config.effort parameter:

• https://platform.claude.com/docs/en/build-with-claude/adaptive-thinking
• https://platform.claude.com/docs/en/build-with-claude/effort

Fixes #281

Validation

• npm --prefix kun test -- tests/model-client.test.ts
• npm --prefix kun run build