fix(agent): avoid promoting tool thoughts as answers

[wolfkill]

Summary

• prevent the chat UI from synthesizing a final answer from trailing thinking once agent tools were used
• keep the existing natural-stop compatibility path for streams that contain only thinking and no tool activity
• add focused Node tests for the promotion guard

Fixes #1371

Testing

• node --test src/components/modelEditorSourceState.test.ts src/utils/agentStreamEvents.test.ts
• npm run build-only

Note

• npm run type-check currently fails on existing project-wide frontend type errors unrelated to this change, so it is not listed as passing.