live-evidence.sh 探针创建 run 时缺 stream_mode——D2/D3 可能被低估

见 commit 说明。这条探针的 POST /threads/:id/runs 一直不带 `stream_mode`，LangGraph join 流默认只回放 values，D2/D3 引擎重评（.harness/state/deepagent-eval/2026-08-23-6f84375c/）因此只看到 metadata+values 两种事件类型判了 0.3/0.3。同一根因此前在 apps/api 侧真实修过一次。本 issue 验证：补 stream_mode 后重新采集，D2/D3 是否真的被低估。

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Start with live-evidence.sh and inspect its POST /threads/:id/runs request, then compare the related apps/api fix mentioned in the issue. Add the missing stream_mode, rerun the D2/D3 evidence collection, and verify whether the resulting event types and evaluation scores change from the reported values.

Written by the indexing model from the issue text.