fix(agent): preserve successful model retry streaming

[JasonW404]

Summary

• Preserve the step label before model output, roll back only failed attempts, and stream the successful retry's raw reasoning/content before commit.
• Keep accepted content and parse/tool execution consistent across SSE, history, and reconnect.
• Add bounded streaming diagnostics and regression fixtures.

Verification

This is a backport of the effective fix from e38fd50, based on origin/main at 02a54cc. It intentionally does not include unrelated develop commits.

模型	开关	完整会话	用户请求	首轮完成/合规	修复后成功	裸文本兜底	错误/未知	兜底含工具语法
Qwen/Qwen3-VL-32B-Instruct	off	20	60	60	0	0	0	0
Qwen/Qwen3-VL-32B-Instruct	on	20	60	40	20	0	0	0
Pro/zai-org/GLM-5.1	off	20	60	60	0	0	0	0
Pro/zai-org/GLM-5.1	on	20	60	57	3	0	0	0
Qwen/Qwen3.5-27B	off	20	60	60	0	0	0	0
Qwen/Qwen3.5-27B	on	20	60	51	8	0	1	0
Qwen/Qwen3-8B	off	20	60	60	0	0	0	0
Qwen/Qwen3-8B	on	20	60	42	14	4	0	3
Qwen/Qwen2.5-72B-Instruct	off	20	60	60	0	0	0	0
Qwen/Qwen2.5-72B-Instruct	on	20	60	59	1	0	0	0
deepseek-v4-flash	off	20	60	60	0	0	0	0
deepseek-v4-flash	on	20	60	45	15	0	0	0
glm-5.2	off	20	60	60	0	0	0	0
glm-5.2	on	20	60	55	5	0	0	0
glm-5.3	off	20	60	60	0	0	0	0
glm-5.3	on	20	60	42	13	5	0	5
glm-5.3-flash	off	20	60	60	0	0	0	0
glm-5.3-flash	on	20	60	17	15	28	0	20
MiniMax-M3	off	20	60	59	0	0	1	0
MiniMax-M3	on	20	60	29	29	2	0	0
qwen3.7-max	off	20	60	60	0	0	0	0
qwen3.7-max	on	20	60	55	5	0	0	0
qwen3.7-plus	off	20	60	60	0	0	0	0
qwen3.7-plus	on	20	60	50	10	0	0	0
qwen3.8-flash	off	20	60	60	0	0	0	0
qwen3.8-flash	on	20	60	51	9	0	0	0
qwen3.8-max	off	20	60	60	0	0	0	0
qwen3.8-max	on	20	60	59	1	0	0	0

关闭态的“首轮”表示旧链路完成，不表示严格协议格式合规；legacy_direct 列在 CSV 中记录其中的裸文本次数。开启态的“首轮”表示通过代码动作解析并得到最终答案。每个完整会话包含三条连续用户请求。