fix(chat): hide leaked provider citation markers

[loto585]

Linked issue

Closes #625

Summary

部分模型或中转服务会将 ChatGPT 私有联网搜索引用标记直接混入普通文本，例如：

citeturn0search5turn0search15

这些标记属于 provider 内部协议元数据，不应显示在用户回答中。流式响应中该标记还可能被拆分到多个 text_delta，导致仅按单个 chunk 过滤无法清理。

本 PR 在文本流和最终 assistant 消息两个阶段清理 provider citation markers，并增加跨 chunk 状态处理，避免这些标记出现在：

• 聊天界面；
• assistant 历史消息；
• 后续上下文请求；
• 持久化会话数据。

工具调用、图片等非文本内容不会被修改。

Change scope

• Modules: agent-gui
• Key paths:
  • crates/agent-gui/src/lib/providers/runtime/messageUtils.ts
  • crates/agent-gui/src/lib/providers/runtime/textOnlyRuntime.ts
  • crates/agent-gui/src/lib/chat/runner/agentRunner.ts
  • crates/agent-gui/src/lib/providers/llm.ts
  • crates/agent-gui/test/providers/message-utils.test.mjs

Screenshots / preview

这是一个聊天文本渲染行为修复，不涉及布局或视觉样式变化。

Runtime preview：

Before:
Answer content citeturn0search5turn0search15 continues here.

After:
Answer content  continues here.

跨流式 chunk 的场景也会被处理：

Chunk 1: Answer content cit
Chunk 2: eturn0search5turn0search15 continues here.

Rendered:
Answer content  continues here.

Verification

• Targeted provider and agent tests: 108 tests passed

  • citation marker 完整文本过滤；
  • 跨 delta 拆分的 citation marker 过滤；
  • assistant 非文本内容块保留；
  • agent runner；
  • hosted search；
  • native web search；
  • stream retry。

• GUI build:

  pnpm --filter liveagent build

  TypeScript 检查和 Vite 构建通过。

• 分支已同步最新 upstream/main，并确认无合并冲突。

Pre-submit checklist

• A requirement issue is linked with Closes #N.
• Synced with the target branch; no merge conflicts.
• The change is focused, with no unrelated modifications.
• No secrets, tokens, or personal data included.
• Docs are updated for changes affecting user behavior, deployment, or configuration.

[github-actions]

PR governance checks passed. Awaiting human review.

[su-fen]

Follow-up pushed: c95c9eea — fix(chat): unify provider citation marker matching

stripProviderCitationMarkers and the stream sanitizer now share one matcher (\uE200cite + one or more \uE202 segments + \uE201):

• Lookalike text without a delimiter is kept on both the stream and the final message (previously the stream path could strip it while replace() would not).
• Unterminated markers / truncated start prefixes are held while streaming, then dropped on finish() / complete-text strip, so they no longer leak back through text_end.

Added coverage for lookalikes, char-by-char streaming vs strip agreement, and unterminated text_end.