Streaming Gate: 字数强约束 (soft/hard limit)

[DankerMu]

Part of #8

Dependencies: #5 (AI Scene Generation), #16 (Workflow Engine)

Description

实现流式生成时的字数强约束——soft limit 发送提示、hard limit 在中文句号边界截断，支持多轮续写。

OpenSpec Spec: `streaming-gate`

Tasks

• 3.1 Implement soft_limit / hard_limit calculation
• 3.2 Implement Chinese sentence boundary detection (regex: [。？！…」』）\n])
• 3.3 Implement multi-round continuation (max 3 rounds)
• 3.4 Persist streaming state to workflow_runs table
• 3.5 Write tests for boundary detection, continuation, state recovery

Acceptance Criteria

• soft limit 触发后 LLM 收到收尾提示
• hard limit 在最近中文句号边界截断
• 字数不足时自动续写（最多 3 轮）
• 流式状态可持久化和恢复