Feat: Add same-action deduplication to prevent repeated firing

[baabanana]

问题

状态判断僵化：用户反馈：例如"如果我坐着/如果我睡醒了，就...”出现重复触发

目标

能够区别状态和动作持续状态中只触发一次。
”小孩在读书的时候，给我发通知-->在看书的期间只发一次通知”每动作一次就触发一次。
“我比个耶就给我发送一次通知-->每比一次耶就发送一次通知”

解决方法

• 重构触发规则条件判断逻辑，新增 is_same_action 判断，解决同一事件被重复触发的问题
• 引入 _last_happened_cache 内存缓存，保存每个摄像头/通道上次触发时的图像序列，用于与当前画面对比
• 重写 LLM prompt，从简单的 yes/no JSON 输出改为数字编码（0/1/2），分别表示：未发生、新事件、同一事件
• 更新 prompt_helper.py，构建 prompt 时附带上次触发的图像帧和时间戳，供 LLM 进行对比判断

修改内容

• config/prompt_config.yaml: 重写 trigger_rule_condition prompt（中英文），增加 is_same_action 判断流程和输出格式；新增
  current_time_prefix、last_happened_frames_prefix、last_happened_time_prefix 等前缀配置

• miloco_server/service/trigger_rule_runner.py:

  • 新增 _last_happened_cache 字典缓存上次触发的图像序列
  • 新增 _load_last_happened_img_seq() 和 _parse_llm_output() 方法
  • 重构 _check_trigger_condition() 中的 LLM 响应解析逻辑，从 JSON 解析改为数字解析（0/1/2）
  • 删除规则时自动清理对应缓存

• miloco_server/utils/prompt_helper.py: build_trigger_rule_prompt() 新增 last_happened_img_seq 参数，构建包含历史帧和时间戳的完整 prompt

  The trigger rule system monitors camera feeds and executes actions when conditions are met. The current implementation has a deduplication issue: rules fire repeatedly while conditions remain satisfied, causing the same action to execute multiple times for a single continuous event.

Summary

• Add is_same_action detection to prevent the same ongoing event from repeatedly triggering rule execution
• Introduce _last_happened_cache to store the last triggered image sequence per camera/channel for comparison with current frames
• Rewrite LLM prompt from simple yes/no JSON output to numeric encoding (0/1/2): not happened, new event, same event
• Update prompt builder to include last triggered frames and timestamps for LLM comparison

Changes

• config/prompt_config.yaml: Rewrite trigger_rule_condition prompt (Chinese & English) with is_same_action judgment flow and numeric output
  format; add current_time_prefix, last_happened_frames_prefix, last_happened_time_prefix prefix configs
• miloco_server/service/trigger_rule_runner.py:
  • Add _last_happened_cache dict to cache last triggered image sequences
  • Add _load_last_happened_img_seq() and _parse_llm_output() methods
  • Refactor LLM response parsing in _check_trigger_condition() from JSON to numeric (0/1/2)
  • Auto-clean cache entries when a rule is removed
• miloco_server/utils/prompt_helper.py: Add last_happened_img_seq parameter to build_trigger_rule_prompt(), building complete prompts with
  historical frames and timestamps