lov-image-translation-errata

# lov-image-translation-errata 把含有原文与自动翻译的截图改成可直接传播的校样式勘误图：读者既能看懂 正确译文，也能看出旧机翻错在哪里，同时保留原图的身份、版式和信息层级。 ## Triggers ### Activate when - 用户说“给这张图片做翻译勘误”“把错误机翻划掉并改正”或“让读者看到旧翻译有多糟糕”。 - 用户提供截图、海报、社交媒体图片或扫描件，要求核对图中原文与译文并生成修正版图片。 - The user asks to “create an in-image translation errata”, “mark bad machine translation”, or “correct the translation without changing the layout”. ### Do not activate when - 用户只要纯文本翻译或一般文档审校，不需要图片制品；使用翻译或文档审校能力。 - 用户只要干净替换图片文字，不希望保留旧错误痕迹；使用普通图像本地化编辑。 - 用户只想分析视觉风格、生成复刻 prompt 或描述图片；使用视觉分析能力。 - 用户只问某个事实真假，不要求生成图片；使用事实校验能力。 ## User Profile (cross-session) Read `skill.yaml` on every run and resolve the shared `user-profile/v1` context. Current request and project context take precedence over Skill records and shared preferences. Persist only direct durable user statements through `scripts/profile_store.py`; never persist source images, quoted private text, credentials, or inferred personal data. The safe default is an unlabeled track-changes treatment: visible strikethroughs and replacement color already communicate correction. Do not add “勘误”, “Correction”, or another heading unless the user requests it or the marks would otherwise be ambiguous. ## Skill Group Composition Read `references/skill-composition.md` before composing adjacent capabilities. This Skill owns the final corrected image and its acceptance criteria. Upstream research notes and downstream rendering tools are optional artifact handoffs, not hidden sibling dependencies. ## Workfl...