classifyAuditAction checked QA patterns first, so requests like
'生成……清单，重点关注法规依据' hit '法规/依据' and were misrouted
to the QA flow without the checklist panel. Match GENERATE_PATTERNS
first, consistent with classifyAuditIntent; style/refine with a
previous result still take priority.

- emit STOPPED terminal chunk on abort so partial text and pending
  auditResult/auditExport stay attached to the message
- fill '已停止生成。' when nothing was streamed, and skip local
  fallback scoring once the user has stopped
- treat aborted signal (including string abort reason) as terminal in
  the request catch
- drop stale stream/error updates in useChat when the target message
  was removed, e.g. after switching conversations

- drop unused allScoredRecords/groups from audit results
- add slimAuditRecord/toPersistedAuditResult to strip derived fields
  (mdContent/searchText) before attaching results to messages
- sanitize legacy messages when loading conversations so stored
  oversized results are replaced on next save

- Add AI-driven audit analysis engine with batched semantic selection
- Introduce shared audit skill (public markdown file) with business
  filter rules, skill upload in prompt config dialog, and rule matching
- Add risk point "exclude and remember" action writing back to skill
- Support configurable Word report style (fonts, spacing, margins)
- Include AI analysis summary and skill rules in Markdown/Excel/Word exports
- Persist skill object via DataService (ai_skill subject)