feat: NLP patch insights + standalone demo mode

- dlib/ai/insights.py: pure-Python NLP analysis that correlates accepted
  AI suggestion operations/keywords/sections with submission outcomes
  (pending_review / published = positive, archived = negative)
- Backend: GET /api/v1/insights route + service + Pydantic schema
- Frontend: InsightsPanel component with bar charts for operation impact,
  section impact, and keyword signal lift scores
- Insights tab added to the version panel; compact preview on doc overview
- NEXT_PUBLIC_DEMO=true makes the webapp fully standalone: loads
  DEMO_DOCUMENTS / DEMO_SUBMISSIONS / DEMO_INSIGHTS from demo-data.ts,
  disables all mutating actions, shows a DEMO badge in the top bar
- apps/webapp/public/demo-cv.docx: static dummy CV (Alex Rivera) for demo
- scripts/gen_demo_cv.py: script to regenerate the demo DOCX
- .env.example: document NEXT_PUBLIC_DEMO flag

https://claude.ai/code/session_01LWxu2qrwY6BRjUFXXn7NiM

apps/backend/fastapi/app/schemas/insights.py

@@ -0,0 +1,34 @@
+from __future__ import annotations
+
+from pydantic import BaseModel
+
+
+class OperationImpactSchema(BaseModel):
+    operation: str
+    total: int
+    positive: int
+    rate: float
+
+
+class KeywordSignalSchema(BaseModel):
+    keyword: str
+    positive_count: int
+    negative_count: int
+    lift: float
+
+
+class SectionImpactSchema(BaseModel):
+    section: str
+    positive_rate: float
+    count: int
+
+
+class InsightsResponse(BaseModel):
+    total_submissions: int
+    positive_count: int
+    positive_rate: float
+    operation_impact: list[OperationImpactSchema]
+    top_positive_keywords: list[KeywordSignalSchema]
+    top_negative_keywords: list[KeywordSignalSchema]
+    section_impact: list[SectionImpactSchema]
+    has_data: bool