Consistent EU doc classification and AI analysis pipeline cleanup

[Copilot]

• Inspect latest PR review thread comments and identify required code/doc updates
• Fix generators.ts leftover legacy references/redeclarations in deep-inspection sections
• Fix Sankey EU/SKR duplicate node/flow handling and keep localized labels
• Fix duplicate const filtered snippet in .github/workflows/news-evening-analysis.md
• Run targeted TypeScript/lint/tests for changed areas
• Run final automated code review + CodeQL scan
• Reply to the new PR comment with commit hash and validation status

Original prompt

---

This section details on the original issue you should resolve

<issue_title>Redesign Article Generator Architecture for AI-First Multi-Iteration Analysis</issue_title>
<issue_description>## 📋 Issue Type
Feature / Architecture Redesign

🎯 Objective

Fundamentally redesign the news article generation architecture (scripts/generate-news-enhanced/generators.ts) so that all political analysis content is AI-generated through multiple LLM iterations rather than being produced by static TypeScript templates with hardcoded text. The current architecture uses template-based generators that produce static SWOT entries, dashboard text, and analysis paragraphs. This must shift to an AI-first pipeline where templates provide consistent HTML structure/styling, but all analytical content (SWOT analysis, political assessments, stakeholder perspectives, policy implications) is generated by the AI agent with iterative quality refinement.

📊 Current State

• generators.ts (1690 lines): Generators (generateWeekAhead, generateCommitteeReports, generatePropositions, generateMotions, generateDeepInspection) produce content by calling generateArticleContent() from data-transformers.ts which uses hardcoded template strings for analysis
• buildDeepInspectionSections(): Creates SWOT entries from document metadata titles and hardcoded SWOT_DEFAULTS map with template strings like "Policy initiative and agenda-setting on %t" — this is NOT AI analysis
• generateDeepInspectionContent(): Generates content sections using document titles/types rather than deep AI-powered analysis of document content
• Content flow: MCP data → static TypeScript transforms → HTML template → output. No AI analysis step exists in the pipeline
• Stakeholder analysis: Fixed 3 perspectives (Government, Parliament/Opposition, Private Sector) with hardcoded SWOT text

🚀 Desired State

• AI-first pipeline: MCP data → document enrichment → AI analysis (iteration 1) → structured analysis output → AI quality review (iteration 2) → AI stakeholder expansion (iteration 3) → HTML template rendering → output
• Templates provide structure only: Article HTML structure, CSS classes, accessibility markup, and Schema.org structured data remain template-driven for consistency
• AI generates all analysis: SWOT entries, political assessments, stakeholder impact analysis, policy domain identification, risk scoring, and watch point generation are all AI-powered
• Multi-iteration refinement: Each article goes through 2-3 AI passes: (1) initial analysis, (2) fact-checking and depth enhancement, (3) stakeholder completeness review
• Configurable analysis depth: CLI flags or workflow inputs control analysis depth (quick/standard/deep) affecting number of AI iterations

🔧 Implementation Approach

Phase 1: Create AI Analysis Pipeline Interface

// New file: scripts/ai-analysis/pipeline.ts
interface AnalysisPipeline {
  analyzeDocuments(docs: RawDocument[], topic: string | null, lang: Language): Promise<AnalysisResult>;
  refineAnalysis(initial: AnalysisResult, docs: RawDocument[]): Promise<AnalysisResult>;
  validateCompleteness(analysis: AnalysisResult): Promise<ValidationResult>;
}

interface AnalysisResult {
  stakeholderSwot: StakeholderSwot[];
  policyAssessment: PolicyAssessment;
  mindmapBranches: MindmapBranch[];
  dashboardData: DashboardData;
  watchPoints: WatchPoint[];
  narrativeContent: string;
  confidenceScore: number;
}

Phase 2: Refactor Generator Functions

• Extract the hardcoded SWOT defaults from generators.ts lines 1069-1086 into AI analysis prompts
• Replace buildDeepInspectionSections() with pipeline.analyzeDocuments()
• Replace generateDeepInspectionContent() with AI-generated narrative that references actual document content
• Keep generateArticleHTML() template rendering unchanged (consistent styling)

Phase 3: Integrate AI Iterations in Agentic Workflows

• Update .github/workflows/news-*.md to include analysis depth parameters
• Add iteration tracking in news/metadata/ for audit trail
• Implement quality threshold checks between iterations

Phase 4: Apply to All Article Types

• Extend AI analysis to committee reports, propositions, motions (not just deep-inspection)
• Each article type gets appropriate analysis depth (e.g., committee-reports = standard, deep-inspection = deep)

🤖 Recommended Agent

code-quality-engineer — This is a major architectural refactoring of the generator pipeline requiring careful decomposition of the existing 1690-line generators.ts into a clean AI-first architecture while preserving backward compatibility with all 12 agentic workflows.

✅ Acceptance Criteria

• New scripts/ai-analysis/pipeline.ts module with typed interfaces for AI analysis pipeline
• All analysis text in generated articles originates from AI, not hardcoded templates
• Multi-iteration analy...

• Fixes Redesign Article Generator Architecture for AI-First Multi-Iteration Analysis #1141

---

💬 Send tasks to Copilot coding agent from Slack and Teams to turn conversations into code. Copilot posts an update in your thread when it's finished.