feat(research): Add research module with manual retry, resume, and lifecycle states

[dakshmhatre434-ops]

Summary

This PR adds a comprehensive research module to Levi, enabling AI-powered research sessions with manual retry/resume capabilities and full lifecycle state management.

Features

• Research Sessions: Create and manage research sessions with configurable depth (shallow/medium/deep) and max subtopics
• Research Tasks: Auto-generated subtopics as tasks with sequence ordering and status tracking
• Research Findings: Extracted findings with confidence scoring, categorization, and source attribution
• Research Sources: Track sources with quality scoring and reliability metrics
• Research Memory: Key-value memory store for cross-session learning
• Research Engine: LLM-powered plan generation, content extraction, and report generation
• Search Integration: Support for multiple search providers (mock/Serper/Semantic Scholar)
• Lifecycle States: Full state management - planning, running, cancelling, paused, completed, failed, cancelled
• Manual Controls: Retry failed tasks, resume paused sessions, cancel running research
• Progress Tracking: Real-time progress updates with task-level and session-level reporting

Technical Changes

• Added 5 new DB tables: research_sessions, research_tasks, research_findings, research_sources, research_memory
• Added research routes and service layer
• Added shared validators for all research schemas
• Added research config properties (search provider, LLM model, API keys)
• Added migration journal entries for research tables
• Mock search provider included for development without API keys

Configuration

New environment variables:

• • Enable/disable research module (default: true)
• • Search provider: mock/serper/semantic-scholar
• • Serper API key
• • Semantic Scholar API key
• • LLM model for research
• • LLM API key
• • Max results per search (default: 10)
• • Max findings per task (default: 20)

Testing

• Type checking passes for all research-related code
• Shared package builds successfully
• DB package builds and migrations check pass
• Server routes properly registered in app.ts

Notes

• The research engine uses a mock search provider by default for development
• LLM integration supports configurable providers via environment variables
• All research operations are scoped to company level with proper access control