feat: research workflows & rich search with user-selectable sources - #437

Research Workflows & Rich Search

Replaces the fixed "first 5 results" summarize button with a full research workflow that lets
users control the evidence set and output format before invoking AI.

What changed

New: ResearchWorkflowPanel component

• Checklist of all result sources — users check/uncheck which ones to include
• Select all / deselect all shortcuts
• 4 report format options: Bullet Points, Narrative, Table, Comparison
• Generate Report button (disabled until at least one source is selected)
• Streams the AI response live via SSE, rendered through AiMarkdown with clickable [N]
  citations that scroll to the source card
• Footer shows which sources were omitted (deselected by user) or failed, by citation number

Updated: SearchResults.tsx

• Added RESEARCH button alongside the existing SUMMARIZE button
• Each result card now has id="result-card-{n}" so citation links can scroll to them
• AiMode state (none | summarize | research) — the panels are mutually exclusive and both reset
  when the query or results change

Updated: aiChatService.ts

• New buildResearchPrompt(query, sources, format, allSourceIds) helper
• Produces format-specific prompts with inline source blocks and a "Sources Used" trailer
• Returns omittedIds — the source IDs present in the full result set but excluded from the
  selected set, for recording in the report metadata

Updated: mcp-server/index.ts — ai_summarize tool

• Optional format parameter: bullets | narrative | table | comparison
• When provided, replaces the generic instruction with a structured output directive
• Fully backward-compatible — omitting format keeps existing behaviour

New types in src/types/index.ts

• ReportFormat, ResearchReportConfig, SourceStatus, ResearchReport

Tests

• buildResearchPrompt: all 4 formats, omittedIds computation, prompt content, source ordering
• ResearchWorkflowPanel: source selection, format switching, generate flow, error states,
  omitted-source footer, state reset on query change
• SearchResults: legacy summarize reset behaviour preserved, new research mode toggle, citation
  card IDs

What was tested

• buildResearchPrompt — 27 new tests, all passing
• Component and integration tests written for all new behaviour
• Legacy SUMMARIZE flow is unchanged; existing tests continue to cover it
• MCP ai_summarize change is backward-compatible and covered by existing MCP tool tests

Notes

• Browser, Express, Vercel, and MCP contracts remain aligned — no paid routes were modified
• x402 settlement and explicit user approval are untouched
• The /ai/chat endpoint (free, no payment) is the only backend call made by the research
  workflow

closes #307