feat: extract LLM prompts and settings into external config

[PavelLoparev]

Summary

• Prompt template system: All hardcoded LLM prompts extracted into prompts/ directory with {{PLACEHOLDER}} syntax, loaded by prompt-loader.ts with in-memory caching
• Configurable LLM settings: temperature, max_tokens, max_iterations now readable from env vars (ANALYZER_TEMPERATURE, CHAT_AGENT_MAX_TOKENS, etc.) with current values as defaults
• Settings UI: LLM Settings tab now displays per-service settings grid, prompt templates directory, and masked secrets (last 5 chars visible)
• Snapshot tests: All 5 LLM services have exact-match prompt and settings assertions that verified zero behavior change during migration

Files

New

• src/lib/prompt-loader.ts — template file loader with caching and {{VAR}} substitution
• src/lib/tests/unit/prompt-loader.test.ts — 7 tests
• src/lib/tests/unit/chat-agent.test.ts — 7 tests (new file, previously untested)
• prompts/ — 9 template files extracted from source

Modified

• src/lib/analyzer.ts, chat/agent.ts, report/summary.ts, report-highlights/service.ts, llm-config/service.ts — use loadPrompt + env var settings
• src/app/settings/page.tsx — new per-service settings grid, secrets display
• next.config.ts, README.md, CLAUDE.md — config and documentation

Test plan

• All 264 tests pass (12 new)
• 8 snapshot tests verify prompts are character-for-character identical after migration
• npm run build succeeds
• Manual: verify Settings page displays new sections
• Manual: run a report to confirm LLM calls work as before