feat(autoresearch): add Layer 4 Save-as-CLI eval with complex zhihu/xhs chains

[jackwener]

Summary

Add autoresearch Layer 4: tests the full "Save as CLI" pipeline (operate init → write adapter → operate verify), ensuring users can reliably crystallize browser exploration into reusable CLI adapters.

Key changes:

• eval-save.ts: test harness (init → write → verify → judge per task)
• save-tasks.json: 26 tasks (20 PUBLIC + 6 COOKIE complex multi-step)
• save-adapters/*.ts: 6 complex adapter files for zhihu/xhs (avoid JSON escape issues)
• save-reliability preset for autoresearch engine iteration
• operate verify fix: no longer hardcodes --limit 3 for adapters without limit arg
• Engine timeout: 180s → 300s to reduce ETIMEDOUT failures

Complex COOKIE chains (6-8 steps each):

Task	Steps	Pattern
zhihu hot+top-answer	6	navigate → hot API → parse big-int IDs → loop fetch answers → strip HTML → merge
zhihu search+stats	7	navigate → search API → filter types → extract question IDs → fetch detail per result → merge
zhihu question+related	8	navigate → question API → answers API → related API → merge 3 layers
xhs search+scroll	6	navigate → MutationObserver wait → scroll 3x → DOM extract with dedup
xhs note+comments	7	navigate → extract meta → scroll container → extract comments → merge
xhs explore+sort	8	navigate → MutationObserver → adaptive scroll → dedup → parse likes → sort → slice

AutoResearch results (5 rounds):

• Baseline: 14 → Final: 26 tasks, all passing
• 3 KEEP rounds (+12 tasks auto-generated)
• operate verify bugfix enabled COOKIE strategy testing

Test plan

• npx tsx autoresearch/eval-save.ts → 26/26
• npm run build passes
• zhihu COOKIE tasks pass (hot-detail, search-detail, question-full)
• xhs COOKIE tasks pass (search-full, note-comments, explore-deep)