Add semantic YMYL classification via LLM extraction

[remete618]

Summary

Two-stage YMYL pipeline that catches implied health/legal/financial content and rejects false positives from metaphorical keyword usage.

Stage 1 (fast): Regex strong patterns for definitive matches (blood type, 401k, DNR order). No LLM call.

Stage 2 (smart): LLM classifies YMYL during fact extraction. Zero additional API calls; the classification piggybacks on the extraction call that already happens for every add().

What it catches that regex misses

• "my chest has been hurting for three days" -> health
• "I owe $40,000 and can't make payments" -> financial
• "I stopped taking my pills" -> medical
• "my ex is threatening to take the kids" -> legal

What it skips that regex falsely flags

• "walked by the bank of the river" -> null (not financial)
• "The Doctor is a great TV show" -> null (not medical)
• "court of public opinion" -> null (not legal)

Files changed

• core/types.py - Add ymyl_category to Fact, ActionItem, Memory
• extraction/prompts.py - Updated extraction prompt with YMYL examples
• extraction/llm_extractor.py - Two-stage: regex override + LLM classification
• conflict/batch_resolver.py - Thread ymyl_category through actions
• core/pipeline.py - Store ymyl_category in metadata, use in YMYL-triggered active retrieval
• core/memory.py - Reconstruct ymyl_category from metadata at search time
• retrieval/temporal.py - Use stored ymyl_category for decay immunity
• tests/test_ymyl_topics.py - 8 new tests

Test plan

• All 163 tests pass
• New: Fact/ActionItem/Memory carry ymyl_category
• New: ymyl_category stored in metadata (and omitted when null)
• New: LLM-classified YMYL memories get decay immunity
• New: Proves regex misses implied YMYL that LLM would catch
• Backward compatible: old memories without ymyl_category fall back to regex