feat: expand fact extraction patterns for personal assistant use

[wtthornton]

Summary

The current 16 extraction patterns in extraction.py are tuned for code-repo contexts ("we decided", "root cause", "convention:"). A personal assistant receives very different input that these patterns miss entirely.

Examples That Fail Today

User Input	Expected Extraction	Current Result
"I prefer mornings for meetings"	preference: morning meetings	❌ No match
"My wife's name is Sarah"	relationship: wife = Sarah	❌ No match
"I'm allergic to shellfish"	health: shellfish allergy	❌ No match
"I always take the 8am train"	routine: 8am train	❌ No match
"I don't like spicy food"	preference: no spicy food	❌ No match
"My dentist is Dr. Smith at 555-1234"	contact: dentist = Dr. Smith	❌ No match

Proposed Solution

Add personal-assistant extraction patterns, ideally gated by profile so repo-brain isn't affected:

Preference patterns

• I prefer ..., I like ..., I don't like ..., I love ..., I hate ...
• I always ..., I never ..., I usually ...

Relationship / identity patterns

• My (wife|husband|partner|boss|doctor|dentist|...) is ...
• I am ..., I'm a ..., I work at ..., I live in ...

Health / constraint patterns

• I'm allergic to ..., I can't eat ..., I don't drink ...

Routine patterns

• Every (Monday|morning|week) I ..., On (Tuesdays|weekends) I ...
• I usually ... at (time)

Tier inference for personal patterns

• Identity patterns → identity tier
• Preference patterns → long-term tier
• Routine patterns → procedural tier (see #procedural-tier issue)
• Temporal/recent → short-term tier

Alternatives Considered

• LLM-based extraction only: Higher quality but adds cost, latency, and API dependency. Better to use deterministic extraction as the primary path and LLM extraction as a complement (with source: agent lower trust).
• NLP libraries (spaCy, etc.): Heavier dependency. Regex patterns are sufficient for common personal expressions.

Effort

Medium — pattern design + testing for false positives across both profiles.

🤖 Generated with Claude Code