fix: replace hardcoded keyword matching with LLM-driven classification

[DeveshParagiri]

Summary

Replaces hardcoded keyword matching with LLM-driven classification at spec/scenario creation time.

• agent_focus_mode (Fragile keyword matching in _normalize_focus_mode() for agent focus detection #100): LLM sets enum (primary_only, couples, all) during sufficiency check instead of runtime keyword parsing
• display_format (Fragile keyword matching in render_intro() for time attribute detection #101): LLM classifies time/currency/percentage format during attribute selection, used in persona rendering
• option_friction (Fragile keyword matching in _position_action_friction() for outcome friction estimation #99): LLM generates behavioral friction scores (0-1) for each outcome option during scenario creation

All three changes move classification to spec/scenario creation time, eliminating runtime keyword heuristics.

Files Changed

• extropy/core/models/population.py - Added agent_focus_mode to SpecMeta/SufficiencyResult, display_format to AttributeSpec
• extropy/population/sampler/core.py - Simplified _get_agent_focus_mode to read enum directly
• extropy/population/spec_builder/sufficiency.py - Added agent_focus_mode to LLM schema
• extropy/population/spec_builder/selector.py - Added display_format to LLM schema
• extropy/population/spec_builder/hydrators/*.py - Pass through display_format
• extropy/population/spec_builder/binder.py - Accept agent_focus_mode, include display_format
• extropy/simulation/persona.py - Use display_format for time/currency/percentage rendering
• extropy/population/persona/renderer.py - Accept display_format_map parameter
• extropy/scenario/outcomes.py - Added option_friction to LLM schema

Test Plan

• All 822 tests pass
• test_agent_focus.py updated for new approach

Fixes #99, #100, #101