Switch to LLM-driven agent with zero hard-coded governance logic

Replace old rule-based agent with pure LLM interpretation system.

Agent Changes:
- Rename agent.py → agent_legacy.py (preserve old hard-coded agent)
- Rename agent_refactored.py → agent.py (make LLM agent primary)
- Agent now interprets constitution to understand authority and processes
- No hard-coded checks for specific users, roles, or governance models
- Fully generic: works with any constitutional design

Constitution Interpreter:
- Updated interpret_proposal() to detect authority structures from text
- LLM determines who has decision-making power from constitution
- No assumptions about voting, proposals, or specific governance models

Mastodon Formatting:
- Improved line break handling for bullet points and paragraphs
- Better plain-text formatting for Mastodon posts

Primitives:
- Added support for admin_approval threshold type

Architecture:
- Bot now uses pure LLM interpretation instead of scripted logic
- Each instance can develop implementation guidelines separately
- Guidelines not included in main codebase (instance-specific)

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

src/govbot/governance/primitives.py

@@ -297,6 +297,11 @@ class GovernancePrimitives:
                 return False
             return (agree / total) >= (2 / 3)
 
+        elif threshold_type == "admin_approval":
+            # Admin decision model - no voting threshold
+            # This type means admin must approve, not vote counting
+            return False  # Requires manual admin approval
+
         else:
             raise ValueError(f"Unknown threshold type: {threshold_type}")