[Enhancement] Feed Narcissus consciousness insights back into therapy prompts

[LarytheLord]

Goal

Narcissus records cognitive states and detects patterns — but this data is never fed back into the LLM prompts. The agent should be self-aware of its own therapeutic patterns and adjust.

Inspired by: GödelOS (strange-loop concept — system observes its own state), CoALA (metacognitive controller)

Current state

In km_agent.py, step 2 (PATTERNS) collects patterns but only passes them as a flat list:

patterns = self.state_tracker.detect_stuck_patterns(patient_id)
patterns_str = "\n".join(f"  - {p}" for p in patterns)

The Narcissus get_self_awareness_metrics() data is returned in the API response but never influences generation.

Fix

Add consciousness self-model to both the ASSESSMENT and RESPONSE prompts:

# Get full consciousness insights
consciousness = self.state_tracker.get_self_awareness_metrics(patient_id)

consciousness_str = f"""
Agent self-awareness (from Narcissus introspection):
- Sessions tracked: {consciousness.get('sessions_tracked', 0)}
- Dominant intervention pattern: {consciousness.get('most_used_intervention', 'none')}
- Patient engagement trend: {consciousness.get('engagement_trend', 'unknown')}
- Detected biases in my approach: {patterns_str}
- Self-awareness depth: {consciousness.get('self_awareness_depth', 'low')}

Use this self-knowledge to AVOID repeating ineffective approaches.
If I've been using the same intervention repeatedly, try a different one.
If patient engagement is declining, shift my therapeutic style.
"""

Add {consciousness_str} to both ASSESSMENT_PROMPT and RESPONSE_PROMPT.

Expected behavior change

• If agent has used cognitive_restructuring 5 times in a row → it should try journaling_prompt instead
• If patient engagement is declining → agent should ask more open questions, reduce structured exercises
• If agent detects its own "confirmation bias" → it should consider alternative interpretations

Impact

• True self-aware therapy agent — adjusts based on its own performance patterns
• Prevents therapeutic ruts (same technique every session)
• Zero extra API cost — consciousness data is already computed, just adding it to the prompt
• This is the "strange loop" — the agent thinking about its own thinking

Depends on

• [KM-Bridge] Fix missing chimera/__init__.py and shims #9 (init fix)