[ENH] Salience scorer — active attention layer for context prioritization

[OppaAI]

Summary

Currently context window is a passive container — WMC turns + EMC episodes concatenated without active prioritization. Adding a salience scorer would make context assembly an active focus mechanism, deciding what actually matters most right now.

Current Behavior

Context = WMC turns + top-K EMC episodes by similarity. No active attention or salience scoring.

Proposed Enhancement

Add a salience scoring layer in MCC that scores each candidate context item before including it:

class SalienceScorer:
    """
    Active attention mechanism for context prioritization.
    Scores each candidate event segment by:
    - Recency (recent > old)
    - Semantic relevance (EMC similarity score)
    - Emotional weight (valence score)
    - Task relevance (does it relate to current goal)
    """
    
    def score(self, segment: dict, query: str, recency: float) -> float:
        semantic  = segment.get("similarity", 0.0)
        emotional = abs(segment.get("emotional_valence", 0.0))
        return (semantic * 0.5) + (recency * 0.3) + (emotional * 0.2)

Impact

• Context = what GRACE is actively paying attention to
• Not just "what's similar" but "what's useful right now"
• More human-like attention and focus
• Foundation for future agentic reasoning

Notes

• Depends on Enhancement [BUG] VLLM_MODEL points to FlashHead which fails on Orin Nano #2 (emotional valence) for full scoring
• Implement in M3 alongside PMC
• Maps to human prefrontal cortex attention mechanisms