Add importance-weighted and recency-boosted recall scoring

[hd719]

Problem

Recall ranking is purely based on vector similarity (L2 distance → inverse similarity). This means:

• A low-importance auto-captured message can outrank a high-importance explicitly stored memory
• A 6-month-old fact ranks equally with yesterday's preference update
• No way to prioritize recent or important memories

Proposal

Implement a composite scoring function:

finalScore = (similarityWeight * similarity) + (importanceWeight * importance) + (recencyWeight * recencyDecay)

Where:

• recencyDecay = exponential decay based on age (e.g., half-life of 30 days)
• Weights are configurable with sensible defaults (e.g., 0.6 / 0.2 / 0.2)

Implementation

• Compute composite score after LanceDB vector search returns candidates
• Re-sort by composite score before returning
• Add config for weights and recency half-life
• Keep pure similarity as a fallback mode

Impact

Significantly improves recall quality for long-running assistants where memory accumulates over weeks/months.