benchmark: single-session-preference category underperforms at 66.7%

[salishforge]

Observation

LongMemEval single-session-preference questions score 66.7% R@5 — the weakest category by a significant margin (next lowest is single-session-assistant at 85.7%).

Analysis Needed

1. Examine the 10 failing preference questions — what patterns do they share?
2. Preference questions likely use implicit/indirect language that FTS struggles with
3. Semantic search (with embeddings) may significantly improve this category

Action Items

• Run benchmark with semantic and hybrid modes to compare
• Analyze failing questions for common patterns
• Consider adding preference-specific indexing or boosting