Benchmark retrieval using the real AI-caller query distribution

Observation (2026-09-04): when coding agents are the primary users of a historical-session search system, human memory-authored questions are the wrong production query distribution. Freeze representative agent continuation, handoff, debugging, and evidence-location queries before tuning; keep humans as blinded relevance/evidence adjudicators. Human-authored questions may remain a separate diagnostic slice, but should not define the main product claim. Source provenance: design review of the Sesh company benchmark and its intended AI-agent consumer.