Get Run Latency Stats

Re-aggregate per-pass latency rows on the fly with arbitrary group_by.

group_by is a comma-separated list of dimensions: provider, model, cache_faithfulness, warmup_method, turn_kind.position, turn_kind.has_image, turn_kind.triggers_tools

include_broken=1 keeps cache_faithfulness=broken rows in the aggregate (drilldown view). Default 0 = headline view that excludes them. Broken rows are always counted separately and surfaced as broken_count on each cell.