Segment the fleet by Job-To-Be-Done from aggregate per-tool telemetry.

~92% of tool usage is local stdio (no call sequence — see /admin/cloud-detail for the thin cloud-relay sample), so real journeys can't be reconstructed. What we CAN do is map each machine's TOOL BASKET (over the window) through the JTBD taxonomy to answer "what are users trying to accomplish". Volume comes from tool_call_events (summed per machine+tool over the window). Meta/plumbing JTBDs (System & setup) are excluded from intent segmentation unless include_meta=true. Unmapped tools are surfaced so the taxonomy stays honest.