Day-bucketed eval-run activity

One row per UTC day over the requested window, for the activity calendar — so a 12-month chart costs ~365 small rows instead of every raw eval run. Every day is present, including zero days. failed uses the same definition as GET /v1/eval-runs?outcome=failed (a completed run with at least one failing score), so clicking a cell and drilling into the list agrees with the chart. Server-side cached for 60s; the window is capped at 400 days.